INTC

U: Intel. Actually, the two big recent news items are both about Intel and Nvidia, right? One is Nvidia investing in Intel. If you read between the lines, it’s really about NVLink. Intel basically ditches OpenCXL and instead uses NVLink so their x86 CPUs can talk to Nvidia GPUs—on servers and potentially on laptops. That could mean dropping the integrated GPU business Intel has been pouring 20 years into, and maybe even axing their discrete GPU unit too—since now they can just bundle RTX IP on the CPU package. And that’s what triggered the 30% jump in Intel’s stock?

V: I mean, I don’t know. It’s tough—Intel’s fundamentals are pretty bad. Nvidia, the government, and SoftBank throwing cash around might keep Intel from going bankrupt for now, but long-term? Hard to invest in. If you just want to trade Intel, fine—there’s plenty of news flow. People speculate Apple might invest. People speculate a “partnership” with TSMC. All of this feels like déjà vu from 2020 when Pat Gelsinger came back as CEO. Stock ran from $40 to $60 then, and honestly Intel was in a way better position back then. The world was still CPU-powered. Today? It’s GPU-powered. So sure, trade the retail-fueled gamma squeezes if you want. But investing in Intel here? I think anyone saying it’s a good idea is either not being honest… or has insider info. I’ve got neither. I’m not gonna sit glued to my screen trading every headline without conviction. So I’ll pass—I’ve got easier ways to make money.

U: But Nvidia does look smart—at least from a gamer’s perspective. Spend $10 billion, wipe out a competitor. Intel actually had something going with their dGPU business—the B60 launched recently and gamer reviews were positive. Sure, nobody really cares about that market, but I think it’s an incubator: if your dGPU works for gamers, your GPGPU will probably do fine too, since compute is inherently simpler. Look at Apple—their GPU is decent for games, which means their GPGPU can perform too if they want. But Intel basically gave up. Can’t blame them—costs too much to develop. Then along comes Nvidia offering an easy exit. It reminds me again why a $500k Trump Club membership “investment” pays off: you get the insider info to trade on. Pretty exclusive crap though. One news drop, and you’ve made the money back.

V: Mhm. I actually put on a small Intel trade back in May. But end of the day, it didn’t fit my style—I wasn’t comfortable holding it, so I dropped it. And my stance hasn’t changed: Intel is extremely hard to invest in. Compare it to AMD. Back in 2016 AMD was near insolvency—trading at $2. Between $2 and $10 it was basically gambling. They whipsawed until 2017 when they finally launched a competitive product—and then the stock 10x’d. For me, I’d never put half a million into a $2 gamble. But from $10 to $100? That’s an investable range. That’s risk/reward I can commit real capital to. Intel today is not there. You can’t just rely on 10% government handouts in a capitalist market. Companies are valued on competitive products and competitive processes. And so far, Intel has shown neither. It is what it is.

U: Thinking about Apple investing in Intel takes some imagination. Nvidia’s deal made sense—worst case, eliminate a competitor; best case, increase Nvidia adoption. And Nvidia doesn’t even have to commit—they could just recommend Intel CPUs to system integrators while still using their own ARM chips. For Apple, though? They tried working with Intel on modem chips—it flopped, so they built their own C1 chip. So if Apple invests, you’d have to imagine a deal structured like Nvidia’s: worst case, remove Intel as a threat; best case, find some synergy. They’re not just going to throw money in and hope for returns.

V: Yeah, from what I’ve heard—Apple’s never going back to Intel CPUs. They’ve got M-series now. And Apple won’t risk Intel’s process tech either—they’re married to TSMC. The most likely angle is Intel’s advanced packaging. But honestly? That’s negligible revenue. Doesn’t move the needle for Intel in the big picture.

U: And from Apple’s side, there’s really no upside either. Packaging isn’t their cost center. They’re not going to gain leverage against contractors through that.

V: Yeah, I mean, right now it’s just speculation. People say other CSPs will follow. But honestly, CSPs got burned by Intel’s poor execution for years. Hard to see them handing Intel new ASIC manufacturing deals. Big question mark there. So sure—if you want to trade on sentiment, knock yourself out. Sentiment is sky-high right now. But me? I’m not touching it.

OpenAI

U: So Nvidia invested $100 billion into OpenAI. OpenAI’s planning to spend $300 billion on Oracle for cloud. What else is in Oracle’s deal? The whole “Project Texas” thing is still alive from the TikTok side—that’s basically the new TikTok/USDS subsidy arrangement, government-backed. Where should we start? Let’s start with Nvidia investing in OpenAI.

V: I mean, I don’t think it’s a bad investment. Basically, Nvidia covers OpenAI’s short-term funding gap before IPO. In return, OpenAI runs primarily on Nvidia’s stack. That’s essentially vendor lock-in—though OpenAI has said publicly they don’t want a single-vendor lock-in, even after this deal. Still, effectively, that’s what Nvidia paid $100 billion for.

U: Yeah, but I don’t think vendor lock-in is a real story. It’s a market narrative, not reality—OpenAI clearly doesn’t want to be locked in. Still, for Nvidia, it’s a good story and a smart investment. We’ve seen this before: in 2018 with crypto, Nvidia “seeded” the market. They’ve learned how to float demand—if you have money, float all the players so demand is sustained. They don’t need to bribe customers to use Nvidia hardware for OpenAI. Just stuff the supply chain with service providers, and the bubble can run much longer—even if it’s already a bubble.

V: So you’re saying it’s a way for Nvidia to smooth earnings, which I kind of agree with. But I’ve seen news suggesting Nvidia might lease GPUs to OpenAI. Honestly? That’s probably fake. Why would they? Nvidia would rather sell chips to Oracle, CoreWeave, whoever. Then on the side, fund OpenAI to buy capacity from those clouds. Much better structure than leasing GPUs directly.

U: Exactly. Putting depreciating assets on Nvidia’s balance sheet is a bad deal. Even car companies know to spin up finance arms to hold the assets and lease them. Holding those directly is terrible. So no—I don’t think Nvidia will ever lease GPUs themselves.

V: Honestly, this kind of arrangement always gets bad publicity. Everyone says “this is just like the internet bubble,” because Cisco did vendor financing back then and investors got burned. So now people scream “this must be the top!” But come on—it can’t be the top. This financing cycle is just getting started. How could this be the peak already?

U: I’ve been trying to find the right analogy for this build-out. The late-90s internet boom doesn’t capture the scale or madness of today. Here’s one: imagine an alternate history starting in 1965. Everyone in the U.S. buys black-and-white TVs. Broadcast companies pop up everywhere. Now imagine TV makers selling directly to broadcasters, and you rent TVs from the broadcaster instead of buying them. TVs depreciate fast—every 3–5 years a new colored standard rolls out, so they ship you a new set. You just keep paying rent. Run that out through the 1980s—that’s closer to today’s craziness and scale.

V: Well, I mean, why do you finance guys all lack imagination?

U: I don’t lack imagination.

V: No, seriously—finance people especially lack imagination. I had this debate with some Twitter account. He was ranting that Nvidia is doomed: China can’t buy, customers are building their own chips, blah blah. He calls this “the top.” I told him, this is just like Apple in 2018. Back then, China pushed Huawei, told officials not to buy iPhones. Developers were protesting the 30% Apple tax. Everyone was bearish. Every active fund was underweight Apple. And then? Apple 4x’d. Today, same setup with Nvidia—despite being the biggest company in the world, active mutual funds are underweight. That’s a massive analogy. This guy laughed at me: “Haha, you’re saying Nvidia will be $18T?” I wasn’t even putting a number on it. My point is: in four years, with money debasement, fiscal madness, and this scale of build-out, >$10T Nvidia isn’t impossible. And vendor financing? Not new. Happens every cycle. Watch Titans of Finance on Netflix—they talk about JP Morgan financing railroads. He lent money to builders and made them buy steel from U.S. Steel, which he owned. Exactly like Nvidia today. Who got rich? JP Morgan. Or take iPhones: early rollout was subsidized by AT&T and Verizon. They financed Apple and consumers. Smaller scale, more consumer-facing, but same arrangement. Nothing abnormal.

U: It’s not abnormal—but it’s still a terrible deal for the middlemen.

V: Oh, absolutely. Like the neoclouds.

U: They take on massive debt financing customers who frankly can’t afford those services. The assets they buy depreciate fast. In an alternate world, those products wouldn’t even exist—cost-performance doesn’t justify them. That’s why it’s a terrible deal for the middlemen.

V: True—and not true. Because those exact arguments apply to the neoclouds right now: Oracle, CoreWeave, Nebius. They are the middlemen, yes. They’re taking on debt to buy depreciating GPUs. But…

U: It’s just hard to tell right now. They’re not making much money—but they’re also not bleeding heavily either.

V: Exactly. That’s the point—it all comes down to execution. The middlemen aren’t pointless. Their bet is: land the customer first, then keep them, upsell, cross-sell, layer value-add services, and actually charge. Then you’ve got a solid business. There are over 200 neoclouds now—half of them are ex-Bitcoin miners chasing gold rush 2.0. Will most fail? Yes. But if even one or two manage to become “the fourth cloud,” that’s the dream. Too early to call.

U: I don’t actually worry about neoclouds yet. They’re not making much, but they’re not hemorrhaging either. Still, someone has to lose money here. Back to my broadcast analogy: in real history, consumers financed their own TVs, so adoption was gradual. Poor families simply didn’t buy. Broadcasters stayed powerful. But in the alternate timeline, broadcasters financed TVs for everyone—including those too poor. They became the losers. Consumers got the service anyway, and the real power shifted to TV manufacturers. That’s Nvidia today: as long as they can push updates every 3–5 years with step-change performance, demand—deserved or not—keeps the race alive. Power concentrates at the manufacturer. It’s bubbly, sure, and will burst someday. But we can finance this bubble for a very long time.

V: Right. We’ll probably overbuild eventually, and at some point supply will exceed demand—the opposite of today. But are we there yet? Not close. Everywhere you look, demand still far exceeds supply. The debt-fueled phase is just starting—Oracle, Nebius, CoreWeave. And a lot of that capacity is effectively backstopped by giants like Microsoft and Google, i.e., by real free cash flow. The first real “debt-led” stress point is Oracle; and if you don’t believe OpenAI will become highly profitable, then you can question the whole chain. But for now, we’re early in debt-backed financing, with room to run. Oracle literally raised ~$18B in 40-year bonds yesterday—5x oversubscribed. Can they raise more? Probably. Can others? Probably. I don’t see the party ending soon on that front.

U: And a lot of that debt gets inflated away.

V: Exactly—and we’re just starting an easing cycle. Much of this debt is tied to benchmarks that are drifting down. Consensus is ~100 bps of cuts over the next 12 months.

U: And this government has zero appetite for austerity. The plan—from start to finish—is to inflate the debt away. So it’ll be a long time before the debt becomes a problem, if ever. In hindsight we might look at depreciated assets and laugh: “An H100 at $30K? Cheaper than my coffee machine. Maybe cheaper than my toaster.” Not a big deal.

V: Yeah. So for the past month I’ve been moving out the risk curve—hunting for more cyclical, junky, high-beta plays on the AI cycle. The neoclouds: Oracle, CoreWeave, Nebius. I’ve owned Oracle before on the pop. I’m even looking at “dumb” PCB providers. We’re at the upswing; my view is you increase beta here, not reduce it.

U: Seems reasonable. The easing cycle just started and isn’t ending soon. There’s even urgency to take on cheap debt—subject to the rate you can actually get, of course.

V: Sure—everyone has to choose their own path. It’ll be bumpy: you can get random 10% air-pockets in this stuff for no reason. But the bigger picture still supports this view.

U: We’re nowhere near the end of this cycle.

Meta TPU

S: OK, our last recording was back in late April. Summer is almost gone, and a lot has happened. The latest news: Meta and Google signed a $10 billion GCP deal. I’m not entirely sure, but it seems related to TPU.

T: Well, I’m not sure Meta really plans to move a lot of training and inference to TPU. If they do, that seems like a very stupid decision.

S: One thing is clear: Google won’t serve the latest and greatest TPU to external customers, no matter what GCP promises. GDM still lacks TPUs for internal research and traffic. Inside Google, GDM’s voice carries more weight than GCP’s, no matter how much GCP wants to be “customer first.”

T: Sure, from Google’s perspective, I agree—they’re capacity-constrained, GPU or TPU alike. But from Meta’s perspective? It makes no sense. If you don’t want 100% dependence on NVIDIA to avoid vendor lock-in, I get it. But switching to TPU is just another, maybe worse, lock-in—with a provider who competes directly with you. NVIDIA doesn’t compete with Meta on models, ads, or social platforms. But if you tie yourself to Google TPUs, you’re literally funding Google’s efforts in areas that compete head-on with Meta.

S: From what I’ve heard, that $10 billion is exploratory—Meta experimenting with TPU. If it were truly meant to cover the next three years, $10B wouldn’t be nearly enough. But it’s big enough for Google to hopefully give Meta preferential TPU access. Still, it’s puzzling why Meta feels the need. They have no TPU expertise. All the PyTorch community collaborates with NVIDIA or AMD; Google is a distant third—or even fourth. Honestly, Apple probably has better PyTorch support now than TPU. From the outside, it looks baffling. I also worry Meta might waste effort again trying to design ASICs for GPGPU workloads. But even then, I don’t think those (old) efforts would fold into MSL. If Meta goes deeper into ASICs, MSL would likely staff up independently. Outsiders can watch that—whether Meta really intends to build their own GPU-like chips.

T: People completely underestimate how hard ASICs are. Especially if you want them competitive with NVIDIA on total cost of ownership. Look at Amazon’s junk—Inferentia, Trainium—nobody wants it. Then there’s connectivity, thermal issues, even basics like heat dissipation. None of it is trivial. It takes multiple painful iterations. Google spent almost a decade and seven TPU generations before they got something halfway decent. Sure, maybe you hire ex-Google engineers and cut some corners, but even then—how do you secure foundry allocation? I don’t know.

S: Maybe as investors we overestimate NVIDIA’s moat. Sure, they dominate on chips, but they don’t monopolize the technology itself. At least four companies—NVIDIA, Google, AMD, Apple—can design and release competitive chips within a generation.

T: On paper, maybe. AMD looks competitive on paper. As for Apple—I don’t really know. You tell me, is it real or just on paper?

S: I’m pretty sure Apple and AMD’s next gen will be competitive. Both are releasing within three months. As for Google, I expected TPU v7 in July or August.

T: They just shared specs at Hot Chips.

S: Okay, that matches. Then v7 should launch this month, though I haven’t seen it. We’ll check. So yes, all these companies can do competitive single-node chips. As investors, we might overestimate NVIDIA’s moat—they’re just in a comfortable spot now.

T: True, but now it’s not just about chips—it’s rack-level solutions: scaling up, out, across. AMD and Apple definitely aren’t there yet. Google maybe, but not the others.

S: Ecosystem matters, sure. But zoom in—NVIDIA can only afford one misstep, maybe one generation (~18 months). That’s luxurious. In this cut-throat industry, one generation is survivable, two is not. So yes, they dominate now, but they’re not unbreakable.

T: Let me push back. What about the B200? Honestly, its ramp was a flop. Supposed to mass-produce last fall, but delayed again and again. Only this quarter did CSPs start ramping. Definitely a misstep. But it also shows how insanely hard rack-level solutions are. NVIDIA figured it out—so from GB200 to GB300 it’s smooth. But for others with no rack-level experience, the learning curve will be brutal. They’re not just “a generation away” from NVIDIA. No way.

NVIDIA

S: From NVIDIA—that’s not what I meant. I meant NVIDIA can afford one generation of misstep. In chip world, that’s ~18 months. So one and a half years, give or take. Anyway, let’s move on to NVIDIA’s story. They just released their quarterly earnings.

T: Yeah. Headlines say results were slightly below buy-side expectations, but I think it was a very strong report—clean beat and clear re-acceleration. Strip out China noise, and revenue was up ~15% QoQ. Guidance is for 17–19% QoQ next quarter. Look at networking: two quarters ago, people panicked that Ethernet ASICs were taking share. But this quarter, NVIDIA’s networking revenue blew out expectations. Margins too—two or three quarters ago gross margin fell from mid-70s to low-70s, sparking fears. Now it’s climbing back toward mid-70s, exactly as management guided: biggest product ramp ever, margins dip, then recover. I’m not the biggest fan of NVIDIA’s CFO—she could explain better—but at least management is being transparent.

S: But NVIDIA is even offering discounts on B300 and GB300 in promo emails—what’s going on? And I’m suddenly getting cold emails and calls from Supermicro sales. Last year, early on, you couldn’t even get through to them—they were overwhelmed with demand. Now they’re chasing customers. That’s a big shift.

T: Well, Supermicro is its own story. After their accounting fiasco, they’ve been suffering on both the supply and demand side. So I’d bet this is more of a Supermicro issue than an NVIDIA issue. Also, I haven’t heard of anyone—besides CoreWeave—getting real B300 allocation yet. No idea what’s going on. Are you on some secret pre-release list or something?

Unemployment

S: What about tariffs and unemployment? And now rates are going lower. California unemployment was 5.5% in August, I think. What’s happening there?

T: First, summer unemployment numbers are noisy—tourism, summer jobs, volatility. I wouldn’t put much weight on them. Second, honestly no one knows what’s going on with policy. On one hand, the OBBB stimulus is still running into next year. On the other, tariffs are clearly inflationary. Back in April we said “nobody knows,” but now it’s obvious: tariffs are inflationary. The only real question is whether it’s transitory or lasting. I hate the word “transitory”—Powell in 2021 scarred it—but no one really knows. The bigger question: can growth accelerate enough to outrun the inflation?

S: I mean, we've been 2% inflation for extended time before COVID. So that's what it's sort of conditioned us to that thought, like 2% inflation year over year, but that's not normal, right?

S: We had ~2% inflation for years pre-COVID, so people got conditioned to think that’s normal. But it’s not, right?

T: Yeah. I asked Gemini to do deep dives comparing 1990s vs 2000s inflation. Results were volatile. On average: early 2000s ~2.5–3%, 1990s ~2.7–3%. Definitely higher than 2%. What matters more is real GDP growth. That’s why Powell is leaning growth-oriented, not just inflation-targeting. If we can hold real GDP at 2.5–3%, we’re fine.

S: I think we’re fine, and markets agree. Housing’s moving. I even saw mortgage rates back in the high-2% range—with some strings attached of course. Markets are anticipating Fed cuts.

T: How is that possible?

S: Honestly no idea. My worry is the Fed chickens out and cuts rates, but the 10-year doesn’t follow.

T: Well, since Jackson Hole, the 10-year’s down ~10bps, from ~4.3% to ~4.2%. Not huge, but not nothing. Still, it’s hard to read. Lots of noise: courts ruling Trump’s tariffs illegal, Trump pushing appeals, etc. I think tariffs will stay.

S: I think so.

T: Yeah.

S: Exactly. It’s a cornerstone of Project 2025, alongside federal overreach. Tariffs will stay—high tariffs for a long time. Inflationary, yes. But will it be one-time or slow burn? My thesis: slow burn. It doesn’t feel elastic. Just vibes, but that’s my read.

T: Talking about vibes, right? Yeah.

GPT-5

S: We haven’t talked about the GPT-5 release yet.

T: Oh, I haven’t even tried it. I’m happy with Gemini. Every query I send it, it does deep research. Overkill sometimes, but it’s cheap, fast, and not rate-limited. I honestly haven’t used OpenAI or ChatGPT in half a year.

S: Actually, GPT-5 is a good release. You no longer pick models—it just works. It has fresher built-in knowledge, so recent questions are answered directly without search. Search always made ChatGPT’s worldview conflicted—what it “knew” vs what it just scraped. GPT-5 has neural knowledge, so answers are cleaner. And ChatGPT’s app is better than Gemini’s—Gemini still has bugs, like losing your entire conversation if you background the app. Basic stuff. People dismiss GPT-5 too quickly. It’s going to serve close to a billion people. People underestimate OpenAI.

T: Who said that? No one I know. Everyone I know has been trying to invest in OpenAI through SPVs, even paying $500B valuations while they’re officially raising at $350B. Who exactly are you talking to?

S: I don’t know. Maybe the people I talk to are too forward—living in the future. But Google’s doing great too. Gemini 2.5 Pro is a very good model. They’re not releasing Kingfall, but 2.5 Flash Image—nicknamed Nano Banana—is excellent at image gen and editing. Lots of strong models out there.

T: We’re living in different worlds. Every investor I know worries Google has structural issues—losing search moat, not aggressive enough against OpenAI, risk of becoming the next Yahoo. Totally different worldview from yours.

S: Maybe some truth. But little things matter. I stopped using Gemini for quick questions—I just use Google Search + ChatGPT. Search has AI mode now, which helps, but sometimes it flips back to classic search randomly. Gemini always insists on searching first. GPT-5 is better for direct answers—it just gives them, fast. Those little things matter. And inside Google, no one builds careers by fixing little annoyances.

T: By making customers actually love the product.

S: Exactly. In big companies, people don’t build careers by making the product nicer for users. They build careers by moving metrics.

T: My experience is the opposite. One day I uploaded a math problem photo to Gemini—it solved it fast, correctly, though the method was clunky. I tried the same with ChatGPT: 4.0 gave a wrong answer, 0.3 froze for ten minutes then gave another wrong one. That experience was so bad I stuck with Gemini. But maybe you’re right, maybe I should give GPT-5 a try.

S: Yeah—and if you use Nano Banana, it’ll even write the answer directly on the picture.

T: On the picture? Oh man. And it wasn’t even a difficult math problem—just a normal one.

S: Yep, surprising. OK, let’s end on that note.

Google

Q: Google reports earnings this week, and it feels like the numbers are already baked into expectations. What’s your take?

R: I’d call it a perfectly ordinary print—nothing spectacular. Expectations were bleak: everyone was down on Search; YouTube was expected to miss; GCP wasn’t supposed to shine. So the “beat” was mostly a function of a low bar. On the call, every analyst—each in their own wording—asked whether Google’s CapEx would change if the macro backdrop deteriorates. Pichai and the CFO repeatedly made it crystal-clear: CapEx stays. If the economy weakens, they’ll squeeze headcount and comp—the same layoff-plus-efficiency theme we’ve seen across tech for two years.

Q: Makes sense. They’re the biggest company facing the Innovator’s Dilemma in our lifetime. Yet they’re well positioned: a cash printer, ample corporate fat, and two legendary founders people forget are still top-10 billionaires. Watching the corporate immune system fight the dilemma is fascinating. Gemini 2.5 is solid, but look at Claude’s “Artifacts” (OpenAI calls it Canvas): a pane to stash generated code or content. Super handy, yet neither the web nor the mobile app has it—no code highlighting, nothing. A perfect mirror of their inertia.

R: They also have some—let’s say—“creative” AI metrics. Sundar announced that AI Overviews are wildly successful, rolling out to all of Search and monetising roughly on par with legacy results.

Q: They have to claim that—it’s the keystone of their Innovator’s-Dilemma survival plan. I always click the citations because I know LLMs hallucinate, but an unsophisticated user might not. AI Preview itself is a destination: it doesn’t push you elsewhere, so no advertiser is bidding for that click-through. Old-school Google tried to eject you from the site; AI Preview wants to keep you there. That’s the dilemma—no sugar-coating it. Internally and externally they must insist it works; otherwise the plan collapses. We outside the PR bubble know it will ding margins, but as long-term investors we also know it’s the only path they’ve got.

R: Exactly. GCP still trails AWS and Azure. AI helps, but the 2020 “cloud optimisation” mantra is coming back. Add accelerated depreciation: CapEx was $50 bn in 2024, guidance is $75 bn this year—over 30 % growth. Free cash flow will bleed; there’s no dressing that up.

Q: Early last year people still scrounged for A100 80 GBs. By January it was obvious: outside tiny training clusters, nobody uses A100 for inference. They keep them for training because the clusters exist and A100 depreciation is gentler than H100’s. H100 doubles performance but rents at $1.50/hour; A100 is $1.00, so no one chooses A100 for new work—even on a two-year horizon.

R: Right—the A100 shipped in 2020, so we’re three years in. Call it done.

Q: H100’s trajectory is different, especially once B200 arrives. There are more H100s than A100s in the wild; two years from now, with B200 mainstream, H100 rental could fall below $0.50/hour. That’s a brutal curve—nothing this steep since the Intel CPU refresh frenzy a decade ago. GPUs went from a four-year life to something far shorter.

R: Quantity matters too. CSPs still depreciate GPUs over five-plus years on paper. Anyone running the racks knows the real economic life is maybe three. That’s right in Jensen’s sweet spot. Cloud providers are squeezed unless they can pass the cost on. Rumour is Google is leasing CoreWeave’s H100 fleet.

Q: Makes sense.

R: Word is the rate is under $1.50 per GPU-hour—basement pricing.

Q: Yeah, bizarre: H100 at $1.50 while a 5090 gaming card rents just above $1. Pricing is all over the place.

R: Only way CoreWeave can go that low for three-plus years is if those servers are fully written off—no financing cost, just power. Then sure, rent them; otherwise the math collapses.

Q: Fair point. If the boxes are off the books and Google signs a long contract, everyone wins. Anyway, shifting gears: Google’s ad revenue shouldn’t be affected by the May 1 “Liberation Day” timing in Q1, right?

Walmart

R: Let’s be clear: Google never gives formal guidance. What we got was a Q1 snapshot. Analysts tried to pry into how ads are trending so far in Q2, but management stone-walled: “We’ve only seen a few days—too early to comment.” In short, don’t read anything about “post-Liberation Day” demand into this print.

Q: Then what about Facebook—Meta? What’s the setup?

R: Meta’s chart is wild. From the YTD peak around $740 it cratered to $500 after Liberation Day—-33 %. It’s crawled back to ~$550, still a hefty discount. If a recession really is coming, that pricing looks rational: ad revenue is the first casualty.

Q: But are we even heading into a recession? Walmart and Target just told suppliers to keep shipping.

R: Right, huge signal. My read: there’s >50 % chance the retailers have an implicit promise from D.C.—either tariffs get paused or resolved before those China containers hit port. Yes, Walmart tells suppliers, “We’ll cover tariffs,” but they still fear demand destruction if price tags jump. No one wants a 2022-style Target inventory write-down. More plausible: the White House quietly assures them “Give us two-three weeks; we’ll sort it.” Coincidentally, the CEOs met with Trump last week. Net-net, odds of a U.S. recession just dropped.

Q: They could just get a carve-out—retailer-specific exemptions.

R: Politically radioactive. It would look like Washington hand-picking winners and losers. Every food-and-beverage importer still paying tariffs would howl. Even GOP rank-and-file would balk.

Q: Politics aside, if tariffs stick, Meta eats it more than Amazon. Amazon’s alphabet-soup private-label sellers just mark the import at $1 and pay $1 duty—no big deal. The hit lands on branded SMBs—the very DTC players who scaled on heavy Meta/Instagram spend. Think fashion upstarts, Anker/Roborock-type electronics, Narwal vacuums. Their ad budgets vanish first, so Meta hurts disproportionately.

R: Fair—Meta would be the biggest casualty. Still, I think targeted exemptions would be a political nightmare.

Q: This administration’s racked up “disasters” all year and keeps polling fine as long as deportations and D.O.G.E. cuts. I doubt SMB pain moves the needle.

R: Deportations don’t spike unemployment. Smacking SMBs might—and then unemployment headlines get ugly.

Q: Which would shove the Fed toward cuts: Walmart/Target hold prices, inflation stays tame, unemployment ticks up—perfect cover to ease.

R: We’ll agree to disagree. I can already picture cable news parading irate shop owners 24/7—“government picked winners!” GOP would bleed, Trump popularity or not.

Q: Realistically, no way we ink a China deal in three weeks.

R: Maybe just a tariff pause.

Q: A deal with India, sure. China—maybe.

R: Let’s call it “maybe.”

Q: Total coin-flip. We’ll see.

R: Speculation aside, “keep shipping” is the only hard data. With that, recession odds look meaningfully lower than two weeks ago. Even if every China-linked DTC brand quits Meta overnight, that’s maybe a 3–4% revenue hit—and auctions re-fill empty slots. Worst-case is 4%, which is extreme. So Meta isn’t as doomed as the tape implies. My bigger worry is their AI strategy…

Meta

Q: LlamaCon is next Tuesday.

R: Next Tuesday—got it.

Q: Yup.

R: Then we’ll finally see their next-gen “reasoning” model.

Q: Speakers are Satya and—pretty sure—the Databricks CEO, plus Meta folks (Chris Cox, Manohar, Angela Fan). So: new reasoning model, plus some shiny consumer features—likely baked into the main apps. We’ve already had “Llama Maverick” and “Llama 4 Scout.” Maybe Behemoth lands, maybe it slips. Rumor mill says Qwen 3 and DeepSeek R2 could drop Monday on the China side, so we’ll see a flurry.

R: DeepSeek already pushed an R1T—basically a bargain-bin R1 model.

Q: End of the day, whatever DeepSeek ships next won’t leapfrog. Even with 996 schedules, you still need bake time. GPUs grind 24/7, but you have to nail data mix, curriculum, RL-fine-tune, etc. Meta’s habit of setting a hard LlamaCon date before research is done makes me nervous. In contrast, Google doesn’t pin model launches strictly to I/O Day. Credit where due: FAIR still publishes gems—yesterday’s paper used an LLM as a black-box optimizer to iteratively guess what’s in a picture from abstract feedback. Super-clever. Meta has deep research; it’s just siloed. ByteDance? No research culture—seven or eight competing LLM teams throwing spaghetti at the wall. Alibaba similar. U.S. players run fewer parallel bets—Google has two-ish big-model teams and, hey, Gemini/Gemma 3 turned out great; lots of shops already replaced Llama with Gemma 3. Google stays research-driven; product follows. Anyway, I’m rambling.

R: Fair point: prices shape stories, not vice-versa. When Meta ripped 20 days straight, the narrative was: “Meta’s crushing AI—models commoditize; he who owns distribution wins.”

Q: That thesis still holds.

R: Exactly. Same fundamentals, but a 30% drop and suddenly “Llama’s dead, Meta’s losing the AI war.”

Q: Distribution still trumps models. Look—OpenAI offered $3B for Windsurf after missing Cursor at $10B. All about user reach.

R: They also waved a hand: “Hey, if Chrome’s up for grabs, call us.”

Q: Political theater. If anything, it strengthens the case for Chrome staying inside Google. Browsers should move cautiously—no overnight LLM injections. I’m still salty Chrome killed JPEG-XL, but I’m a Firefox user so…

R: Meta still has the TikTok cloud. Honestly, TikTok will just keep operating until the next administration—no ban, no forced sale, no clean resolution.

Q: True, but Beijing explicitly said TikTok isn’t on the tariff table—you can tweak tariffs anytime; TikTok can only be sold once. TikTok keeps calling DC’s bluff, hinting at self-shutdown. Even when it disappeared from app stores for a month, the devs kept cranking. Uncertainty is toxic, yet China won’t swap TikTok for lighter tariffs—full stop.

R: Exactly—let TikTok slide and precedent says SCOTUS rulings are optional. Slippery slope.

Q: It’s the selective-enforcement circus: Congress said “delist TikTok,” but DOJ says, “delist and we sue you.” So Apple and Google keep TikTok and are technically breaking the law daily. Next president probably won’t litigate either. No private right of action, so the law is toothless. Total mess.

R: Exactly—dangerous slope all around.

News

O: We were sick last week, so this is our first real catch‑up—which felt like “we’re so back, we’re so over” on repeat. Let’s start with tariffs and recap what happened in the past two weeks.

P: Tariffs? Yeah. Clearly they’re not going well for the U.S., right?

O: From our last discussion, China did a 145% levy, and the U.S. responded by raising it to 245%.

P: They’ve exempted semiconductors and consumer electronics—for now at least.

O: We’re expecting tariffs specifically on semiconductors, and as of today there shouldn’t be any trade deals. Would you agree?

P: Right—it’s going to hit many U.S. industries hard: electric vehicles, solar panels, and so on. An embargo on rare earth materials would be devastating.

O: Solar panels aren’t really impacted—since the U.S. doesn’t manufacture them; they’re all made in China. So solar-based climate initiatives suffer—but climate isn’t on the agenda for the next three-and-a-half years anyway. It’s effectively off the table, so no one will invest in it, and we’ll just deal with the fallout. Electric vehicles—and Tesla in particular—will face challenges because Tesla does produce batteries and cars here in the U.S. But that doesn’t mean they can’t pivot to other import sources. The U.S. market still functions under its own trade rules, though deregulation might bring mining back.

P: Think about it—if you’re an investor eyeing a rare‑earths company, you need stable policy support and continuity. You don’t want to pour capital into an industry that takes years to ramp up, only to have the China embargo lifted after a couple of years, margins collapse, and you go bust.

O: Maybe they’ll innovate production—like shale gas: shut down when it’s uneconomic, ramp up when profitable, with zero cost when idle. Businesses can be nimble that way, so I’m somewhat optimistic. Anyhow, back to tariffs. Over the weekend, our info lagged—people thought Jensen’s lunch with Trump meant NVIDIA’s bans (including the AI diffusion rule) would be lifted. But by Wednesday, NVIDIA confirmed a China‑market write‑down: the AI diffusion rule remains in force, so they can’t sell H20 to China without a license—effectively an export ban. Jensen must’ve known by Sunday—he was already in China when the news broke.

NVIDIA

P: Writing down $5.5 billion stings, but it doesn’t alter Nvidia’s fundamentals. What’s far more concerning is Washington’s hawkish export‑control stance. Nvidia answers to shareholders, and bans make its business harder. I doubt this is national interest—Hopper and A100 (Ampere) are four‑year‑old platforms, lacking unique Chinese IP. They sell because CUDA’s ecosystem is dominant. Banning mature tech makes no sense. If you embargo the latest gen, you can argue for a U.S. lead, but targeting old tech reeks of bad faith. Look at the White House Science and Technology Director—hawkish on controls, with ties to Palantir and Peter Thiel, and David Sacks backing stricter bans. I suspect they’re using “national security” to cripple Nvidia, commoditizing GPUs—Huawei or whoever benefits. They aren’t fools—they’re protecting vested interests and hurting Nvidia on purpose.

O: So we’re manufacturing our own “Cisco moment” instead of letting the market create one. Remember: Cisco was undone by cheap Huawei gear, plus the U.S. helped sink Nortel. Not exactly friendly to Canada. Anyway, hobbling NVIDIA won’t automatically commoditise GPUs. You might get talent outflow that seeds new startups—that’s some silver lining—but what you really get is bifurcation. Huawei will sell to China at near‑NVIDIA prices, the U.S. will bar Huawei parts, so China‑based firms serve China, NVIDIA and U.S. startups serve the West, and everyone else picks a camp. That’s bifurcation, not commoditisation. And trying to engineer such a moment always fails—history shows those “playing god” are doomed. Twists and turns will foil the power players’ plans.

P: Speaking of which, why do you think Jensen went to China this week?

O: Rumors abound—but I doubt them. Leaks say Jensen implied Nvidia chips don’t have to be fabbed in Taiwan or the U.S., hinting at a deal with SMIC to bypass controls, or an IP transfer. Porting a 3 nm design to 7 nm requires massive rewrites, and Nvidia likely won’t allocate U.S. resources. MediaTek, though, is a trusted non‑U.S. partner—they’ve worked with Google and Nvidia. It’s plausible but a long shot, and it would infuriate Washington.

P: There are precedents—AMD licensed x86 IP to a Chinese firm (Hygon/Haiguang). So it’s not unprecedented.

O: That was a different era—Xi’s first term, U.S.-China relations were in a honeymoon phase, and the “China threat” narrative was muted. AMD was fighting for survival; banning x86 exports would have killed them. Today’s political climate is far less forgiving—a similar license now would provoke fierce opposition in Washington.

P: My point is you need to be savvy with licensing. At GTC, Nvidia discussed licensing NVLink to MediaTek—you could rebrand GPU services as an “ASIC/XPU service.” If companies like AVGO can serve Bytedance-driven Chinese firms, why can’t Nvidia?

O: Jensen’s worst nightmare is losing CUDA mind‑share. He doesn’t need NVIDIA chips in China, he needs CUDA talent in China. Chinese docs on warp specialisation vs. multi‑stage are the best I’ve seen—deep dives, architecture nuances, real examples. Hopper favors warp specialisation; earlier GPUs favored multi‑stage; Chinese devs document that better than anyone. Bottom line: the bigger your CUDA developer bench, the better, and China may already have more CUDA devs than the U.S. (We have more React devs—yay SaaS.) If those engineers ever say “Bye CUDA, hello Ascend,” NVIDIA’s moat evaporates.

P: It’s not about selling billions of chips to China—what matters is that everyone remains on CUDA.

O: Software knows no borders—everything’s open source. Chinese CUDA optimizations help U.S. developers, who then run them on Nvidia hardware. Nvidia’s moat only strengthens.

P: Sure—but shutting out Nvidia from China achieves the same. So can Nvidia simply license CUDA to MediaTek instead?

O: They don’t need to license full CUDA—just share enough specs for MediaTek to build CUDA‑compatible hardware, following Nvidia’s instruction‑set updates. That’s crucial, since no third party can replicate CUDA yet. It’s positive if executed, but Washington would freak out. Jensen is the only Asian among the top ten billionaires—he’s a lonely figure.

P: Is he still in the top ten?

O: He’s dropped to around #14. Everybody ahead is white—except one Indian billionaire, who now surpasses Jensen. He’s in a lonely spot. I suspect some in this administration want to undercut high‑achieving Asians—it’s grim for U.S. diversity.

P: Indeed.

Llama 4

O: Llama 4 dropped two weeks ago, and benchmarks reveal Llama 4 Maverick uses special optimizations for LMArena. LMArena ranks models by head‑to‑head user preferences—Llama 4 Maverick scored ~1400—yet the public release scores lower, since they removed the “friendly” prompt tweaks (compliments, colorful prose, emojis). LlamaCon is next week—we’ll likely see a dedicated reasoning model. Overall, Meta seems to be slipping behind: O3, O4 Mini, Gemini 2.5 Pro each have trade‑offs.

P: I love O3. It blows my mind. If progress stays this fast, maybe I can retire in two years—or even sooner.

O: Of course, these models still struggle on certain tasks. O3 Mini’s multi-modal and tool‑use capabilities exceed expectations for a smaller reasoning model, but knowledge gaps remain. Gemini 2.5 Pro outperforms on some programming edge cases—Apple Metal code, for example. Even CUDA code generation trips O3 up, inserting syncthreads incorrectly. Knowledge limitations worry me, and China’s pace—e.g., DeepSeek’s R3 March ’24 update—overshadows Llama. Personally, I favor Mixture of Experts: with fixed compute, MoE packs in more knowledge (parameters) and speed (activations). I see LLMs as knowledge encoders, not AGI itself. The U.S. recession likely delays AGI here.

P: Recession aside, let’s close out on Llama 4. There’s still hope at LlamaCon, right?

O: Reasoning models thrive only with deep coding knowledge, so I’m keen to see it in action. LlamaCon is imminent—Meta pioneered RoPE and extended it for longer contexts, benefiting everyone. Partnerships with Microsoft and Amazon are rumored, too.

P: That sounds reasonable.

O: I wouldn’t overinterpret it—VPs reach out all the time. Zuck can absorb it; Reality Labs, VR/AR, and GenAI apps are firm cost centers.

P: Sure. But calling them all “laggards”—Amazon, Microsoft, and maybe even Meta?

O: Microsoft isn’t a laggard—Satya locked in strategic options (OpenAI stake etc.). Microsoft holds the cards. AWS invested in Anthropic. Facebook, however, can’t use OpenAI, Claude, or Gemini, so must build its own stack. That places Meta in a unique spot, despite “laggard” rumors.

P: Right—let’s ditch “laggard.” None are pushing the absolute cutting edge. For any CSP, supporting leading models is logical—the API market will commoditize. Reports say Meta can’t bear all the CapEx and is courting Amazon and Microsoft for a cost share.

O: I disagree—Zuck is willing to fund it, given Reality Labs’ budget. Of course, if a recession hits, U.S. companies will cut CapEx quickly.

P: Indeed—it’ll be interesting to see whether they slash OPEX (headcount) or CapEx first.

O: You mean operating vs. capital expenses?

P: Yes—operational cuts, headcount.

O: They always trim OPEX, so CapEx cuts make sense too. I suspect AGI progress here slows as a result. Maybe China will achieve AGI first—they face an aging‑population crunch and won’t delay.

P: You mean the aging population?

O: Exactly—the aging demographic. China has top STEM talent, often with limited job options. I’m pessimistic the U.S. will outrun China to AGI given our instability. Did you see China’s robot half‑marathon?

P: Yes—it was disappointing.

O: Exactly—there’s often hype. Without proof, skepticism is warranted.

P: I’m not bashing Unitree, but they set a poor precedent: more marketing than product polish, courting Xi Jinping’s attention while the marathon performance underdelivered.

O: I’ve followed Chinese robotics for years. Unitree’s G1 has strategic issues—too short, marketed as a dev kit, not a finished product. Same story we saw with Ubitech (ended up selling toy robots). Some engineers left for Robosen, which now makes those awesome Hasbro Transformers—buy one now; after a 145% tariff they’ll be triple the price. Unitree launched in 2020 with a $2k quadruped—a breath of fresh air: cheap, open, no bureaucracy. U.S. universities pounced because a DIY Mini Cheetah costs $4–5k in parts alone. That success locked Unitree into dev‑kit thinking, and they’re missing the humanoid window Tesla’s aiming at. But China’s sheer scale matters: for every flop there’s a hit, IP leaks fast, and the whole ecosystem levels up (see Roborock vs. Dreame/Ecovacs). Even if Unitree fails, some other shop will nail a humanoid, lead a few years, then everyone will catch up. That rapid diffusion just doesn’t happen in the U.S. now.

P: Mhmm. So the takeaway is we’re fucked. I’m not alone—many share this bearish view, though I might be overdoing it.

Trade War

O: That doesn’t guarantee China will win the trade war.

P: We shouldn’t sugarcoat it—China’s economy has weaknesses, and they’ll sacrifice much in a trade war. I’ve joked the Chinese will drink their own urine before America’s blood, but that’s grim. Xi probably prefers a deal—he’s building leverage with Europe, Japan, Southeast Asia, and waiting for U.S. inflation, unemployment, or recession to pressure Trump. Despite assurances of “no recession” and “90 deals in 90 days,” no such deals exist yet.

O: Yet markets remain calm despite port crises and halted shipments. Do you think they’re pricing in a man‑made supply shock—or ignoring it?

P: Folks assumed U.S. retailers and companies stockpiled three to six months of inventory, but they likely have only 45–60 days. Many thought Trump would back down as part of his “art of the deal,” fueling optimism and keeping the S&P around 5,200 instead of plunging to 4,500. Surveys show over 50% expect a hard landing, with sentiment swinging from record bullish in December to record bearish now—but positioning remains lofty, revealing a battle between emotion and reason.

O: The Right calls this “deprogramming”—freeing yourself from the belief that U.S. hegemony and bull markets are permanent. Once deprogrammed, you see people’s true priorities. Besides coffee beans, what else would you stockpile?

P: Not sure. We’ll be fine—we’re not buying a car anytime soon.

O: Appliances—swap out broken dishwashers, washers, vacuums (even if made in Europe). Anything else?

P: Plenty, but skip toilet paper—it and kitchen towels are U.S.-made. I’ll check where the wood pulp comes from.

O: Probably Canada. I’ll look into it—it might be tricky.

P: We’re not building a house, so we’re fine—though builders will suffer as input costs spike.

O: Time will tell. I trust American ingenuity—people will work around restrictions. California’s long coast means smuggling is an option. Entrepreneurs will adapt.

P: What a day—over 50 years later, smuggling remains an option. We’re living interesting times. Ready to call it?

Tariffs

M: It’s like the boot has finally hit the ground on tariffs this Wednesday. According to the tariff formula, the purpose is to balance trade. If you ask ChatGPT to crunch the numbers—assuming everything else stays the same—the collected tariffs would exactly wipe out the U.S. trade deficit, or even those of all countries. Of course, there are some very interesting twists in between. I noticed that some small enclaves—because last year they bought a few U.S. aircraft engines with a trade deficit of over three million dollars—ended up with a 99% reciprocal tariff. There’s even an extra tariff on U.S. military overseas bases, and the funniest case: Heard and McDonald Islands adding a 10% tariff. Setting those aside, I think Bitcoin’s bounce back last night might have given everyone a bit more confidence—that it can carve out its own trend instead of being dragged down along with everything else amid global risk-off selling.

N: No, I think it’s some kind of weird short covering. Today—being the 5th day of this sell-off—we’re seeing strange behavior. For example, gold is down 2–3%, and even the yen (traditionally a safe haven) is falling. They should be rallying along with treasuries, right? Instead, both are down and seem to conflict with one another. The yen is down, yet not because of a stronger dollar (since gold is also down), and the USD remains strong. I suspect this is just a liquidity event or forced selling, with self-inflicted pain from forced liquidations. After all, it’s a risky asset—a direct measure of people’s risk appetite. We can all agree we’re in a risk-off environment, right?

M: Yes.

N: So, I don’t think there’s going to be a single day with one specific, large flow that can drive an independent trend.

M: I think that if Bitcoin—still a risky asset—suddenly graduates into its own independent trend this weekend, it could trigger some very interesting consequences. Being a risky asset that’s locked in, it can’t be easily sold off, which might lead to defaults. Since 2022, U.S. banks (including Goldman Sachs) have even started taking Bitcoin as collateral, and we’re now seeing credit-derived loans tied to it. I need to research more on what these loans really mean for the market. But if Bitcoin breaks out on its own this weekend, it won’t be good news for the overall market.

N: Of course—especially for the stock market, that’s obvious. To elaborate, you believe that Bitcoin and crypto are assets even Trump and his family keep a close eye on, right?

M: Yes.

N: If they were to have an independent rally over just a single weekend, it would give them the feeling that they’re invincible—because the asset they’re focused on would keep rising on its own, right?

M: And then, from a political perspective, if people find that they can’t sell this risky Bitcoin asset, they’ll double down on selling off other assets as a cover. When assets can’t be sold, more liquidations follow. We might not know which one will be the next Lehman Brothers, but I worry that a Lehman moment is just waiting to happen.

N: Okay, we don’t even have to go that far. I’m just saying this gives them an invincible feeling. You can clearly see how this negative feedback loop works: they act like they don’t care about anything and can just do whatever they want—as if they don’t give a damn. If this attitude carries into next week, I think some people are already calling it Black Monday—much like October 1987, when the stock market fell around 20% in a day. I mean, it’s not impossible.

M: I think it depends on what happens this weekend. Right now, the “Crypto Bros” are all shouting “Buy the Dip,” while the “Financial Bros” remain a bit reserved.

N: The mood is extremely negative. Since yesterday, they’ve been discussing negotiation tactics—but they’re still holding on to hope, like Vietnam saying its tariffs will drop to zero and the UK starting negotiations with the U.S. If these sentiments hold, Monday could even be a bounce-back day.

M: But I don’t even have to warn you—you’ve seen the TikTok news, right?

N: So?

M: It’s going to extend for another 75 days, which means there won’t be any deal or negotiations with China this weekend—or even in the coming weeks. This uncertainty is just going to drag on for the next 75 days.

N: I think China is a lost cause—we should brace for a price hike of around 30%. Yesterday, many checked with Chinese suppliers, and they’re only willing to budge by about 2%, which isn’t nearly enough to cover a 34% jump. The Chinese government says there will be a stimulus package, but it won’t be huge. Companies will negotiate for the U.S. to shoulder the price increase instead of relying on Chinese stimulus. So, while last week we were unclear about how inflationary these tariffs would be for the U.S. market, this week it’s clearer that most of the impact will likely be absorbed by U.S. consumers, since it’s a universal tariff.

M: So you can’t really shift production to lower-tariff regions to escape this; the maximum relief might be around 10%. Moreover, previous tariffs have already driven supplier prices so low that they can’t afford further cuts—meaning the U.S. has to absorb the cost. This isn’t a new scenario; just yesterday, more Chinese suppliers were discussing it. In fact, there were rumors as early as last week that Walmart was pressuring its Chinese suppliers to lower prices.

N: In a typical scenario, costs might be 20% plus a currency offset of 5–7%, and if suppliers lower their prices by that amount, it might not seem too bad. But this time is different. The dollar index (DXY) year-to-date is probably down 7%, meaning the dollar’s purchasing power is lower while costs remain high. I agree with you—and if everyone faces the same tariff, there’s no reason to negotiate because everyone knows it applies universally. Besides, preparations for this have been going on for a very, very long time.

M: China has been bracing for a cumulative 60% tariff for a long time—and now it’s effectively at 64%, right in the range they anticipated.

N: Yeah, they’re thinking, “Worst case: if we stop trading with you, we’re fine.”

M: Looking at it another way, it’s understandable. U.S. agricultural products have plenty of substitutes—mainly because they’re cheap. Take Utah, for example: it produces lots of farm goods that end up being sold to China rather than consumed locally. For farmers, that’s profitable, but for the local population, it isn’t necessarily a good thing. So aside from farmers taking a hit, it might not be all bad for those regions.

N: I don’t think it’s surprising if China retaliates. If the U.S. only adds, say, a 30% tariff on China, China might not retaliate because over the past few years they’ve shifted production bases to India, Vietnam, and elsewhere. But now—even those regions face high tariffs—making it very hard for Chinese companies to operate. The Chinese government might think, “Since you’re at war with everyone, I’ll just team up with everyone else to take you on.”

M: It’s unlikely they’ll align with Europe or Canada, but they might find more common ground with Asian or some developing countries.

N: I don’t think it’s a long-term alignment, but when negotiating with Trump, he might try to cozy up to Europe for a bit—I think that’s possible.

M: For Europe, let’s be honest: an alignment with the UK, France, or especially Germany is unlikely—Germany’s automotive industry is in direct competition with the U.S. However, countries like Italy, Spain, and Portugal—previously part of China’s Belt and Road—might be more open.

N: Hungary.

M: Hungary was part of the Belt and Road, but due to Biden’s policies, these countries have pulled out.

N: Now they’re rejoining the Belt and Road.

M: That’s not out of the question either, but another boot hasn’t dropped yet—how exactly will tariffs on chips be levied? There’s still little clarity on that. So that’s another issue looming. Even before that happens, I’m already sensing some momentum toward buying Nvidia.

N: This is straightforward. First, companies like Nvidia have solid margins—enough to absorb a shortfall in the worst-case scenario. Second, they enjoy significant pricing power; their products are in high demand and sold mainly to large enterprise customers. In contrast, memory companies like Micron or Microchip deal in commoditized, consumer-facing products with thinner margins and weaker pricing power, and their consumer demand isn’t as robust. So if you have to choose among semiconductor companies based on pricing power and quality, I’d lean toward companies like AVIDIA, AMD, or Avago. But let’s be honest: if a recession hits, the entire semiconductor sector could tank—possibly by as much as 20% in two days. There’s a reason for that.

M: But if these tariffs are here to stay—which seems highly probable—there will be a slow electronics upgrade cycle. All electronics might see a 5–10% price increase (for example, a $799 iPhone could jump to $899). And if the upgrade cycle is slow, more AI computing will shift from edge devices to the cloud. At the end of the day, I believe…

N: I have a question: why, in an upgrade cycle, would computing shift from the edge to the cloud?

M: It’s a slow upgrade cycle—that’s exactly why. The idea of an “AI PC” is becoming obsolete, so people will move more computing to the cloud. And we all know you can’t levy tariffs on bytes. This means some companies—like Oracle and OpenAI, which plan to build data centers in the U.S.—will be at a disadvantage. Meanwhile, data centers in places like Osaka or Malaysia won’t face high equipment tariffs and will be more competitive.

BTC

N: I have a question: what does “American-made Bitcoin” even mean?

M: It basically means that Bitcoin miners in Texas use Texas-grade electricity to mine Bitcoin right here in the U.S., and those bitcoins…

N: Does it have to be Texas?

M: Well, Texas is where most Bitcoin mining operations are located because many believe they’ll get long-term tax benefits. Also, while Texas electricity isn’t exactly cheap, the spot pricing lets miners operate during the cheapest hours—making Texas ideal for Bitcoin mining.

N: I have another question: how do you know if Bitcoin is American-made? Does it come with some kind of stamp?

M: First, everything on the Bitcoin blockchain is public so we can trace which bitcoins were mined by whom. That said, there isn’t a tagging system yet. So if U.S. miners claim their Bitcoin as tax-free, it is tax-free on the chain. However, once Bitcoin crosses to off-chain systems (like mixers such as Tornado) and then comes back, it’s hard to tell if it remains tax-free. For miners, though, it’s quite easy to claim tax benefits in the long term.

N: I have a question: for Bitcoin already in circulation, how can I tell if it’s American-made or qualifies for an exemption?

M: For older bitcoins, it’s harder to tell since you can’t track which mining pool they came from—it might even be mined on someone’s personal computer in 2010. Fortunately, most of those coins are locked up anyway. Newer bitcoins, mined through pools, are much easier to trace, so we can determine a significant portion as American-made.

N: Interesting. But if I have a coin in my Bitcoin wallet right now, can I tell whether it’s American-made?

M: Yes—if a coin was mined recently and can be traced back to a specific miner, you can tell. But as bitcoins go through numerous transactions, even if they started out as American-made, there will come a point when that status is lost and taxation kicks in. Essentially, every Bitcoin can be traced back to its mining block, and if that block was mined using U.S. electricity, then that Bitcoin is considered U.S.-made—but only until it goes through enough transactions to obscure its origin. Then it might be “tainted” and lose its tax-free status. Ultimately, there has to be a cutoff for tax purposes.

N: Yeah. Now, here’s another question: If I mine Bitcoin in Iran where electricity costs almost nothing (which many miners do for the margin), that Bitcoin obviously wouldn’t be American-made by rule. But how does that work in circulation? Does it mean Bitcoin effectively forks from that point?

M: No, it’s not a fork. Bitcoins mined in Iran will simply flow back into circulation. Unless U.S. authorities decide to ban Iranian mining—which might force a fork and lead U.S. (and likely Chinese) miners to exclude them—that’s not happening. So they just become part of the circulating supply and are recognized as non–American-made, subject to export controls. Essentially, even if it’s technically illegal, once it’s in circulation, authorities won’t pursue criminal actions over it.

N: But then you’re effectively creating two different types of circulating Bitcoin, right?

M: No, no, no.

N: If I have the option to buy Bitcoin, I’d naturally choose American-made Bitcoin for the tax benefits.

M: And then?

N: So in circulation, Iranian Bitcoin would likely have a permanent discount relative to American-made Bitcoin.

M: It could be.

N: And that discount effectively creates two separate circulating assets.

M: It might appear as two systems off-chain, but on the Bitcoin blockchain itself, one Bitcoin is still one Bitcoin—you can’t actually exchange one for the other as separate assets.

N: What does that even mean?

M: Off-chain, however, it’s conceivable that one Bitcoin might be valued at $80K while another is at $70K. In that case, you’d need some kind of smart contract to obscure its origin—but I don’t think that’s practical.

N: No, it still confuses me. If you can swap U.S.-mined Bitcoin for Iranian-made Bitcoin on-chain, but off-chain they trade at different discount rates, wouldn’t that create a risk-free arbitrage opportunity? I could just convert my holdings and exploit the difference.

M: That’s why I think it’s not possible. On-chain, the origin is always identifiable—whether it’s Iranian or U.S. Bitcoin.

N: So, as I was saying, you’ve effectively forked the two types—meaning that even on-chain, no one is willing to swap them directly.

M: On-chain, imagine one wallet holds U.S.-mined Bitcoin and another holds Iranian-made Bitcoin. If one wallet sends U.S.-mined Bitcoin, the exchange of value happens off-chain—and that might reflect a different price. Essentially, there is no pure on-chain exchange between different types of Bitcoin. They might have different values on decentralized or centralized exchanges, or when converted to Ethereum, but on-chain one Bitcoin is still one Bitcoin.

N: Now consider a scenario where every transaction is in Bitcoin—for example, I buy something with 2 Bitcoins and you owe me 0.5 Bitcoins in change.

M: No, no—there isn’t an exchange mechanism like “giving change.” It’s digital cash. Using Bitcoin as a medium of exchange creates issues because you wouldn’t normally use an appreciating asset for daily transactions. For example, if you buy a pizza, the store might price it in U.S.-mined Bitcoin so that you only need one coin, but if you use non–U.S.-mined Bitcoin, you might need two coins for the same pizza. This doesn’t mean there are two separate circulation systems; it just creates two different price points for the same item—underscoring why Bitcoin is a poor medium of exchange.

N: So basically, you’re creating two different kinds of assets.

M: Yes.

N: So we start with one asset, but it ends up splitting into two distinct assets.

M: Exactly—it's not that you value one Bitcoin differently from another, but rather…

N: Okay, so if I’m a U.S. citizen, I’d definitely want to buy U.S.-mined Bitcoin because of the tax benefits.

M: Because you’re a patriot.

N: Anyway, what’s the implication? If everyone mines Bitcoin in the U.S., does that mean we’ll run out of electricity?

M: No, no, no. First, people have some very twisted logic about this. They claim that mining Bitcoin means we need to build more power plants—and that by doing so, Bitcoin can absorb the excess power, which in turn makes electricity cheaper in the U.S. It’s a very convoluted rationale, but if you assume building a power plant is free, it kind of makes sense.

More Tariffs

M: In the past couple of days, more news about tariffs has emerged—including reports of internal conflicts. For example, Elon and Bessent don’t agree with the proposal currently being announced.

N: Bessent is questionable, but Elon has already distanced himself.

M: However, Lutnick, Paul Navarro, and Trump himself seem to be the driving forces behind this framework. Even Sacks hasn’t spoken up in a long time.

N: Does staying silent now mean disapproval?

M: Uh, I think they’re still strategizing—figuring out how to get involved in the discussion without alienating their billionaire friends. At the end of the day, this is part of a “Great Reset.” Next week, some people might even start jumping off buildings, and as small and medium businesses go bankrupt, there will be…

N: Let’s take a step back. Your read is that if these tariffs remain in place for just a week or two, it could cause many industry supply chains to collapse—leading us into a massive crisis.

M: Right. We’ve already discussed that—in negotiations between suppliers and importers, suppliers refuse to absorb these tariffs. That leaves price hikes as the only option, which will undoubtedly be inflationary.

N: In the short term, it acts like a consumption tax. Initially, it’s inflationary, but over time it will depress demand—pushing us toward a recession, or worse.

M: Then you end up with winners and losers. For instance, I think the niche industry of building AI data centers in the U.S. will take a hit—especially for new companies like OpenAI and XAI, which have already invested billions in U.S. data centers.

N: Here’s the deal: whose deal did OpenAI broker—Larry Ellison? And who’s behind XAI—Elon? Even if Elon gets pushed out of Washington for disagreeing with the trade deal, a $300 million donation still carries weight, and being a lifelong friend of Trump counts for something. Those two will probably be fine. Even if it isn’t a tariff reduction or exemption, they can likely secure government support. I’m just speculating, but it looks like this new regime is effectively picking winners and losers—they’ve got a trump card, to some extent.

M: I think it’s tricky because taking government subsidies isn’t popular among conservatives. That’s why I feel if these tariffs stick—even for three weeks—they’re likely here to stay. In that case, many small and medium businesses will go bankrupt, and I doubt the government will support them.

N: The closest analogy is COVID—a massive supply chain disruption. But the key difference is that during COVID, central banks and governments responded super fast with monetary and policy measures: extended loans, grants for small businesses, and near-zero interest rates. However, according to Powell’s comments on Friday, the central bank isn’t planning to cut rates now.

M: They’ll only cut rates once a major bankruptcy forces their hand—but that’s just cause and effect. It’s not an overnight decision.

N: I mean, unless a big bank goes bankrupt, the Fed will likely just sit it out. Small business bankruptcies aren’t really the Fed’s problem. If a major bank collapses or the repo market tightens drastically, then the central bank would need to step in. Otherwise, I don’t think the Fed will get involved.

M: For the short term, I agree—it’s unlikely that building AI data centers in the U.S. will receive subsidies or preferential treatment if tariffs rise by, say, 20%. For many small companies like Lambda Lab, CoreWeave, or Nebius, it’s much easier to build data centers outside the U.S.

N: I think most companies will simply build their data centers in Canada. Why not?

M: Everyone except for OpenAI and XAI (who have already invested heavily) and Facebook—which has the disadvantage because it hasn’t built many data centers outside the U.S.—will likely opt to build abroad. Other companies are better positioned because they already have a strong multinational presence.

N: You mean Google?

M: Yes—Google, AWS, Microsoft, and other cloud service providers are well positioned to shift orders from the U.S. to countries like Malaysia, Japan, or even parts of Europe and Canada. After all, you can’t levy import taxes on bytes and bits.

N: I think a 20% hike—or even close to 30%—is too steep. The tariff on Taiwan is around 27%, right? That’s too much for end consumers to fully absorb. Either the supply chain or the ODMs, who are already operating on razor-thin margins (around 7–10% gross margin, with most at 7%), would have to bear the brunt. It might end up being split between companies like Nvidia and the end user.

M: No, I think it’s hard to imagine a scenario where that cost is passed on to the retail side for computing.

N: No, I mean enterprise.

M: It’s tricky—CSPs can absorb a lot. For example, AWS charges about $10 per hour for an H100 per card, which is huge compared to the nominal price of $1.5 to $2 per H100. They have absurd profit margins and can absorb the cost. However, enterprise users, who are very sensitive to price increases (since it affects all their projections), might see prices rise more slowly than consumer products. This could push us toward a recession due to demand elasticity. That’s why I’m planning to stock up on coffee beans tomorrow—maybe a year’s supply. There is no supply of 5090s now, so I might have to take one for the team. And yes, an iPhone might end up costing an extra $100, which is acceptable.

N: But did you see the estimates? Some analyses suggest iPhones could be about $300 more expensive under the new tariff regime. If Apple only passes $100 on to the consumer, then they’d have to absorb $200 themselves—or split it with their suppliers.

M: It really depends. I might even cancel my annual upgrade contract with Apple and switch to an outright cash purchase—because here in the U.S., tariffs apply, whereas they don’t elsewhere. I’d rather skirt the tariff (almost like tax evasion) than pay an extra $300. Realistically, I think we’ll end up paying an extra $50 to $100 on average. However, those stuck in AT&T contracts or unable to pay full price for an iPhone will suffer. People with more financial flexibility can work around it, which might become the norm here in the U.S.—I experienced similar tactics with small items when I was in China.

N: Yeah, I think consumer electronics components are in even worse shape than, say, AI servers. But at the end of the day, if we head into a recession and demand drops, everything will be in trouble.

M: Do you want to talk about Lama 4? It’s set to release tomorrow.

N: Oh, okay. What are the expectations?

M: Lama 4 is going to be based on a MOE (Mixture of Experts) architecture. For the initial models, don’t expect them to be exceptional—that’s the consensus. They’re just the first step; reasoning model will come later. Since it’s a MOE model, don’t expect to be able to host it on your own. It’s clear that most of these advancements are headed to the cloud. Tariffs only worsen the situation—discouraging electronic upgrades and leaving insufficient compute power. In short, models are getting larger and will be hosted in the cloud, which raises issues of privacy and security. People are willing to compromise on those issues simply so they can use the technology. For edge devices, tariffs and the slow pace of efficiency improvements mean they can’t keep up. We still haven't reached AGI yet, so people are not going to reduce the size of the model or make it dumber in order to fit on a device. People have to make the model smarter and smarter and smarter. Even if that means it has to use more compute, more memory, and the model itself is going to be bigger, with more training data—everything just has to scale up. The only complication is that AGI is delayed due to the tariff, which is the biggest open question.

CRWV

L: CoreWeave.

K: CoreWeave is in a very tough spot. The market hit a short-term bottom in mid-March, rebounded for almost two weeks, and then lost all its gains in just two days—a brutal ride. For private companies, especially those underperforming SaaS firms that raised in 2021, if they go public, I expect their valuations to go down. It is what it is.

L: Do you think Nvidia being the anchor investor is a huge red flag?

K: I think people are being overly pessimistic and critical of CoreWeave. It’s essentially a rental business—it doesn’t add as much value as established cloud service providers (CSPs). But if you look at AWS when it went public in 2012, it was the same basic model. Value-added services don’t materialize overnight; they take years to build. CoreWeave is capital intensive, just like all CSPs. I’m not saying I’ll invest in CoreWeave, but there’s a lot of misunderstanding and excessive criticism about the company and its business model. Plus, Anidia’s involvement is painted very negatively, which is a bit unfair. As a provider, Nvidia benefits from a diverse customer base, which in turn weakens the negotiating power of giants like GCP, Azure, or AWS. It’s a simple strategy.

L: But back to CoreWeave—here’s an interesting point about U.S. companies: they need to hit the market at the right time. In countries like India or China, time-to-market isn’t as critical as the ability to raise capital. For instance, if you look at AWS when it went public in the early 2010s, it was essentially an infrastructure service company—exactly what CoreWeave offers. They don’t let you manage the infrastructure; they handle it for you, yet they don’t provide much else. Back in 2010 or 2012, cloud providers (CSPs) weren’t clear on what extra services to offer atop basic infrastructure. Take Google App Engine—it turned out to be a huge failure. At that point, CSPs were just beginning to transition from pure infrastructure to value-added services, and the window for building that out was closing. This shows that U.S. capital is impatient. For example, replicating Amazon’s model of warehouse management and delivery is relatively easy—Alibaba and JD managed to do it in China years after Amazon. Yet, in the U.S., Shopify had to cut back quickly because venture capital isn’t as patient with such ventures. In short, if CoreWeave wants to evolve from just providing GPU compute to adding value-added services, it might get punished by both an impatient market and talent.

K: I don’t disagree, but let me add two points. First, CoreWeave’s business model is similar to Amazon’s early AWS—we all agree on that. U.S. investors are impatient, but they do love a compelling story. I don’t know much about CoreWeave’s founders or management, but if they can craft a strong narrative, they might gain some leeway. Remember early AWS? Their pitch was, “We already have the infrastructure; we’re just selling spare capacity,” essentially treating capex as free. CoreWeave’s story might be tougher to spin, but with the right angle—and with falling interest rates lowering refinancing costs—it could work. More importantly, think about OpenAI. In the long run, OpenAI won’t rely on traditional cloud providers—they’ll build their own data centers and might even become a CSP themselves. My wild speculation is that, one day, OpenAI might even buy CoreWeave.

L: For that to work, it depends on the company. A consumer product company generally enjoys much higher margins and can afford to delay full-stack integration while focusing on expanding its market. However, if a company remains an API provider in a low-margin, high-volume business, it must compete aggressively on price. In that scenario, owning the entire stack to deliver the cheapest API becomes an advantage. Nvidia’s situation is intriguing—it challenges the common belief that manufacturers can’t capture most of the value. In the computer industry, we’re used to manufacturers’ margins being squeezed (think of Intel sacrificing margins to partner with software giants like Microsoft, or Apple only regaining strong margins after its mobile turnaround). Nvidia’s rumored acquisition of Lepton.ai hints at a strategy similar to the airline industry, where a few players (like Airbus and Boeing) capture most of the value while the airlines themselves, being commoditized, compete on minor details such as customer service.

K: It all comes down to customer experience. For many services, the experience is nearly identical between providers. For example, when I fly—even if the experience is subpar—a four-hour flight is just that, and most Americans fly fewer than ten times a year. Similarly, in AI, I’ve experimented with different providers—Perplexity, OpenAI (and skipped Grok because I’m not a fan of Elon Musk)—and it’s clear that after a few iterations, OpenAI’s refined responses are far superior. So for now, there’s enough differentiation among providers. But will that hold when we eventually reach AGI or ASI? Possibly not.

L: We don’t know. As we approach the singularity, things become unpredictable. That said, Nvidia is placing numerous side bets. Unlike AMD, Nvidia not only manufactures chips but also publishes research on model training and inference tools, and they’re even expanding into cloud services by offering toolkits to help bootstrap infrastructure. Much like how airlines operate, these side bets are expensive—they require hiring top researchers. It’s a long-term strategy. Anyway, shall we talk about tariffs?

Tariffs

K: Yeah, sort of. I can update you on the news, but there’s not much more I can add.

L: Do you think the tariff situation will worsen as we approach April 2nd, or will it improve?

K: April 2nd is coming up—next Wednesday. I’m not sure, but it seems 50/50. Some believe it will be the worst day, as tariff clarity might trigger tough negotiations that eventually improve conditions. Meanwhile, headlines like “Trump doesn’t care” suggest he wants to raise tariffs—since his first term’s tariffs were too mild, and now he plans to generate $600 billion annually to “balance the deficit.” Honestly, I’m in Camp 1—I think things will start to improve after April 2nd, though I worry that the consensus might be off.

L: So, is that the consensus? Many believe tariffs will worsen consumer costs, but the financial view is that clarity is coming—and we’ve had two months to prepare.

K: I think finance types might be overly optimistic—or maybe they’re just rationalizing their support for Trump. The financial consensus is that peak uncertainty will hit on April 2nd, after which clarity will lead to negotiations. Expect some back-and-forth: Trump may say outrageous things, then his team will calm the market. After April 2nd, there might be minor bumps, followed by poor economic data in May and June, with June potentially being the bottom when the Fed can finally cut rates. Trump might even backtrack slightly, claiming victory by promising deregulation and tax cuts. So, while a modest bounce might occur, most expect another 7–10% drop. Honestly, no one really knows.

L: I agree. None of us are tariff experts in this generation, and we really don’t know the U.S. pain threshold—especially with all the media manipulation. People might endure a lot more pain before breaking, or maybe it’s even beneficial because a recession forces a balanced budget.

K: I disagree with the idea that the government would intentionally trigger a recession or mini-recession. Even if Trump and his entourage seem reckless, no one wants to spark a recession—it would wreck their midterm prospects and blow up the deficit. Even if his brand isn’t strong, they wouldn’t want to “let the genie out of the bottle.” I remain somewhat optimistic—I’ve hedged my bets—but intentionally triggering a recession just doesn’t make sense.

L: The window for triggering a recession is shrinking. If they intended to do it, it would have to happen within the first six months—and we’re already in April, leaving only about two months.

K: That’s probably why many predict a bottom around June or July. If they can’t execute their plan, things will be painful—especially since midterms are about a year away from the recession’s end (if there even is a clear end). You can force a recession, but you can’t control its duration. I think they’re very cautious about intentionally triggering a recession. Just consider the Great Financial Crisis—it lasted years.

L: Two years to recover.

K: Actually, it took even longer. Unemployment remained very high for at least four years. Not every recession is like 2020, with both supply and demand shocks that allow for an immediate bounce back.

L: In today’s hyper-political climate, it’s hard to pin down a single issue. The lesson from the Great Financial Crisis is that high unemployment can be politically devastating. During COVID, efforts to avoid an unemployment spike ended up causing inflation—which, in turn, influences elections. People have the memory of a goldfish: they only recall the most recent election loss, not events from 10 or 15 years ago.

K: Oh, okay. Essentially, all his policies will end up hurting both inflation and unemployment.

L: We don't know. Inflation happens when the government starts printing money.

K: No, it’s when import prices rise.

L: And if people can’t afford it, prices will eventually drop.

K: Oh, God. And then, regardless of the situation, corporate margins will be crushed.

L: And that’s why unemployment spikes.

K: Yes, and let’s be honest—over the past two years, white-collar and skilled worker unemployment has been rising, partly due to AI. Government support and service providers have kept things in balance, but if spending stops, even service providers will face layoffs. Ultimately, unemployment will rise, but no one really knows the timeline. People just try to pin dates to events to make sense of them. One thing I’ve noticed lately is a shift in tone among even the more moderate Trump supporters—those “All-in podcast bros” are now downplaying the so-called “Trump tariff prisoners.” They used to insist that whether you loved or hated Trump, you had to give him credit for adapting to circumstances. But come on—the 80-year-old white guy who’s been touting tariffs for 40 years isn’t exactly known for adapting. They act as if they know everything, almost godlike, yet now even the elite supporters seem to be backing off. Who knows what that means for his base?

Apple Intelligence

L: Apple Intelligence is facing significant headwinds, and there are rumors it might acquire Thinking Machines Lab.

K: Oh my God, that’s crazy—didn’t they just get founded about a month ago?

L: Yes, but I don’t think it’s far-reaching. If Steve Jobs were still at Apple, they might have already acquired Anthropic. I’m not sure how they’d handle the shares held by Amazon and Google, but Jobs had a knack for persuading them. Internally, Apple failed this project, and it’s hard to see a recovery from that. Let’s not forget: Apple is a massive company with deep pockets. Yet if you only allocate, say, $2 million a year for external research, those hires often underperform—that’s just how bureaucracies work. The only escape is to make a big acquisition, forming an independent team that isn’t hampered by internal politics—allowing Apple to shell out, say, $100 billion without serious scrutiny.

K: Right, for sure.

L: I think if it is Jobs, that’s already happened. It’s hard to swallow that Apple would have to acquire a company for $100 billion, only for someone to later offer a $30 billion deal for an all-star team. That’s an attractive proposition—an independent, 20–30 person team that doesn’t need to fully integrate with Apple’s existing structure. There’s some credibility to this rumor, and I believe Apple needs to make a bold acquisition to fix its Apple Intelligence issues.

K: But are those people really excited about leaving OpenAI to join Apple?

L: I don’t know. At one billion dollars per person, everyone on the team is a billionaire. It’s hard to say.

K: Yeah, if you put it that way, I’d definitely be reluctant.

L: Exactly. If Apple lets them work independently with the sole mission of making Apple Intelligence work, it could succeed. In today’s world, a billion dollars goes a long way.

K: But how much more time do you think they have? Apple Intelligence is delayed until 2027, right?

L: Well, Apple isn’t facing technical constraints—the technology exists—they’ve essentially boxed themselves in by insisting on running models on the device, which limits both compute and memory, and by restricting their cloud usage. They’ve also limited the data available for training. However, an independent team could work around these issues and fix Apple’s problems, making the deal very attractive.

K: I’m not sure. The privacy issue is a major roadblock for AI—especially for Apple Intelligence. How can they overcome that?

L: I believe they can solve it. The issue isn’t that models must run on the device; it’s about deciding which tasks run locally versus in the cloud, and how to pass context between them. These are solvable challenges if you have a strong research and engineering team. Apple doesn’t need to completely overhaul its privacy-preserving design—just assemble the right team and let them work without excessive internal interference.

4o

L: Even if Apple Intelligence is currently failing, there’s still a lot of exciting progress elsewhere. For example, OpenAI’s new 4o image generation model has been making waves online recently.

K: Yeah, it’s a very impressive product.

L: Yeah, definitely. It’s interesting how these trends reemerge every few years. I remember style transfer around 2016, then filters on TikTok and Snapchat, and now people are rediscovering similar techniques with OpenAI’s work. It’s both weird and fascinating—and it’s proving useful for creating slides, infographics, and charts directly with image generation instead of relying on language models.

K: For me, the most intriguing part is what comes next. If this is all it does—flooding my Twitter timeline with headlines for two days—it’s not very interesting. But if someone follows up, there’s a huge market opportunity, much like Adobe’s.

L: I think the market ultimately goes to whoever has the best models. But models will eventually be commoditized. We’ve seen this with large language models—once you have a good model, the downstream applications (even prompt engineering) become secondary, and the model providers become indispensable. This industry is fascinating because consolidation occurs even as commoditization spreads. Ultimately, it’s hard to see where the remaining value is, aside from chip makers like Nvidia. That’s why they make money and why investors are anxious—they’re unsure who will capture most of the value as every claim to a “perfect model” is quickly undercut.

K: I don’t want to rehash the Nvidia debate again—we’ve done that many times. There’s nothing new to add beyond the usual noise and FUD (fear, uncertainty, and doubt). So who will lose? It seems clear that existing SaaS companies—basically doing rudimentary “prompt engineering” without leveraging LLMs—will be the losers. What exactly are these SaaS companies doing, anyway?

L: I think their moat lies in business relationships and relationship management—think paid dinners, conferences, and kickbacks that keep customers loyal. These advantages are enduring and unlikely to be disrupted by AI or LLMs.

K: I actually disagree strongly. We have to be clear about the time horizon. In the short term—one to three years—relationship-driven businesses might hold up. But over five or ten years, the trend changes. From my experience in asset management, active management has gradually underperformed indexes because, in essence, it offers the same value, driven by relationships and kickbacks. Over a decade, indexes capture far more market share, squeezing fees and margins on active management. Similarly, existing SaaS companies may enjoy short-term advantages, but if their products aren’t compelling, their business will decline. Margins and sales will shrink, and they won’t be able to attract top talent.

L: I agree in the long term, but SaaS companies have a window of two to three years to go public. Everything is compressed—they need a recession and an IPO window for early investors to exit. It’s fascinating to watch what happens next.

K: Notice how those who were once very cautious about CapEx ROI have suddenly become supportive—Trump acted quickly, and many were so afraid he’d wreck the stock market that they missed their chance to exit. This ties back to CoreWeave: despite its hype as the new AI cloud, it had to cut its valuation in half and run flat. That’s similar to what happened with many SaaS companies—they ended up staying private. Maybe that’s why figures like Chamath are suddenly saying, “Oh, Trump is adaptive!” You wish.

L: Maybe. Alright, let’s wrap it up here.

Recap

J: We're back after two weeks. Let's discuss the market crash over the past two weeks.

I: What? A market crash? It’s more like identifying a short‐term bottom, I think.

J: I suspect the market might again be delayed for April 2nd. So, I think…

I: Think about it—April 2nd could go either way. It might spark a rally or trigger another downturn. It’s a 50-50 situation. My point is that if the policy becomes clear, businesses can finally analyze and plan, which would remove uncertainty and potentially fuel a rally. On the other hand, if we see more flip-flops and poor communication, it could lead to further declines. Essentially, this is a pure variance event with no inherent directional bias.

J: So, do you believe the tariff will be fully carried out?

I: It depends on who you ask. Some are optimistic, others aren’t. Overall, I’d say that both net and risk exposures are extremely low—everyone I know is fully hedged and holding plenty of cash right now.

J: Right, and I believe that by April 2nd everyone—including importers and exporters—will be better prepared than they were at the beginning of February. Just as we discussed, I think the short-term and long-term impacts of tariffs aren’t as black and white as everyone assumes. We no longer have “tariff experts”; our expertise now lies in importing, exporting, and international marketing. Thus, the real impact of tariffs on global market adjustments isn’t entirely clear. We’ve already had two months to prepare for this adjustment, so the initial short-term impact of tariffs won’t be as terrifying as people imagine.

J: That’s why I believe it’s clear—they’re definitely adding tariffs, targeting the top 15 trading partners of the United States. The aim is to balance trade and boost government revenue, paving the way for tax cuts later this year.

I: Exactly—but once the policy details are released, businesses can start planning. Right now, everyone’s in freeze mode, which isn’t ideal. That’s why I believe that April 2nd could trigger a rally just as easily as it could spark another downturn.

J: I think the Fed provided quite dovish guidance on rates for this year.

I: Well, yeah, but honestly, I wouldn’t call it particularly dovish. It’s more of a calm tone—especially compared to the chaos on the government side—not exactly eager to act. One friend put it best: if any government official or federal agency doesn’t give a damn about what Trump says, it is the Fed.

J: Well, the Federal Reserve isn’t a government agency, but its chair’s term ends in May 2026, and his successor will likely be very loyal to the current administration. I don’t think market factored that in yet. The current administration seems determined to undermine the US dollar’s global reserve status—and even the Federal Reserve itself—because even with a chairman appointed by them, the Fed isn’t closely aligned with the executive branch. Clearly, they want to exert more influence over the Fed and the dollar’s reserve status. It’s fascinating to watch. So, who do you think will be the new Fed Chairman?

I: I’m not going that far yet. I wouldn’t claim with certainty that they want to abolish the dollar’s reserve status. Look, my take on Scott Bessent is that he’s taken a free fall over the past three months, but he’s not that stupid. He talks about using stablecoins to bolster the US dollar’s reserve status—that’s just nonsense. And I think they already have someone in mind: Kevin Warsh, the former Morgan Stanley banker who married into the Estée Lauder family. He’s emerging as the front-runner for Fed Chairman. He’s been a consistent critic of previous Fed policies—a real hawk aligned with both Scott Bessent and this administration. I’m curious to see how it plays out—after all, if you’re a hawkish leader, you’d typically strengthen the dollar, but Scott seems to want it to weaken. We’ll just have to wait and see.

J: I think the policy aims to bring manufacturing back to the US—essentially turning the US into a pure export country like China. But then ask yourself: if you export goods, you end up importing other currencies. Do you really want China holding a pile of US dollars to buy US treasuries? And do we want to hold Chinese Yuan bonds and treasuries? It doesn’t add up. That’s why there’s talk of returning to gold as the reserve currency for international trade. And if that fails, Bitcoin is on the table as an option. But again…

I: Oh my God, don’t even get me started on Bitcoin. The whole concept is utterly ridiculous to anyone with even a basic understanding of finance. You seriously want to trade the US dollar’s reserve status for something you have zero control over?

J: The US and China are the two main governmental Bitcoin holders, with China holding roughly 1% more Bitcoin than the US. Both governments can sell their Bitcoin without restrictions. That makes Bitcoin interesting—if your biggest adversary also holds a significant amount, international trade and markets could get messy. Bitcoin—like Tether—has proven its longevity over the past decade despite the scams that often plague financial products. Longevity is the ultimate credential in finance (remember, Madoff operated for 30 years). In a way, this administration is also trying to prove itself. I’m cautiously optimistic about Bitcoin’s potential as a reserve currency if the current administration truly commits, despite their other priorities.

GTC

J: Alright, let’s change the subject to GTC. One interesting point is that GTC is taking place in the Bay Area—and so is GDC—even though the mainstream media completely overlooks GDC. Instead, they focus on NVIDIA’s GTC, which, although small a few years ago, now seems to overshadow everything. And Jensen—he did an impressive job in that two-hour presentation. For an internal presentation, it was excellent: no teleprompter needed, and he even called out all the names of external partners. I’m most impressed by that. While this works well internally at NVIDIA, as a public presentation it invites comparisons with legendary salesmanship—think Jobs or even the more colorful Steve Ballmer. Nowadays, presentations by Elon, Lisa Su, Zuckerberg, and Jensen just don’t capture that same spark, and it seems many founders have lost that essential sales skill.

I: It’s not so much about skill—it’s about charm.

J: Of course, Jensen isn’t like most founders. He’s been in this game for 20—no, three decades now. I even said 10 years ago that he’s the best CEO, not just because he’s the founder and leader of the most valuable company in the US, but also because, think about it: when you compare him to Zuckerberg—would Zuck wakes up every day: “I’m going to make tons of ads and targeting vulnerable people and making money from those people, by letting them making wrong purchase choices, that's make me so pumped”—and Bezos: “I’m going to reply 100 emails with emojis every morning so that those employees get on their toes, and I feel so empowered and pumped to work”—Jensen, on the other hand, wakes up every day ready to create the best, most powerful computer chips. He’s truly living his dream. NVIDIA is far ahead of the competition and, evidently, making a lot of money, which clearly makes him very happy.

I: Well, the market seems to disagree.

J: I think they’re in a tough spot. We’ve mentioned that the beating won’t stop until the moral improves. And he even provided a roadmap—initially two years’ worth—which is…

I: Never heard of that, okay? Really, never heard of it.

J: It’s extremely rare.

I: He didn’t give a two-year roadmap—he gave a four-year roadmap.

J: There are many things he might not even know how to execute properly, but the roadmap is undeniably long. Usually, at conferences, they outline a two-year plan, so it’s rare to hear a CEO publicly discuss a four-year roadmap. That’s intriguing. I don’t think the market will buy it—the sell-off might continue until expectations are met quarter by quarter or until a major event occurs (such as a relaxation of export restrictions to China). I do believe Nvidia chips will face tariffs between 10% and 25% by April 2nd, meaning everyone will pay a bit more for gaming cards or GB200 series products. How this will affect data centers—say, whether Lambda Lab might set up more in Japan rather than the US—remains to be seen. But the government has been clear: expect 10% to 25% tariffs on chips, especially on imports from Taiwan.

I: Yes, but on the other hand, it seems there’s some give and take—they’re not fully enforcing the chip diffusion rule.

J: Right. So, two forces are at work here. On the plus side, I actually believe export controls on China will be relaxed.

I: Why? What’s the evidence?

J: No evidence.

I: Then why? I’m trying to figure out whether this is an educated guess or just wishful thinking—because hope doesn’t drive investments; facts do.

J: Well, this administration is vindictive. The export control on NVIDIA chips is a Biden-era policy, and basically, anything from that era—good or bad—will be rolled back before any new measures are introduced. That’s the angle I’m playing. I don’t think people like Marc Anderson or any VCs really have a say here; it’s all about the vindictive nature of the current administration and Biden’s policies.

I: Who—almost—are the policy advisors? Who drafted these policies, and who are the technical advisors behind them?

J: I’m not sure—I’d guess it’s the usual big-name suspects.

I: Interesting. It would be huge if they ended up not implementing the chip AI diffusion rule—or even if they relaxed export controls on China.

J: Yeah, and as we discussed, the only worry about this administration is whether they might be racist—but their actions suggest they’re not. Take the TikTok rules, for example: they’re basically saying that TikTok just needs to implement the proposed Project Texas from two years ago without having to sell to Oracle. They only need to maintain a stake and let an intermediary handle part of the data center. That’s Project Texas in a nutshell. If they’re so lenient with TikTok—a Chinese-owned company—it essentially signals a rollback of Biden-era policies. That’s a positive sign for me; it shows they aren’t prejudiced against Chinese or Asian businessmen.

I: But don’t you realize your view is highly non-consensus? Every sell-side analyst I speak with is pricing an 80% chance that export controls will tighten—not loosen.

J: And that’s not the only issue. I think next year ASML will also export EUV machines to China.

I: Just because the US has pissed off Europeans as much as China?

J: Not only that—these restrictions don’t make sense because China will soon have its own EUV machines. Besides, the US lacks the appetite to buy ASML’s most profitable EUV systems. Intel is struggling and isn’t purchasing new equipment anytime soon, although TSMC is planning some for US expansion.

I: They’re in the midst of a capex contraction.

J: Exactly. ASML will need to find alternative buyers for its equipment, and China is an obvious candidate—it’s expected to have domestic EUV production by the end of next year. They need to sell while they can, and if ASML starts selling early next year, it might at least slow China’s progress with EUV. That’s good for ASML and helps maintain its technical edge over China.

I: I’m not as optimistic. I think the US could simply say, “If you sell, we’ll impose sanctions and kick you out of our payment system.”

J: I don’t think it’s a big deal for the US government. Their focus is on domestic manufacturing—not on controlling every business transaction. US government actions are pretty straightforward these days; if they’re unhappy, they tend to use direct force (as seen with the Houthis in Lebanon recently). It’s highly unlikely that ASML’s actions would trigger such a response since it won’t affect US manufacturing, which is their main concern. And if anything, remember—the Chinese are exceptionally good at bribery. That’s another positive from this perspective. If the voting base isn’t fully aware of the nuances of third-party business dealings, they can be placated. That’s how I see the current administration’s actions.

I: Following the same logic, I can see a scenario where the export ban on Nvidia chips to China might be relaxed—or even lifted.

J: Exactly. It’s not going to affect Chinese or American jobs. Moreover, if Nvidia ramps up domestic chip manufacturing—as Jensen promised at GTC—and ends up exporting chips to China, that could even benefit the US. It all comes down to the “vibe” policy making. Now, let’s talk about Tesla.

TSLA

I: Yeah, I don’t even know what to say about this stock—it’s pretty pathetic right now.

J: I think the headline here is that you never see a company go bankrupt solely because of a boycott.

I: Yeah, for sure. Think about it—Toyota was boycotted heavily in China in the early 2000s. I still remember news reports of people trashing cars and even burning them in front of the Japanese embassy. And yet, Toyota ended up selling extremely well in China, capturing a large market share until domestic EVs emerged. Have you ever heard of any company being boycotted to death? I don’t think so.

J: A company only goes bankrupt by mismanaging its finances—spending too much, failing to generate or raise funds, and enduring prolonged negative cash flow until it dries up. It’s never a boycott; it’s all about the numbers. The crux is that Elon can raise money for nearly anything and is a master of financial manipulation who never pays the price for it. It’s unwise to bet against him right now.

I: So you’re saying you’d bet against him? I don’t want to do that—remember SolarCity, Funding Secured, and those Saudi investors lining up, begging Elon for money even before he was connected with the sitting US president. You’re right; Elon never has trouble raising funds or shuffling his assets. That’s the bottom line: Tesla might not go bankrupt, but does that mean the stock will skyrocket?

J: I don’t think so either. It’s a tricky situation. SolarCity never really benefited common shareholders, and if Tesla takes that route, Elon might profit as a trader while common shareholders could get wiped out. Tesla is essentially a meme stock—hard to bet for or against because you never know when sentiment will shift. There are plenty of Tesla fanboys, but their numbers seem to be shrinking; I’ve even seen regular investors divest from Tesla. I doubt any major Tesla breakthrough will occur in the next two years, which is concerning given that this administration will be around for another four years. One thing’s clear: Tesla’s Robotaxi and humanoid projects won’t materialize in two years—maybe five at best. Technology advances rapidly, especially in robotics. Look at companies like Boston Dynamics and Unitree. Unitree might be early winners, but they still haven’t crossed the “early adopter valley”—that deep gap between initial enthusiasts and mainstream acceptance. I’ve owned Unitree’s first product, the Go1 quadruped, and even their G1 Humanoid still shows early hardware quirks. It’s notably shorter—around five feet—making its dynamics simpler than Tesla’s Optimus or Boston Dynamics’ Atlas (which are closer to a normal adult male height of about six feet). While being shorter helps with movement, it also limits utility for factory or household tasks. Unitree still needs to cross that valley. I remain cautiously optimistic about Unitree, especially if Korean investors continue to back them. Never underestimate the Koreans—Samsung may not always have a blockbuster hit, yet they’re still number two globally.

I: Let’s get back to Tesla. I want the stock to work—it was our closest bet to become the next Apple. I don’t own it anymore (except for a few souvenir shares), but I want to see it succeed. In my view, Tesla is a stock that overshoots sentiment on the upside and undershoots on the downside, making it a potentially very profitable trading play. The problem is, I can’t seem to find a catalyst on a reasonable horizon. For example, at last night’s all-hands meeting, Elon showed numbers of 5,000 robots this year and 50,000 next year. The energy was there, but anyone with half a brain knows that the humanoid is still a product searching for a market.

J: I think it’s not that the product isn’t promising—it just takes time to grow the market. Like any product, it needs time, and not just two years, since no product will hit the market in two years. The first sale might occur as early as next year (which is the most optimistic scenario), but then the market will take time to develop. Remember, it took nearly a decade for the first Tesla Roadster to lead to the mass-market success of the Model 3 and Y. Even the iPhone, which was relatively simple, took about four years from launch to runaway success. Unlike Nvidia—where the catalyst is clear (lifting export controls on China could spark a rally)—it’s hard to pinpoint what will drive a Tesla rally. Even if their robotaxi service launches in Austin this year and expands twice as fast as Waymo, covering 20 cities by next year would require an extra 100,000 vehicles, which is just a fraction of Tesla's manufacturing capacity. Still, these business lines remain in an incubation phase, and it’s unclear when they’ll truly take off.

I: Also, it’s going to be a cash drain on their income statement and balance sheet. They’re not just selling cars and booking revenue—they’re stuck with global taxes and hefty SG&A expenses. And if they don’t partner with companies like Uber or Lyft, they’re on their own, which means significant investment.

J: But that could actually be a silver lining. First, Elon is never short on capital—he can always raise money. Second, although these robotaxis are on the balance sheet, they’re also assets that can be “managed” to show higher value. As Elon mentioned, it’s all about FSD. With FSD, these assets might appreciate rather than depreciate, giving him room to inflate their value. Granted, the numbers might not be huge compared to their current holdings of roughly 100,000 vehicles (the leasing program on a similar scale), but consider that the more expensive Model Ys might book 30,000–35,000 in value after two years, while the cheaper Model 2 vehicles might only value about 20,000—numbers they won’t show because Elon insists it doesn’t make sense to sell them. So, the hope is that these vehicles, thanks to FSD, will become appreciating assets and boost the balance sheet.

I: Ultimately, no one invests in Tesla purely based on the numbers—the metrics never really made sense (except maybe briefly between late 2019 and early 2020). It seems that the negative news flow needs to subside, or at least the stock shouldn’t overreact to it, so that it can begin to respond to positive developments. In any case, it’s just something I’m keeping on my watch list.

META & GOOG

J: So, you want to talk about Facebook and Google? What’s the reason?

I: I mean, if you look at the charts shared by both retail and institutional investors, they suggest that U.S. big tech valuations aren’t as outrageous as some claim.

J: Right.

I: Especially after that 20% drop, people started saying that the best picks among the so-called “Meta 7” are Amazon, Meta, and even Google—at least according to some. In my view, however, the consensus leans toward Amazon and Meta (roughly 5 out of 6), while others like Apple seem to be considered as having negligible growth. Tesla—well, we just talked about it—and Nvidia are seen as purely cyclical plays peaking at the top of their cycle. As for Google, I’d put it in the same category; investors are actively buying Amazon, Meta, and Google on the dip. I’ve even heard that after the rally in Chinese ADRs and Hong Kong stocks, some Chinese funds are selling off stocks like Baba and Tencent to reallocate into Meta, Amazon, and Google. Honestly, I’ve mulled over all three, but it’s tough for me to commit to any at this point.

J: Yeah, we’ve talked at length about Amazon’s leadership issues. For Meta, there’s also the question of their direction with the Llama model—nobody really knows what to make of it, which only adds to the stock’s visibility problems, even though it’s been climbing steadily. As for Google, it’s the most intriguing of the three. It’s been a money-printing machine for ages, but now it’s facing the classic innovator’s dilemma. Google set out to organize the world’s information, and now, with large language models emerging as the most competent organizers, they’re trying to catch up to the innovators. Internally, they’re banking on the “Google Gemini” answer feature—when you search, it gives you a direct answer instead of a list of links, which they believe will boost revenue by increasing click-throughs. I’m not sold on that narrative, but that’s the hill they’re willing to die on; they’re not backing down and are determined to win the battle for organizing the world’s information.

I: To wrap up for me, my personal use of Google has probably dropped by about 90%. Nowadays, whenever I have a question, I simply open ChatGPT—whether it’s the mobile app or the web version—and type in my query, because it gives me clear, well-formatted answers. Even when I do use Google out of habit, I rarely click beyond the top “Gemini answer” and ignore the usual 10 blue links. Granted, I’m not your average consumer—I’m more of an early adopter—but if, say, 10% of the population cut their Google usage by 50%, that would roughly translate into a 5% revenue deceleration for them.

J: That drop probably hasn’t shown up on their earnings report yet—and you might never actually see it reflected there. They have plenty of ways to ensure that revenue doesn’t fall, such as emphasizing sponsored links over organic clicks. They manage this carefully because it’s one of the strategies to overcome the innovator’s dilemma. With their “money printing” machine, they’re statistically positioned to be outliers, giving them the flexibility to navigate these challenges.

I: Do you think that information aggregators that heavily depend on Google for lead generation—say, Reddit—will have their traffic secured?

J: No, I don’t think so. I believe you’re not really talking about Reddit but rather about a company like Pinterest that relies heavily on Google for its traffic. Pinterest—earning most of its revenue through that channel—is the most vulnerable. However, companies like Airbnb and Reddit aren’t as dependent.

I: The reason I mentioned Reddit is that they noted this week that a recent change in Google’s algorithm has affected their revenue, so they’re adapting accordingly.

J: An algorithm change is bound to affect the revenues of companies like Stack Overflow, Reddit, Pinterest, and hundreds of others that rely on SEO for traffic—Reddit included.

I: I know. I’m just trying to figure out if this is a structural issue or merely a temporary blip. If it’s just a blip, then it’s not a big deal.

J: I think these companies will have to adapt to a world where SEO isn’t as valuable as it is today. In my view, Reddit isn’t the most vulnerable—Pinterest is. You rarely see people visiting Pinterest directly; they usually get there through Google’s SEO magic. Meanwhile, companies like Airbnb seem to be in a stronger position. Even if they suffer from reduced SEO traffic, their strong brand recognition will cushion the blow. Ultimately, companies that fail are the ones that mismanage their finances. If you can anticipate the drop and manage your finances well, strong brands like Airbnb should fare better than competitors like Vrbo.

I: Well, Booking.com also has strong brand recognition. I get what you’re saying—this is a reconfiguration, and there will be winners and losers. Some companies are better positioned to come out on top, while others aren’t. Understood.

TSM

H: TSMC, I agree with everyone’s opinion. This deal represents the best outcome for TSMC in terms of negotiation. First of all, not having to partner with Intel is a good thing—Intel’s involvement would be a problem. Let’s be honest: TSMC essentially bought Intel’s fab, but only by getting a discount on equipment prices. Otherwise, it wouldn’t make sense. They claim Intel's 18A is equivalent to 2nm, but it’s an entirely different process. There's no way TSMC could run that process efficiently. Moreover, there’s another benefit: even though it’s still a pushmark, through this deal, TSMC can secure a full exemption from all tariffs. Considering the trips he made, I believe he must have reached some kind of soft agreement with the Trump administration—including a commitment to invest three years’ worth of CAPEX in building a factory in the US (despite TSMC’s annual CAPEX being only about 38–40 billion).

G: I think that doesn’t demonstrate any agreement on Taiwan’s security. In my opinion, Taiwan’s security isn’t even on the negotiation table. And even if it were, following the US’s global retrenchment strategy, if China’s domestic economy continues to struggle by 2027, an attack on Taiwan might become possible. Conversely, if China’s economy is doing well and is well-integrated globally, there’d be no need to attack—if they did, it would mean Xi Jinping lost his mind. On the bright side for TSMC, this deal is a very bullish sign for me. I was especially worried about one thing in the administration: racism. although we know Trump isn’t a racist (he’s just an actor from Hollywood, it’s hard to be a racist there), this administration’s Project 2025 includes many racist individuals. I was concerned about whether bribery from Taiwanese would even work without them being looked down upon for having the wrong skin color. One particularly chilling thought: among the top ten US billionaires, Jensen is the only Asian. The rest are all white.

H: Speaking of which, is Jensen still among the top 10 billionaires?

G: You’d have to check again. Being at the top has its drawbacks. I think if TSMC manages to pull off this bribery successfully—and assuming Supermicro also passed—it’s a very bullish sign. It even hints at whether NVIDIA’s export controls will relax if the bribery is just right.

H: That is lobbying.

G: Okay, it's not really lobbying—lobbying means you pay your congresspeople and senators. This is direct bribery. But anyway, that’s beside the point; it’s just how the executive branch and business synergy operate in this administration. It might surprise everyone, but with the right bribery and lobbying, this administration could have a change of heart—from enforcing strict export controls on NVIDIA to relaxing them. I believe that would be a permanent bullish turning point for NVIDIA, especially since even in its bearish downward spiral, its profits remain strong.

H: I mean, you do hear a lot of chatter about that, right? People say we shouldn’t focus on China; we should concentrate on running our own race at 100 or 200 percent, which I agree with. You can’t just try to limit your competitors—you have to compete at your best. That’s it.

G: All that chatter comes from some VCs. Their vested interest is that if China re-enters the game, they’ll have more attractive investment opportunities—especially the ones we discussed last time. Investing in developing models isn’t very appetizing for VCs because it’s like throwing tens or hundreds of millions into a void without knowing if your model will succeed, or if the application layer will benefit. Yet, you can’t simply stop investing in new models because they unlock new capabilities in the application layer; without them, there’d be no new applications at all. It’s a very tricky position—investing huge sums in models while knowing the money might never yield a return. But if China re-enters the picture, it’s a different story. They’ll have plenty more investment opportunities; they can invest in China or leverage Chinese capital to invest in models. And nobody talks about it. Even now, there’s a lot of Chinese capital in U.S. VCs as limited partners, but no one discusses it because you’re not supposed to. VCs remain very receptive and respectful toward China. That’s why you hear them advocating for lifting the sanctions on NVIDIA. Ultimately, from my perspective—and we have huge holdings in NVIDIA, so take that with a grain of salt—I think we should lift the sanctions on NVIDIA chips, as these sanctions are becoming more absurd over time (especially with the 5090D and all the 50-series debacles). Continuing these export controls and tariffs makes little sense; lifting them would benefit competition and ultimately undercut Chinese competition. I also believe that Western investors fail to realize that China has a degree of market economy. Although even after lifting sanctions, China can continue investing domestically, the support it receives with sanctions in place is completely different from that with sanctions lifted.

H: Yeah, sure. I think they understand that. Everyone is saying that if you rely on them, they’ll use government funding and subsidies to build the entire domestic semiconductor ecosystem—from fabs upward, I believe that's right.

G: What I'm saying is that even without sanctions, the government will still pour money into domestic semiconductor production and leading-edge chips. But if you lift the sanctions, the market will boom because foreign companies competing for talent in the domestic market will muddy the effectiveness of government investment—and that’s good for foreign enterprises. At this point, I think the sanctions on application chips aren’t very useful. We might still keep the sanctions on 7nm fab equipment, because those days are numbered. China is expected to have its EUV equipment ready between the end of this year and early next year, but actual production is likely projected for around 2027. So whether you lift export controls or not, it probably won’t affect that timeline much.

AVGO

G: I don't quite understand why everyone is fixated on the 2027 number from AVGO while NVIDIA is dealing with 2025 and 2026 figures.

H: For the most part, I think people don’t buy NVIDIA’s sustained CAPEX story. They see NVIDIA as a hardware company and wonder how a hardware company can consistently demand a 75% gross margin—it just doesn’t add up. The moment they see a margin drop, even if it’s due to efforts to ramp up Blackwell, they immediately think “those days are numbered.” Competition is fierce—other companies can design their own ASICs, and the compute landscape is shifting toward inference, which doesn’t require a Ferrari-level product; you can run inference on something more like a Honda. (To be clear: “Ferrari” here refers to NVIDIA, while “Honda” refers to an ASIC like Grok or similar.) That’s why people question whether NVIDIA’s peak earnings and margins can be sustained. The market is essentially saying no, even though NVIDIA’s VMware-like recurring revenue and connectivity business give it some leeway. Plus, its AI ASIC segment is growing rapidly from a low base, with the idea that it will eventually become a $70 billion market capturing 50% of it—that’s the rhetoric for 2027, even if I don’t fully buy into it.

G: I think the fundamental issue is that we’re dealing with two types of companies. One type takes on custom contracts, while the other designs and builds products based on a multi-year roadmap. For public investors, this creates a duration mismatch. For companies that secure customer contracts—especially in ASIC design—the time horizon is about two to three years, from contract initiation to full execution. For example, Broadcom already has visibility on its expected revenue in 2027 from contracts signed in 2025. But for product companies, whose competition is other product companies, public investors have no clear idea what the final product will be or how the market will react. It could be a huge success or a massive failure. That duration mismatch is just too hard to swallow.

H: I talk to investors regularly, and the mismatch I see is that people still view NVIDIA as a hardware company—much like the debates about Apple in 2016. I’m not saying NVIDIA’s model is identical to Apple’s (not yet, anyway), but back then critics argued that Apple, as a hardware company, should have been valued similarly to Samsung—trading at mid-teens PE instead of mid-20s—and many predicted Apple would eventually drop to those levels. The same dynamic is playing out now. People view AVGO’s revenue as recurring and nearly immune, almost like a SaaS model in hardware, while others insist, “Show me the product, and I’ll evaluate you year over year.” Then I have to hedge against the risk that the next product could see a drastic drop in demand—a huge left-tail risk where the product fails and inventory piles up. That’s the biggest mismatch between these approaches.

G: I agree—running a product company is incredibly challenging. Look at Tesla: they had a few good products, then spent five years without a new release before dropping the Cybertruck. We have no insight into their product development process—why the Model Y didn’t get an incremental facelift, then suddenly did—making global sales drops hard to explain. It’s likely due to retooling for the Model Y combined with the general dislike for Elon affecting sales. We won’t have a definitive answer until March or April. By then, we should know if it’s Elon’s influence or just temporary Model Y issues. In any case, running a product company is as tough as it gets—just consider Apple over the past 15 years. They must ship a new iPhone every year with significant changes, and despite staring into the abyss, they manage to sell through clever marketing, revised payment plans, and new initiatives. Apple uses many tricks to keep the upgrade cycle going, and Tim Cook doesn’t get enough credit for that. On the Android side, the leading company is Samsung thanks to a broad product lineup ensuring smooth sales cycles. Yet, each cycle features a different leading product. In contrast, Tesla has dropped the ball—announcing no releases is easy, but maintaining a consistent product cadence is extremely hard. That cadence, along with recurring revenue, ultimately determines whether a product company succeeds or fails. Look at NVIDIA’s 50 series chips—they have many issues, yet they had to release them after two and a half years because delays aren’t an option.

H: I agree. Think of it like fast fashion—most hardware companies are judged product by product. If you have a huge hit, it’s almost destined to fade because the odds of back-to-back hits are slim. Typically, investors wait during a bust, then when the next hit comes, the profits are enormous. I believe what you’re saying—and I implicitly agree—is that founder-led product companies (like Apple and NVIDIA, where the founder is deeply involved in product design) are an exception. They might have minor missteps, but they rarely recover because they’re too bogged down in managing every detail, unlike leaders who’re detached from day-to-day operations.

G: Tim Cook isn’t the founder, after all. There’s a spectrum among hardware companies. For example, Samsung is a hardware company but isn’t seen as a product company in the same way. The same goes for Micron and even AMD—which is more hardware-oriented—while NVIDIA and Apple (and even Tesla) are firmly in the product company camp. AVGO, on the other hand, leans more toward hardware. So, the observations you made apply more to product companies than to hardware companies in general. Many hardware companies don’t release consumer-facing products; they manufacture components for others to integrate and sell, which makes them harder to evaluate as standalone entities.

H: But I mean, we’ll see how this cycle turns out for NVIDIA. However, one thing I’m still trying to wrap my head around is that it’s not so much about NVIDIA—it’s more about the broader macro backdrop.

Scott Bessent

G: I think that's what we need to discuss right now. We've talked enough about semiconductors—let's move on to macroeconomics and talk about Scott Bessent. He’s been appearing on TV frequently these days.

H: I mean, I don’t even know what his motivation is at this point. During Trump 1.0, whenever Trump announced some crazy policy, Steve Mnuchin would go on TV to calm the market. But Scott Bessent? Oh my gosh, he does the exact opposite—trying to spook the market even more with all his talk about a detox. And then he says things like, “the stock market doesn’t really matter; you can be up 20% for two years in a row and still get voted out.” What is he aiming for? Besides that, he looks absolutely terrible on TV—he comes off as an idiot, even a retard, and he speaks so slowly. If you’re going to spout all that contradictory nonsense, at least do it fast, for crying out loud.

G: That's what I was thinking. He probably hired a public speaking coach recently, which explains his odd TV presence and slow speech—he’s trying to apply the lessons from his coach.

H: I listened to his podcast right before he was nominated for Secretary of the Treasury, and he sounded completely different—smart, seasoned, educated, a true veteran of the financial markets. But on TV, he comes off as if he’s saying, “I’m so good, you have no idea what you’re in for,” and, “I don’t care about you.”

G: I think Bessent is seriously underrated. He provides a rare, unfiltered glimpse into what this administration is really thinking—no sugarcoating—and that raw honesty unsettles the market because it's just the truth. In contrast, figures like Marco Rubio or J.D. Vince blend market-friendly sound bites with their own spin and the administration’s messaging, making their speeches seem palatable until you spot the glaring gap between their words and what actually happens—leaving you to wonder, "What's really going on?" But I suspect Bessent’s time in the administration is numbered; he’s downright terrible on TV. Meanwhile, Rubio might stick around longer—especially after his clash with Elon—because reality TV loves a good spectacle, and the president seems to favor that kind of show. And when it comes to the recession, you're missing a key point: those who hit their prime during the Clinton years—mostly white, male, and supremely confident—treat a downturn like a minor hiccup. Elon, Bessent, and Trump all come from that era (though Elon, as a younger immigrant, operates on an entirely different level), and they genuinely believe they can bounce back from any downturn. That said, I have this superstitious belief—there’s a sort of mystical force at work, a collective consciousness, or even a God-like push—that favors certain narratives while disliking others. You know the kind: a tale of a hero fighting a villain or slaying a greedy, evil dragon, where the hero always wins, no matter what. Yet when it comes to the other kind—an overconfident man predicting the future—that never pans out. Look at Alan Greenspan’s endless boom prophecy, which ended with the Great Recession; we’ve seen it time and again. There are forces—strange, inexplicable forces—that ensure narratives of overconfidence in predicting the future, as appealing as they are, simply don’t come true. I totally understand their perspective: if we somehow navigate this detox phase, get the budget deficit under control, slash Social Security and Medicaid, and let some of the...

H: Trump said he's not going to cut all of that.

G: It's going to happen—you'll see a wave of boomers pass away, and the economy will improve as the age imbalance is corrected, with more young people ready to work (even if they aren’t college educated and are primed for manufacturing jobs). It might seem like everything will magically work out, but I believe that if a narrative rests on one man confidently predicting the future, it just won’t pan out. It’s not rational; it’s superstitious.

H: I just don't trust any of those fuckers—I think they're going to screw up big time. Do I believe the situation is as dire as they claim? Probably not, because I’m not that bearish. Everyone I talk to seems overly pessimistic, expecting the market to tank, and non-US investors are even more bearish. They compare the current US fear to China’s 2022 lockdowns and irrational policies. But honestly, if you go out to dine, travel by plane, or visit places like Hawaii (which we just did), you don’t see any signs of a consumer slowdown—and that’s what really matters for the economy.

G: Hawaii isn’t a fair comparison—tourist numbers there haven’t recovered since the Maui fire. You can have a booming economy while the stock market remains depressed. Look at China after the 2015 crash: its GDP grew at 6–7% per year, even though the stock market lagged. The US is more complex, but what if the stock market stays down without the economy falling into recession?

H: It's unlikely.

G: Unlikely, but what if it does happen?

H: No, you have to consider the composition of wealth. In China, 70–80% of wealth is in real estate, which is why after 2021—when domestic developers went bust and property prices tanked—China entered a massive deflationary period where the wealth effect disappeared and spending stalled. In the US, over 70% of family wealth is in equities, making the situations quite similar. My point is, if the stock market crashes, consumer confidence will suffer, triggering a deflationary cycle that stifles economic growth. You can’t engineer a prolonged stock market drawdown in the US without it impacting consumer confidence and the real economy.

G: I think that by simply scaring public investors, they might trigger the recession they’re aiming for.

H: Dude, that’s why I’m not feeling that bearish. First off, we’re at a 4.25% Fed funds rate—so by slightly hurting consumer confidence and introducing recession talk, you can push down inflation (since inflation is largely about expectations and confidence), giving the Fed room to cut rates. Secondly, oil prices have dropped significantly, saving consumers money. And third, without rehashing all the data (I’m not a macro investor), the overall signs suggest I’m not overly bearish. I do believe those old white motherfuckers will eventually screw up big time—but not right now.

G: I think you're painting a rosier picture than even they hope for. You’re saying that we are going to…

H: I mean, at least in the short term.

G: We’re going to have a pretty painless soft landing, with inflation, interest rates, and even the 10-year treasury yield coming down. That’s why they want to engineer a recession—if they can trigger one without an actual recession, why even aim for it? Then the economy would start growing again. I think that’s a rosier picture than what they were even hoping for. Let’s end on that positive note.

H: Let me add one thing before we wrap up on that positive note. I don’t think their actions will cause much short-term damage, but they are seriously undermining the US dollar’s status as the world’s reserve currency—which will have long-term implications.

G: I don’t even think they want the US dollar to remain a long-term reserve currency, considering they’re setting up swap agreements between USD, USDT, and USDC.

H: That's the fucking dumbest thing, right?

G: That swap agreement essentially gives them the right to print as much USD as they want. There’s no regulatory oversight for USDT and USDC—USDC is self-regulated domestically, and USDT is self-regulated internationally, meaning there’s basically no regulation at all. So, if they establish a swap agreement between USD, USDT, and USDC, I see it as even more bullish than spending tax dollars to buy Bitcoin for a reserve. It basically means they can print unlimited USDT and USDC, and if that agreement comes to fruition as rumored, it shows they don’t really care about maintaining the USD’s reserve status. It’s simply not their concern. Maintaining a global currency isn’t a central priority for this administration.

H: Okay, let's end on that note.

NASDAQ

E: NVIDIA earnings? I don't even want to discuss it. It's such a non-event—pretty much in line with expectations. Sure, they missed on networking and gaming but exceeded expectations on Blackwell, even though they fell short on their own gross profit margin guidance. It was all well anticipated, so I don't understand why there's a big reaction. It's just a very boring, transitional earnings report. The market seems to be driven more by overall degrowth. For example, some are now circulating a narrative that Trump and Scott Bassnett are trying to force shock therapy—first triggering a quick recession to push down the long-end yield curve, then lowering interest rates, and finally sparking a long bull equity market. Honestly, I think they're just forcing a narrative on Trump and his administration—I doubt they even have a clear plan.

F: I believe there are some brilliant minds in his administration, but Trump himself remains an unknown quantity. We still lack clear visibility into this entire administration. Although Elon tweets daily, it’s hard to discern what’s really going on. From my perspective, tariffs on Canada and Mexico should have been imposed by February 1. As America’s second and third largest trading partners—accounting for roughly one trillion dollars in total U.S. imports—the plan under “Project 2025” is to transform the U.S. from an income-tax-dependent nation into one that relies on tariffs. With annual income tax revenues of around three trillion dollars, they’d eventually need to slap 100% tariffs on these top partners to offset roughly 1.5 trillion in tax revenue. That would drop the effective tax rate to levels reminiscent of the 1920s, the Gilded Age. But since they didn’t act in February, it seems they’re worried about the stock market’s reaction. Now, in March, we’re still confused about what’s happening inside the administration. There’s no coherent view on what the high-level planners behind Project 2025 are doing—that’s the biggest uncertainty. You just can’t model this administration’s behavior.

E: Well, it’s not just you—everyone feels this uncertainty. That’s why you see the market being sold off day by day. In this environment, I’m definitely not ditching my hedges, and it’s exhausting. Look, momentum stocks—the wild, retail-heavy ones—have tumbled, though they might catch a brief recovery. But it’s far from over. The past few weeks have seen extreme degrowth in TMT sectors and extremely bearish sentiment. Even though the NASDAQ is only down 8% from its high as of yesterday’s close, that shows just how leveraged and frenetic retail has been. Every rally seems to get soaked—a pretty bearish sign. Today, I even dusted off my October–December 2018 charts to remind myself how volatile things can get in a short period. I’m not saying we’re exactly repeating that playbook, but unless we see a clear pivot from the current administration or the Fed, I doubt we’ll see any renewed “animal spirit” in equities anytime soon.

F: I doubt the Fed will pivot given the current CPI and Core PCE numbers, don’t you think?

E: Yeah, that’s fair. Honestly, I don’t think they’re strictly data-dependent. Powell seems obsessed with engineering a soft landing. I believe the key is that if he sees the economic and macro indicators declining—and then if we get one or two weak job reports—he won’t worry about inflation; instead, he’ll have room to cut rates. That’s just my take, though I could be wrong.

F: I'm not sure. The Fed’s mandate includes managing inflation, so it can’t just let the economy boom indefinitely. I also think this administration is keen on engineering a recession. From what I understood before the February pullback on Canadian and Mexican tariffs, it seemed they were trying to trigger a recession as quickly as possible so that by midterms they could present a rosy picture of the U.S. This aligns with what Elon Musk has said—that you’ll feel some pain initially, and then things will get much better. It hints that the U.S. might pull even deeper into isolationism, meaning it won’t burn money overseas but will instead impose tariffs, essentially triggering a shock therapy recession domestically. After the tariffs, everything would eventually readjust, and the economy could start growing from the bottom up with inflation falling. But then again, the worst-case scenario might not just be a recession—it could be a depression. I’m not sure.

E: Anyway, regarding investments or trading, right now I’m only holding the stocks I have the highest conviction in. I’m also raising cash whenever there’s a mini rally, and I’m keeping about 50% of my cash uninvested for now.

F: And that’s crucial, because if you look back at 2018, you had to make money while the market was down for an extended period—almost a year.

E: No, no. 2018 was relatively painless overall. Sure, the drawdown was brutal, but it only lasted about three months—and then the market bounced back quickly after the Fed pivot in January 2019, which resumed the rally.

F: Given that, I think a Fed pivot is possible, especially since Powell’s term ends in January 2026. Worst-case scenario, we have just over a year left, and the rally in 2019 ended after Yellen’s term, correct?

E: Yes, after Yellen’s term ended and one year into Powell’s term—when Powell was raising rates, the market freaked out, there was a tariff trade war with China, and hedge funds experienced degrowth—I don’t think the market has actually priced in a recession. It just appears to be a typical hedge fund degrowth in the TMT sector, maybe with a bit in healthcare. For instance, banks, financials, and payment companies like Visa and MasterCard keep rising daily. In a true recession, I wouldn’t expect them to continuously rise.

F: But those stocks are supposed to reflect a recession. In an actual recession, they would fall rather than continuously rise—they aren’t precursors to a recession.

E: Well, the stock market trades on expectations. If a recession were expected, you’d at least see banks being sold off heavily. At least for now, it doesn’t seem to be reflected in prices.

F: Maybe they should.

E: There was chatter this week suggesting that maybe we should buy the long end of the bond market—essentially, TLT—as a recession play. Some even claimed that during Trump’s first term, his key performance indicator was the S&P 500: whenever it dropped 1% or 2%, he would intervene—tweeting, calming the market, or pressuring Powell to cut rates. Now, they’re saying that for a potential second term, his KPI might shift to TLT, with the goal of lowering the long end of the yield curve and, once again, pressuring Powell to cut rates—so he can claim victory in reducing the fiscal deficit and normalizing the equity market. Just some chatter, really. It seems people are trying to push the narrative that Trump is engineering a recession, but I don’t think that narrative has been fully priced into the market yet.

F: True, that chatter has been around for a while now. We need more proof to be certain, but all we know is that the next three to six months are going to be turbulent. Everyone should feel comfortable taking some short positions in the meantime.

AI

F: By the way, Anthropic and OpenAI both released new models this week.

E: Yeah, but here’s a minor point—can’t they come up with better names? Why does it have to be 3.7 and 4.5? It sounds more like a patch update than a major model release.

F: I believe the reason GPT‑4.5 exists is that it doesn’t quite meet the expectations for GPT‑5—either it falls short or they haven’t allocated enough resources to reach that next level. The word on the street is that each GPT version requires over ten times more compute and training data than the previous one. For instance, GPT‑4 was trained on roughly 20 trillion tokens, so by that logic, GPT‑5 would need 200 trillion tokens—a huge leap. Realistically, GPT‑4.5 was probably trained on no more than 100 trillion tokens, using around 50,000 CUDA cards. It might be one of their better experiments in scaling toward GPT‑5, and they felt it was good enough to release. But what’s more interesting is Anthropic’s model—everyone seems quite pleased with the 3.7 version.

E: What’s their trick? Why are they so good at coding but not as proficient in other areas? Did they discover something special?

F: I think the initial feedback on Anthropic’s 3.7 is that it’s quite good overall. It’s not just excellent at coding—it’s also strong at text classification, processing tasks, and many real-life applications. It’s like having a highly competent secretary who can manage your Excel spreadsheets, prepare documents, perform translations, and handle various tasks. While some haven’t probed its knowledge base extensively, it appears that both Grok‑3 and GPT‑4.5 still have more world knowledge than Claude. Their training data composition is certainly interesting, though we don’t know all the details yet. Moreover, they’re much better at generating SVGs and vector graphics using just the language model. They must be using some form of synthetic data for that, and they seem to excel in post-training techniques. They likely have unique methods for synthesizing data in a way that’s more digestible for large language models—a common challenge—and possibly a more refined data filtering pipeline and evaluation metrics than what’s publicly known. These improvements essentially create a moat that’s hard for others to replicate over time.

E: Interesting. By the way, do you think Anthropic is currently training on Amazon Trainium, the ASIC card?

F: I’m not sure; I don’t think so. For researchers, such migrations would be a distraction in the race to AGI. The key is to have a pipeline that works and can scale—you don’t want your task list bogged down with moving from one system to another or other busywork, as that would hurt your research quality and output, which is crucial nowadays.

E: So… don’t they have minions to handle that? I mean, I know they probably have a team of star researchers, but they might also have a bunch of B, C, D, E, or F players assigned to this kind of migration work.

F: It’s hard to imagine a top research firm employing B, C, D, E, or F players. If that were the case, it would indicate bigger issues than just migration tasks—it speaks volumes about their internal struggle.

E: But how many people do they have right now?

F: They have just over a thousand employees.

E: That's a lot.

F: However, fewer than 200 are actually working on large language model research; most of the rest are focused on product-related tasks.

E: Yeah, okay. I’ve heard conflicting information about Trainium and Amazon’s ASIC efforts. On one hand, if you look at TSMC allocations, it’s clear that Marvell—Amazon’s partner for designing Trainium 2—has been aggressively ramping up capacity. I’ve also heard that a Taiwanese AI chip company was chosen as their partner for Trainium 3, and now there’s talk of Trainium 4. It seems they have a long-term, ambitious roadmap for this ASIC, suggesting they expect significant ROI or utility. They even stated it’s intended not just for inference, but for large-scale cluster training. On the other hand, I’ve seen data indicating that AWS recently cut Trainium ASIC pricing by 20%—though that might be less relevant. It’s perplexing; aside from a few insiders who have access to or use the chip, I haven’t heard much. So what’s the plan? Why have such an ambitious roadmap if the current version isn’t a runaway success?

F: In chip design, you need a long-term roadmap because the process from design to production—and having the software ready—is a very long pipeline. That’s why TPU v1 and TPU v2 didn’t achieve runaway success until they became competitive with Nvidia and found their niche in TPU v3. Essentially, it’s easier to design tolerances at both the chip and software levels for TPUs. Smaller firms with access to TPUs prefer using them for training large numbers rather than relying on Nvidia GPUs, which require a full-time engineer to babysit when training with thousands of GPUs. GPUs can overheat and need checkpoint restarts, whereas TPUs handle failover seamlessly. Managing a few thousand TPUs is much easier than managing the same number of GPUs. On a smaller scale—say, around a hundred GPUs—things run smoothly with PyTorch, but if you need 10K GPUs, it’s nearly impossible to secure them since Google uses its TPUs for its own training. That leaves Nvidia as the only alternative. Regarding Amazon, a long-term roadmap is a great sign; it indicates continued investment in their own chips, which may eventually pay off. As for the recent price cut, I don’t see it as a bearish signal—their current pricing isn’t competitive anyway. For example, their L40, a lower-end Nvidia GPU for data centers, is priced at $3 per hour—the same as the H100 on Lambda Labs, which has already doubled in price since the DeepSeek R1 hype (from around $1.7–$1.8 to $3.4).

E: But I think there’s a key distinction here: Are they not competitive in pricing because they choose to prioritize margins, or because they simply lack capacity? Maybe they literally cannot offer a more competitive price due to supply and demand dynamics. All I’m saying is—was the price cut for Trainium meant to boost competitiveness and gain traction, or is it because there isn’t enough demand?

F: For H100 pricing, I think it’s a mix of both—a chip shortage and a desire to maintain a healthy margin on their investment. They don’t intend to use the H100 as a tool to attract new cloud customers; it’s simply easier for their existing customers to onboard GPUs when offered. Moreover, once onboard, those customers are less likely to switch providers, allowing them to charge a premium on the H100. Additionally, we know that Jensen is trying to balance all cloud providers—ensuring that not only the big three get all the GPUs but also smaller players, forcing them to justify their demand by showing how their customers generate revenue. He’s doing everything he can to avoid a “Cisco moment,” wanting to know exactly who is using his GPUs and whether they can use them sustainably to generate returns.

E: I never imagined back in the day that Amazon wouldn’t be very competitive on margins. I guess we’re entering a new era, wouldn’t you say?

F: Yeah, I think U.S. companies have become complacent regarding competition, so they aren’t actively competing on margins. That’s a pretty bearish sign for the U.S. economy overall, because if you’re not pushing on margins, you’re not innovating. You end up like GE during the Jack Welch era—focused on margin and profitability to the point of stifling innovation, and kill the golden goose.

VC

E: I don’t know—speaking of that, one piece of news caught my attention recently: Perplexity is launching a $50 million seed investment VC fund. I find that very bizarre.

F: I think it’s fine—it’s not exactly unique. OpenAI has its own fund with convoluted terms; it’s structured to look like an official OpenAI fund, but it’s really Sam’s personal fund for startups. Similarly, MidJourney’s David Holz has his own fund under the MidJourney name to invest in other startups. It’s a way to cultivate their ecosystem, and Perplexity is doing the same. It really depends on how the fund is structured and how investors feel about it.

E: I mean, if I were a VC investor, I’d feel ripped off if I invested in a company that then used my money to set up another VC fund. Essentially, it’s like they’re subleasing my capital. Why would I be okay with that if I trust my own ability to generate great returns?

F: I think for VCs, if an investment is considered successful, they give you a lot of leeway because they don’t want to antagonize you, which helps secure future investments in you. They also spend more time trying to recoup investments from failed startups than squeezing extra millions from the successful ones. In short, I don’t think they really care about that—that’s why they can get away with doing things like this.

E: I think they should start to… I’ve noticed that the entire VC industry seems to be undergoing a structural shift—although less visibly than the software engineering world, where job postings spiked during COVID in 2020 and are now unwinding amid frequent layoffs. Meanwhile, it seems that old-school VCs are turning into “boomer VCs.” Do you agree?

F: I think it’s a very interesting topic. From my perspective, during the internet era you never really knew who would succeed—many new VCs took chances and became super successful. In the so-called golden years of VC around 2010, there was a minor revolution that was easier to track and justify, as most investments were in companies valued around $2 billion or similar mobile internet firms.

E: Yeah, that’s a different case—think of mobile internet companies like Instagram or Snapchat, or B2B SaaS companies where simple formulas (like how fast you can double your revenue) make the case.

F: Exactly. For established VCs, that era was golden because they wielded enormous power to determine success or failure—especially with SaaS companies—allowing them to invest and profit handsomely. Now, we’re back in a more wild-west-like internet era where there’s no consensus on a winning formula for the next billions. Consequently, many smaller VC funds are making interest investments, and non-VC funds—who aren’t as focused on immediate profits—are making big investments in major players like Thinking Machines Lab, SSI, and even Lambda Labs. These big players don’t want the strings attached that come with VC money; they prefer cheap, no-strings-attached capital from non-VC sources. This puts established VCs in an awkward position, which is why their investments tend to be smaller. A notable exception is A16Z, but even they often make unconventional investments, operating more like YC with significantly more capital—or at least, that’s the image they project.

E: The current state of the VC industry reminds me of the evolution of public markets. Initially, you had old-school value investors using traditional metrics to buy stocks. Then came 2000 and, after 2008, a wave of mobile and SaaS companies led to a decade of stellar performance—investors looked for “compounders” with high growth potential. Essentially, these investors were like value investors but focused on metrics such as revenue growth acceleration. Then in 2022, a correction squeezed the slack out of the system, and growth-at-a-reasonable-price investors didn’t recover from the drawdown. Now, for the past 2 years, it’s been a wild west—with AI dominating everything from GPUs and ASICs to cloud providers, utility companies, and specialty equipment firms. You’re either riding the AI wave or you’re not—you're screwed. It seems that only those investors who aren’t tied to one winning formula and who keep an open mind—focusing on making money rather than being thought leaders—have survived. Meanwhile, many VC firms in Silicon Valley, especially the “boomer VCs,” are more obsessed with projecting thought leadership (appearing on podcasts, writing multi-page Twitter threads) than actually making money.

F: I think you’re misunderstanding what VCs do. They’re not just investors—they’re promoters. They promote themselves to attract capital from their LPs and also promote the startups (especially the toughest ones) so that these companies are willing to accept their money. Essentially, they’re more about promotion than traditional investing.

E: But what value do they actually generate? If they’re just promoters, as you say, they don’t really add value for their investors or for the companies they invest in.

F: Investors are passive, while VCs are active—they scout for deals and ensure they get a seat at the table. For them, investing isn’t just about putting money into companies; it’s a privilege to participate in these opportunities.

E: Sure, but if that’s the case, why do they spend so much time promoting themselves instead of actively hustling?

F: Because they need to promote themselves to attract capital from passive investors (their LPs) and to secure a spot in the rounds of the hottest startups in the Valley.

E: But based on what you’re saying, it feels like the middle player is getting squeezed pretty hard these days.

F: Yes, that’s what’s happening. In the AI space—which is still a bit uncertain—if you want to be a successful VC in the next three to four years, you can’t just be a good promoter (as might work in SaaS); you need to think differently. You must invest based on a unique philosophy to secure truly interesting startups, not just the hottest ones. Alternatively, you can have a lot of capital to throw around with no strings attached to get into those rounds, because the hottest startups today care more about the strings attached to their funding than anything else. That’s why they get money from firms like Fidelity and Vanguard—entities with vast amounts of capital from retirement funds that aren’t under pressure for active management. That kind of capital is ideal for startups, and if given the chance, they’ll take it.

E: Or even corporate VCs—like Nvidia throwing in a billion here or there—why would they care? It’s all about synergy, right?

BABA

C: BABA

D: Regarding BABA’s performance, I don’t find it that bullish. In a bear market, these results would likely drop the stock 5%–10%. First, the improvement in their CMR comes from raising the take rate—not from genuine GMV growth—which means it’s essentially “milking” existing merchants instead of expanding the base. Second, when you look at their cloud business (specifically the grocery segment), even using Baidu Cloud as a benchmark, it’s performing even worse. Moreover, part of that cloud revenue comes from internal use, so the overall numbers aren’t that impressive. On top of that, BABA has historically relied on share buybacks. Now, with CAPEX already hitting 200 billion, they admit that buybacks can’t continue as before—which ultimately hurts shareholder value. It’s a different setup, and although some interpret this earnings report as super bullish, it’s really all about animal spirits right now.

C: But don’t you think that Alibaba Cloud is an entirely different ball game compared to American counterparts like AWS, GCP, or Azure?

D: Why? They have even lower margins.

C: Because in the US, you have the concept of “neocloud” – anyone can spin off a new cloud (think Nebius, Lambda Lab—by the way, Lambda Lab recently raised funds—and CoreWeave, among others). In China, however, spinning off a neocloud is far more challenging due to regulatory burdens. In the US, if you want to rent a cloud instance, you just sign up on AWS. In China, you must first complete real-name verification—sending your ID details to Tencent Cloud, Huawei Cloud, or Alibaba Cloud—before you can launch an instance. Because of these hurdles, providers can tack on many surcharges. For example, bandwidth on Alibaba Cloud (and its peers) is charged in increments (starting at 100 Mbps and increasing in steps), unlike the flat-rate interfaces in the US. In short, the entry barrier for building a cloud in China is much higher, meaning Chinese providers have stronger “fences” than their US counterparts, who are now under pressure from neocloud initiatives—especially when it comes to GPU build-up.

D: Yes and no. I think it’s a bit counterintuitive. While I agree that the barrier to entry for cloud services in China is higher than in the US—thanks to tougher regulations—the AI boom is still a significant driver for cloud adoption.

C: That’s exactly the issue we discussed before. As a Cloud Service Provider (CSP), the traditional competitive edge was to “compete on cost” – to drive margins down. But ever since AWS set a benchmark, everyone believes the strategy should now be to “compete at the top” by offering value-added services like S3, and even more services such as database hosting.

D: Well, that approach might be optimal in the short term but is bearish in the long run. If you’re aiming to build a high-margin business through economies of scale, you need to be the lowest-cost provider.

C: Exactly. But I believe that in the US, the only business effectively competing on margins is Amazon’s retail segment. They have always driven margins down. Before Jassy took over, Bezos did a great job in that regard; now, however, Jassy isn’t performing as well.

D: I agree. However, my counterpoint is that while the CSP model isn’t as dominant in China due to lower barriers to entry, US cloud providers enjoy much higher margins. This is likely because the US has a robust ecosystem of SaaS and enterprise software, whereas in China neither consumers nor enterprises are accustomed to paying for subscription services.

C: Maybe. I think it’s because we aren’t entirely clear on how the cloud business operates in China. We don’t have a proper bird’s-eye view to tell whether Chinese cloud services are driven by SaaS (like in the US) or if they’re more government-focused. We only understand the US perspective, where a decade of SaaS growth has allowed AWS, Azure, and GCP to charge exorbitant fees for additional services. Even those SaaS models seem to be in a ridiculous position. The other day, I saw Datadog’s pricing: it used to be per seat, but now they display a full page of offerings in tiled format—a 5 by 7 tiles—with each tile costing around $10 extra. You have to crunch the numbers just to figure out your bill. I don’t think this business model is sustainable in the long run; it’s egregious.

D: They’re just trying to milk you.

C: Exactly. It is “enshittification”—you build a good product to attract users, then you cram in extra monetization opportunities until the core user experience suffers. Anyway, let’s talk about Microsoft.

MSFT

D: Oh man, Microsoft is such a peculiar company. They currently occupy a unique strategic position—about a year or two ago, they were the AI darling after investing in OpenAI. Everyone saw them as the exclusive provider of the OpenAI API and Co-Pilot.

C: They still offer Co-Pilot, but now they’re facing competitors like Cursor, Cline, and several others. I’m not sure what happened to Co-Pilot—I’ve never used it, but there’s a lot of competition now.

D: Everyone’s trashing Co-Pilot. But let’s not even get into that—let’s focus on the OpenAI partnership. Back then, it probably added around 20% to Microsoft’s valuation premium. Now, a year to a year and a half later, it looks like they’re getting cold feet and pulling back on that partnership.

C: There are even rumors that Grok 3 is outperforming OpenAI because OpenAI is still relying on its old base model for new reasoning fine-tunes—mainly because Microsoft isn’t willing to invest in the data centers needed to train the next generation of GPT models.

D: Those are the rumors, but I think we’ll see how it plays out soon. That’s why some say OpenAI goes to Oracle—really just Oracle and SoftBank. It’s curious what made Satya change his mind. I don’t get it—if you look at Azure’s growth, AI-related revenue is showing triple-digit year-over-year increases, which is its only bright spot. In fact, Microsoft’s stock price largely depends on whether Azure is accelerating or decelerating. In that sense, Azure needs OpenAI far more than OpenAI needs Azure. Perhaps that’s why Microsoft’s stock hasn’t moved much over the past year.

C: Right. It’s tough to bet against Satya right now. Satya has been a darling since the OpenAI days and still commands that status. Betting against Microsoft is risky. They continue to make curious moves, like that recent rumor you mentioned.

D: Today? The rumor is that Microsoft is backing out of its data center build-out deals. I’m not sure how to interpret that or whether it’s even true. For one, it contradicts the latest earnings call where Satya mentioned that Azure is capacity-constrained. Secondly, it doesn’t match the upbeat tone he used in his recent podcast discussing exponential growth in compute power. And third, even if it is true, it might be fine—OpenAI recently partnered with SoftBank and Oracle on the Stargate project, so it makes sense for them to redirect some of their compute to that alliance. Perhaps Microsoft is shifting from building huge, centralized clusters to a more distributed data center model. But I don’t want to speculate too much since this is just a rumor on Twitter—I even checked with a friend at Millennium, and they hadn’t heard anything concrete. If true, it could serve as another DeepSeek-like catalyst. There are just too many nuances to this rumor.

C: But you all keep saying that hyperscalers are the ultimate winners.

D: I think that’s a lazy answer. Sure, in the end, hyperscalers might be the winners—no one can argue with that—but…

C: I believe some argue that vertical integration will win out.

D: Vertical integration of what exactly? Software? Everything? Essentially, everything.

C: But I actually disagree. I think vertical integration almost counters human nature—no single company can excel in everything. Even if a company is driven by a strong leader, it has its areas of strength and weakness. It’s nearly impossible for one dictator to manage both upstream and downstream integration perfectly. Look at Apple or Tesla: they’ve built their empires with strong leadership, yet both show clear cracks—Apple’s software, for instance (I recently swipe down to access the iPhone taskbar and saw a messed-up UI with unfixed bugs), and Tesla’s battery tech has faced delays (as noted by BYD). This all highlights the inherent vulnerabilities of vertical integration.

D: Even so, consider the time and money Apple invests in vertical integration—it’s not a one- or two-year project. Remember the Qualcomm fiasco? Apple aimed to cut Qualcomm out as their Wi‑Fi provider, for example.

C: But they did manage to pull it off. For instance, with the iPhone 16e, they’re now using their own designs for both Wi‑Fi and cellular connectivity.

D: I know, but how many years did it take? They first did that around 2018, so who knows how long such transitions will take in the future.

C: Because initially, they tried partnering with Intel—but after Intel failed to deliver, Apple poached their engineers and moved the design in-house. The point is, when you build a company and pursue vertical integration, you end up with a multitude of business units (BUs). With so many BUs, it becomes nearly impossible to evaluate performance, and the whole process turns into a bureaucratic, political game. That’s why I believe vertical integration fundamentally counters human nature—it just doesn’t work well.

D: I get it, and I’m not disagreeing with you. My point is that even if Apple is determined to achieve vertical integration and won’t give up after the first try, it still took many years to complete. How many years did it take them?

C: Exactly, 6 years.

D: That’s exactly my point. It’s impressive—because any other company attempting this would likely fail.

C: Yeah, most companies would have given up by then.

D: Even if a company has the cash flow to support vertical integration over many years, it still took Apple about six to seven years for even a minor component. When people say, “Vertical integration will win,” they’re not accounting for the internal complexities, bureaucratic dynamics, and long timelines involved. Plus, it depends on the pace of specialist build-out. Sure, all hyperscalers want to bring chip design in-house, but NVIDIA isn’t just idly watching its market share slip away—they’re innovating and building strong competitive barriers. They just need to focus on what they do best.

C: But I believe NVIDIA will eventually have its “Cisco moment.”

D: Oh yeah, definitely. Jensen is doing everything he can to escape that.

C: Plus, he’s been there before—during the crypto boom, they overloaded the channel and the chips didn’t sell. Jensen is determined to avoid a repeat of that. But that’s human nature: you can delay it, but the “Cisco moment” is inevitable.

D: Yeah, though probably not anytime soon.

TML

C: There’s some good news for NVIDIA this week—the Valley’s worst-kept secret: Mira’s startup, Thinking Machines Lab, was unveiled on Monday.

D: Also, how much money do you think Mira’s startup raised?

C: Rumor has it around $200–$300 million.

D: I think that’s on the lower end, given their ambitious plans and the caliber of their founding team.

C: It’s an all-star team—people who worked on OpenAI infrastructure, folks from PyTorch, and of course John Schulman (whom you already know). With a founding team like that, it’s clear they will build another foundational model. They aim for superintelligence—but not by a direct, head-on approach (as Ilya described with SSI). Instead, Thinking Machines Lab is taking a more product-driven, gradual approach, validating their product first before eventually reaching that level. It’s a different strategy.

D: But I was just thinking that $200–$300 million seems low. They likely have 20 or more people on their team—roughly $10 million per person. Their combined net worth probably exceeds the funds raised in this round.

C: Sure, that’s just a street rumor—they probably raised a bit more, or at least have additional offers lined up. It’s intriguing, and I’m optimistic about the company. After all, the naming conventions in AI startups are telling. OpenAI eventually became “closed,” and Stability AI turned out to be, well, unstable. “Thinking Machines” is a strong name; in the context of the Butlerian Jihad, thinking machines were responsible for wiping out billions of human lifes. They’re going to create a safe AI—unlike SSI, which might end up being the most dangerous company we’ve ever seen (wink-wink).

D: I agree—having more labs like that is good for NVIDIA because it diversifies their customer base and increases pricing power. But do you think they’ll try to train a foundational model from scratch rather than simply building on an open-source model like R1?

C: I believe they will.

D: Why?

C: Because they have the talent to do it—and training has become relatively inexpensive these days. If you don’t follow Elon’s approach and maintain a principled stance, they could likely train a model for around $50 million or less. Moreover, having your own training pipeline offers numerous advantages. For example, a recent DeepSeek paper on native sparse attention shows that it can be up to ten times more efficient than using a full attention model during inference but requires training a new model from scratch. The training capability is especially valuable since there aren’t any good, from-scratch, open-source multimodal models available. With eventual moves toward reinforcement learning, building a foundational model from scratch might be far more cost-effective. So if you have the money, it makes sense to try.

RUS

C: There’s a looming issue we discussed—the Russian sanctions and its relation to cryptocurrency.

D: I think there’s a lot of chaos in crypto this week. For one, Argentina’s President Milei released a memecoin on Saturday, and within three or four hours, everyone in his inner circle cashed out.

C: Are we sure he actually cashed out? I’m still on the fence about this—I suspect he might have been played and is just that insiders are cashing out. For instance, with the Trump coin, in the Crypto Ball event, everyone already knew it would launch the next day—it’s basically the ultimate insider trading. Next time there’s a crypto event, everyone should attend; it’s an opportunity to make serious money.

D: And on top of that, Solana was hit hard over the weekend—dropping from around 240 to 160. Why? I think people got angry and started to realize that this might be a scam.

C: I don’t know—I'm really confused by all this. Crypto appears to have the strongest backing from the Trump administration, yet it hasn’t delivered much. Solana and Ethereum are like competing twins: Ethereum is based on a variant of JavaScript, while Solana is built on Rust. They use different implementations (one uses EVM; the other relies on a glorified BPF), but both have founders alive holding significant stakes. In the end, when we consider things like the Trump coin and the Melinda coin, we wonder if the World Liberty Foundation will support Solana—but it turns out they’re investing heavily in Ethereum too.

D: Did you see the news today? Bybit exchange was hacked, and about $1.4 billion worth of ETH was stolen. I’m not sure if it was a true hack or just an embezzlement, but $1.4 billion in Ethereum is missing.

C: Yeah, but there’s more bad news for crypto, which makes me question what Trump was thinking. Right now, the only asset class still trading in Russia is cryptocurrency. I believe that once the US lifts sanctions on Russia and trade normalizes, the money currently invested in crypto will shift into other asset classes.

D: Yeah, I don’t disagree. It’s pretty obvious that speculative money from crypto has flowed into Russia’s stock market—which skyrocketed this week—as well as into Chinese stocks. There was a lot of hype around Chinese ADRs, Hong Kong stocks, and even mainland Chinese equities last week, while crypto (even Bitcoin) hasn’t seen much marginal buying. We all know Asians love to gamble in crypto, but now that money is moving elsewhere.

C: And let’s not forget Binance, OKEX—no matter where they operate, at the end of the day, they’re fundamentally Chinese.

FED

D: Right now, the market looks a lot like it did in 2018—the big drop then seems to be repeating itself. First, remember how Trump kicked off the trade war with tariffs? At first, that made the market very jittery. Then, after the US-Mexico-Canada negotiations, everyone started to ignore it. But now, he’s continuously rolling out new tariffs and cranking up the pressure, and the decoupling between the US and Europe is becoming quite obvious. So, everyone is gradually starting to price in that risk. And on top of that, with a pretty hawkish Fed, the overall market backdrop is full of uncertainty. Worse than in 2018 is that inflation is still stubbornly high.

C: But it shouldn’t be stagflation. I mean, with tariffs, it’s hard to imagine the economy merely stagnating—it's more likely to plunge into a recession. And without government spending during a recession, you wouldn’t really see inflation.

D: Well, I mean, sure—it starts as stagflation, and then inflation eventually drops, leading to a full-blown recession. People worry because, with stagflation (persistent inflation with no growth), the Fed loses room to cut rates. That’s the biggest concern right now. OTOH, nce growth stalls, inflation naturally falls—unless you can imagine a scenario where inflation never crashes, which seems unlikely. Essentially, as soon as market expectations become very cautious, inflation will drop.

C: So, if things are like 2018 now, what should we do?

D: Well, I think there have already been some early signs. For example, last week with reports about government spending cuts, Palantir was the first to drop by 10%—or even more. That drop in Palantir then triggered a broader pullback in stocks within the same factor group—like Applovin, Robinhood and others—which have been soaring this year, mostly driven by retail investors. Now, those stocks are broadly down by roughly 20%. So, I will be very cautious from here. The overall backdrop isn’t favorable, and if NVIDIA’s earnings also disappoint—and if market data continues to look weak—I wouldn’t be surprised if the market sees a decent correction of about 10%. It’s not unimaginable. Therefore, it might be wise to add some hedges. The simplest approach would be to use QQQ or SPX hedges. For a more aggressive tactic, you could consider individual stock hedges—I was thinking about buying puts on Supermicro. Supermicro has risen significantly year-to-date and faces a big risk: will it meet its deadline on the 25th? Additionally, it acts as a hedge for NVIDIA’s earnings too, since Supermicro is technically a customer of NVIDIA; if NVIDIA’s earnings are poor on Wednesday, Supermicro will likely drop as well. Of course, you can pick other stocks, but my idea is to add a bit of QQQ puts, some long positions in 10-year treasuries, and a few individual stock hedges like Supermicro puts to protect the downside.

INTC & TSM

A: First, Intel's Foundry business is spiraling downward. There isn’t any potential buyer besides TSMC—Samsung Foundry is in terrible shape, and GlobalFoundry lacks advanced process technology. Essentially, only TSMC can step in.

B: I think there are only two ways to take over: a joint venture or a direct purchase. However, I doubt either option will be straightforward. My guess is that if the Trump admin can make money, they’ll first carve the business out of Intel and then decide whether to recapitalize it. Essentially, they’d acquire a small stake—for example, people like Howard Luternick might get in—and then later sell that stake to TSMC, effectively paying the Trump admin insiders. It might be a full sale (which would be ideal for TSMC), but it might also end up as a JV if things don’t go so smoothly.

A: I believe it doesn’t need a full recapitalization; it just needs these investors—or the capital behind them—to buy in at a very low price, with TSMC also taking a stake. With TSMC’s operational excellence, they can revive the business and turn it into a cash-generating machine. Those low-priced shares would then appreciate in value, potentially allowing the business to go public again (perhaps as “American SMC”).

B: But to revive this business, you need it to turn a profit on a longer time horizon. You might think about making money in four years, but even that may not be enough—you need to be profitable before the midterm elections. So the time horizon is really tight (less than two years). That means TSMC would have to invest using its own shares, which is ideal: ideally, someone in the Trump admin holds your stock so that it just keeps rising. Of course, nothing’s ever perfect.

A: But I think their options are very limited—even if they don’t want to do it this way. There are barely any local players in the U.S. who can revive this mess; you can’t expect Apple or Samsung to take it on. Essentially, only TSMC is capable. And let’s be honest: TSMC doesn’t really want this business anyway. They’ve said it over and over—“We don’t need this.” Besides, the normal approach for TSMC would be to build a plant in the U.S., gradually relocating staff. That’s optimal for them. Taking on Intel’s messy business is like building another plant in Arizona—which, given the incompatibility of Intel’s processes, is almost like starting over.

B: Essentially, it all comes down to spending money. They might even have to fire some people (despite having to “respect” certain opinions) to pave the way. In the end, it’s just paying to buy a path—that is, spending money to secure goodwill from the current admin.

A: Right—and that includes the tools. With TSMC suddenly managing multiple plants, they still have to purchase equipment.

B: Moreover, TSMC’s capital expenditure wouldn’t really make sense here. If it’s a joint venture, the CAPEX can be treated as an investment—meaning it won’t directly hit TSMC’s CAPEX line, even though the cash still goes out.

A: But if TSMC invests in this, it harms its current shareholders—who are basically in the clear, just waiting for Intel to collapse. Why should TSMC spend money to nurture a competitor? So even in a JV, TSMC would insist on being the controlling partner, because its current shareholders (who aren’t the same as the management team) won’t approve otherwise.

B: So that’s the crux of the issue. I don’t think the Trump admin is truly trying to fix things—they’re just looking to cash in. How they do that is a brainteaser, but I believe they would require at least 100-fold returns. It’s really up to TSMC’s management to navigate this tricky situation.

NVDA

B: And NVIDIA—NVIDIA stock has been climbing for over a week now.

A: Right. Essentially, they’re making up for that big gap which is nearly closed by the 27th. The volume isn’t huge, but U.S. analysts are still worried about the earnings report on the 26th. They fear an “air pocket” between the Hopper and Blackwell product cycles—if Hopper’s demand or sales drop too quickly, and Blackwell’s shipments are hit by supply chain issues, the revenue will suffer, and even next quarter’s guidance might be weak. Fundamentally, U.S. analysts are nervous about both this quarter’s results and outlook. Personally, I’m not overly worried—expectations have already been lowered considerably, and everyone knows about the supply chain issues with Blackwell and the subdued demand for Hopper. It’s more a price correction than a demand surge, and liquidity has improved. (Back when, both specialists and generalists argued, but by January 27th, nearly all the generalists had exited NVDA—leaving only the specialists.)

B: Now, if you look at all the neocloud providers, their rates are roughly between $1.59 and $2.30 per hour for H100s.

A: That’s still lower than I expected. I thought we’d see a price increase, but apparently not.

B: Also, any price increases might take some time to reflect because it’s really hard to do a channel check right now. There are so many shipment channels—you might see weak shipments from Lenovo, for example, while other channels are still strong. Especially with Supermicro, it’s hard to track where everything is going.

A: However, recent news appears positive. Major companies like HP have already shipped their products, Foxconn said its supply issues are solved with shipments exceeding expectations, and some Taiwanese factories are also shipping. Plus, Supermicro—and even Dell—has mentioned a $5 billion XAI GB200 order. You can indeed see that on major neocloud platforms, the GB200 (or B200) products are rolling out.

B: But still, these products aren’t accessible—ordinary folks can’t get them.

A: Aren’t Corewave, Lambda Labs, and similar platforms already online?

B: Commoners can’t access them; only those with long-term partnerships might have a chance—but it’s still a long way off.

A: But China’s DeepSeek move is fascinating—they’re deploying loads of inference capabilities. I heard that everyone’s been frantically buying up Hopper inventory; for instance, Tencent placed an order for 200,000 H20 cards because WeChat has been integrated with DeepSeek, right?

B: Right. They’ve always been pretty adept at smuggling. Just the other day I saw on Xiaohongshu that NV72 chips were already being smuggled into the country—there were photos and videos showing them in containers, and you could even DM to buy them. Pretty impressive. So, I think there are still plenty of Supermicro systems around. By the way, has Supermicro released its earnings report yet?

A: No, it hasn’t. Their filing deadline is the 25th—if they don’t file by then, they risk being delisted on February 25th. So, Supermicro is a variable, and expectations for their earnings report are low. I remain cautiously optimistic, and I think NVIDIA should be as well. As for Supermicro, things are messy: if they end up shipping some inventory to China in violation of the ban, dealing with that will be very problematic. Their immediate priority should be securing an accounting firm that can sign off on their report—by now their BDO should be on board rather than pulling out at the last minute.

B: Supermicro does have some controls in place. First, when you buy one of their machines, you must sign an agreement promising you won’t violate regulations (i.e., you must keep it local). Second, the machines come with geolocation locks—the IPMI (their board management system) checks the IP address when online and will lock itself if it detects a Chinese IP. Of course, these measures can be bypassed, but to claim that Supermicro chips are being smuggled into China without their knowledge is a stretch—it’s hard to imagine thousands or even 30,000 units being moved at once.

A: I think it’s very difficult to smuggle in several thousand or 30,000 units at once.

B: I’m not sure, but on Xiaohongshu they’re selling systems that start at 32 or 64 cards, so the quantities are significant.

A: I think China’s current wave of NVIDIA procurement hasn’t been fully modeled. I once read a sell-side report saying that cloud CAPEX is positively correlated with stock prices—a positive feedback loop. The higher your stock goes, the more you invest in CAPEX, which in turn boosts the stock even further. It seems the Chinese market is just beginning such a cycle: as soon as a company announces integration with DeepSeek, its stock can jump 20–30%. After years of bearish conditions, Chinese companies see this as their chance and rush to integrate—quickly grabbing NVIDIA chips to avoid lagging behind.

B: Essentially, it’s like using investors’ money to buy cards.

A: Not exactly—it’s called positive feedback. Investors reward innovation, which then fuels further success by great companies (wink-wink).

META

B: Meta has been on the rise for over twenty days now.

A: Twenty days—that’s when I started thinking Meta’s stock price has become somewhat detached from its fundamentals.

B: For shareholders, it’s like living in constant fear of the stock rising even higher.

A: Yeah, but honestly, I feel most people view Meta’s stock through a narrative lens rather than focusing on fundamentals. Technically speaking, do you think Meta’s model has fallen behind?

B: It depends on Llama 4, doesn’t it?

A: Yes—but when exactly will Llama 4 be released?

B: DeepSeek R1 has been trained within three weeks, so Llama 4 will undoubtedly be a reasoning model—but it depends. I think the biggest challenge for U.S. model makers is regulatory compliance. With LLMs it’s somewhat easier since, say, book data isn’t bound by as many rules. But there are still nitty‐gritty details. For example, Apple buys its own data, it can only secure about 1–2 tillion tokens. (They can’t even use YouTube transcripts!) In contrast, DeepSeek’s data is thoroughly processed and cleaned—filtering out low-quality bits—to yield around 14 trillion tokens. Imagine how much obscure data might be included (maybe even “borrowed” from places like the Chinese National Museum). But for a giant like Meta, non-compliance isn’t an option. That’s their challenge. Early movers like OpenAI and Anthropic have legacy models that aren’t compliant, and they can synthesize new, compliant data from these models—giving them a first-mover edge. Meanwhile, Llama 2 started with non-compliant data, forcing Llama 3 to adjust; now Llama 4 must do the same. Data compliance is a huge hurdle, not to mention whether to use a dense model or a mixture-of-experts approach. I don’t think you should focus solely on whether Llama 4 leads or not.

A: I don’t need it to be ahead—I just need it to catch up.

B: Exactly—he doesn’t need a model that catches up because Zuck’s idea is to commoditize their competitive advantage. Whether it’s Facebook or DeepSeek, once it’s commoditized, that’s that.

A: I know—that’s your classic argument: once the model is commoditized, the model layer loses value and the distribution layer captures it. But what exactly is Facebook distributing? Who is it reaching, and how does it make money from distribution? I feel like no one can explain that clearly.

B: It’s because you can imagine a middle-aged woman in her 40s or 50s playing a casual gardening game—or a runner game—on her phone. You wouldn’t expect her to be having deep, philosophical conversations with GPT on her phone, would you?

A: The only scenario I can really picture is shopping.

B: Facebook’s main app is essentially dead among young people—and even among the more discerning—since everyone’s now on Bluesky, X, Instagram, Threads, etc. However, Facebook still has Instagram, Threads, and WhatsApp. I think WhatsApp is the most likely platform for distribution. But as of now, no one really knows what the product will look like—not even Zuck. We all have to figure that out.

A: Could they possibly emulate WeChat? Even WeChat’s product form remains a bit of a mystery—no one really knows, so they’ll have to brainstorm.

B: Up to now, the only “product” you really see is one that essentially takes people’s jobs. Zuck wants to lay off 20% of the workforce each year—reducing a company from 50,000 employees to just 500—while still being a trillion-dollar enterprise. You can just do the math on how much each person is “worth.” But if that’s the model, then the only practical use for large models is to fire people and replace them. Yet, Facebook isn’t a productivity app; they tried Facebook Workspace, it didn’t work, and they moved on. It’s not that they failed—it's just not in their DNA to chase that kind of revenue. We simply don’t know if Workspace will succeed, and even if it would, the company isn’t really focused on that type of profit (they’ll let CRM and Slack handle that).

A: So do you think Facebook, if it wants to make money, still has to rely on its core strength—advertising?

B: Exactly. In advertising, you need a user interface—an entry point like GPT. With GPT as your interface, it’s incredibly easy to insert ads. You simply put your ad into the system prompt, instruct it to insert the ad in the right place, and it does so seamlessly—you might not even notice it. But if you don’t have that entry point or brand, you’re at a disadvantage.

A: I agree—but here’s a question: what about video and image models? Wouldn’t that be even easier? In the past, you’d have to shoot a short skit and formally insert an ad to target a product. Now, you could simply say in your prompt, “I need this image (or text-over-image, or text-over-video) to include, say, soy sauce brand A here and brand B there,” and have it clearly labeled.

B: That’s too simplistic. The ultimate aim of advertising is to optimize conversions—whether the purchase happens immediately after seeing the ad or gradually within 30 days. Your approach is overly convoluted and lacks a clear target metric. It’s hard to tie, say, which soy sauce brand was used in a generated video. Sure, it might have meaning, but ultimately if we’re using generative AI, not having someone shoot a video, we would care a lot about cost and measurability, so I have a thousand other ways to optimize my marketing. For example, I used to run Instagram ads using ten influencers. Now, I could use software to generate a thousand different influencer profiles—covering various ages and ethnicities—and target the group most likely to be influenced, then continuously serve them my ads.

A: So this is really about lowering content creation costs rather than improving targeting or conversion.

B: Actually, it’s about improving conversion—by generating a greater variety of content, you can more quickly identify and target the group that’s easiest to influence. In the past, creating such varied content was much more challenging.

A: I think there’s another positive for NVIDIA—Grok 3 is about to launch. I’m not entirely sure what Grok 3 is, but everyone seems to see it as the ultimate test of the pre-training scaling law for large models (especially among investors who only have a superficial grasp of them). My lesson from DeepSeek is that most investors don’t really understand what large models or deep technology are—whether they’re generalists or specialists. Yet there’s a consensus: Grok 3, built on a 100,000-GPU cluster, will prove whether the scaling law still holds. If Grok 3 outperforms current models, it confirms that NVIDIA’s training demand remains strong. In short, if Grok 3 scores well, I believe NVIDIA’s stock will get another significant boost. But circling back to Meta—what exactly are they distributing? No one seems to have figured out how to monetize that, and frankly, I don’t understand it.

APP

B: Applovin is in the ad distribution business, right? I think their traffic is pretty crappy. Look at Google—Google’s AdWords far outweigh AdSense from back in the day. Fundamentally, an ad network is all about selling your impressions, which is a pretty shitty business. If your traffic were stellar, you’d just run your own ads. The internet is winner-takes-all: if you’re good, you grow huge, and most ad network traffic ends up coming from mediocre software or games banding together. In short, if your traffic isn’t already impressive, no matter what you do, it’s like trying to decorate a crap. Even if you take everything at face value, the key question remains: can you sustain 50% growth?

A: Right—that’s fundamentally the issue.

B: But if you’ve already captured all the available impressions, there’s only so much you can overcharge—your growth is basically capped, determined by how much you can charge per install.

A: And we’re not talking about 50% year-over-year growth—it’s 50% quarter-over-quarter growth.

B: Right—I think the company can maintain its growth momentum, but how they’ll sustain it remains unclear. Ultimately, if you need to keep up unsustainable growth, you’ll have to buy back some of your traffic and cook these spends away from your books. At the end of the day, you can’t fake the cash flow; you’ll always have external data showing how much money is coming in and going out.

A: Right now, buying traffic is very efficiently priced—good traffic is expensive.

B: Exactly. I think they’ll essentially start “paying themselves” if they have to maintain 50% growth. For instance, in e-commerce, if you sell a shirt for $40 on Facebook and spend about $20 on ads, then on Applovin you might set it up so that you only effectively spend $1 per click—essentially subsidizing you by $19 to get the ad out. That $19 wouldn’t be a direct subsidy; it could be structured as a coupon (say, after buying a $40 shirt, you get a $19 coupon) to incentivize repeat purchases.

A: Aren’t they already doing something like this?

B: I simply don't know.

A: Look at what Lauren Balik posted on X—she mentioned that they’re already using gift cards to incentivize e-commerce purchases.

B: Right, but the point is you can see exactly how much money is flowing in and out. While companies will inevitably obscure some of that—since as they grow, they spend on CAPEX and R&D and don’t pay dividends—the cash flows are still trackable. In this industry, it’s crucial to monitor where the money comes from and where it goes. That said, right now, whether you go long or short, you’re likely to be off the mark—the timing just isn’t right. So, stay away.

Subscribe now

Higher capitalization requirements for BABA / Ant to shed risks with the banks A regulatory requirement for credit information sharing between states, banks, regulatory bodies and other private enterprises Some transparency and additional pro-competitive requirements that would benefits their competitors such as JD and PDD

We could also see some experimental regulatory requirements that eventually will benefit big players than smaller ones, much like the regulatory requirements introduced in 2018 to Games. What you won’t see, as the western investors worried, are CCP members on the board or cap tables. There are more innovative structures to exercise controls than the plain cap table or Board seats.

The question to ponder, more interestingly, is how the synergy would work out for ARM and its potential buyers.

NVIDIA

NVIDIA has some presences in mobile. From its ill-fated Denver CPU to the modest success with Tegra X1 in Shield and Nintendo Switch, for one, it is not lack of trying. So far, NVIDIA has established itself as an interesting player in the mobile space, while not gaining meaningful market share in the otherwise vast mobile market.

On the other hand, NVIDIA’s data center business has enjoyed a strong growth trajectory in the past a few years. It has established itself, rightfully, as the go-to choice for any high-performance computing needs.

It is likely that the acquisition of ARM won’t be for the mobile market share. With a strong CPU design portfolio like ARM, NVIDIA can have more flexibility in its high-performance computing package design. Right now, NVIDIA’s flag-ship design is DGX A100, with its CPU being AMD Epyc. The past few years, in the high-performance computing world, the role of the CPU is transitioning from computing powerhouse, to a computation coordinator. This is particularly obvious from the accelerated development pace of PCIe 4 and PCIe 5 standard.

That won’t be nearly enough for Jensen Huang. Mobile-based solutions such as NVIDIA Xavier have been based on the shared memory design for quite some time. While there are challenges particularly on cache coherency model for such design, a CPU-GPU shared memory design would alleviate many pains for software developers to unlock the computation power of such a machine. The recent deep-learning models have been hinting a future of at least two orders of magnitude more parameters. To accommodate these new trends, a bigger departure from current PCIe based interconnect and form factor for accelerated computing co-processor is needed badly.

Apple

One thing interesting about Jensen is that for the past two decades, he always has the right insight. However, his endeavors were expensive for the company to pursue, with an often stretched credit line.

Apple has far more cash reserves for such an expensive purchase. But an ARM purchase would be a defensive move. Apple’s roadmap, while heavily ARM-based, doesn’t require radical departure from what ARM’s currently doing. The cost of inaction to Apple, would be a new adversary owner of ARM. From the past track record, NVIDIA's ARM would likely not prioritize any Apple’s request.

Apple had made defensive purchases in the past, and has the habit of hoarding good technology companies for many years until its technology has matured. Apple’s ARM purchase would be boring, but a strategically important one for them.

Google

A Google’s ARM would be a distracted one. There are many product lines within Google that would benefit from an ARM purchase, from its mobile phone, Chromebook to its data center. The distraction and the likely politics interplay will likely terminate these endeavors prematurely. This will be an OK turnout for its competitors such as Apple. A Google’s ARM won’t be anywhere near as adversarial to them as NVIDIA's purchase.

A good asset like ARM would have many potential buyers. Many of them I didn’t cover here but equally likely. To me, as an interested observer, the most exciting one would be NVIDIA by far.

Crude/China demand induced manufacturing recession, similar to late 2015- early 2016 - peak to trough drawdown about 15%;

Covid19 uncertainties and potential dent to consumer confidence. Depending on the policy response (both policy and fiscal, don’t expect it to have a lasting effect of more than two quarters) past episodes imply a negative 10-15% drawdown.

Two effects are not additive. They can partially offset each other. Lower energy price means less inflation pressure, room for loose monetary policy, savings for consumers. They can also compound each other just through worsen sentiment and reflexivity. Our conservative guess right now is a 20-25% market wide drawdown with high volatility and potential to undershoot. No system wide liquidity crisis in sight at the time of writing, so we are not calling for a “sky is falling” >50% draw down distressed scenario. Our current portfolio holdings of Semis are relatively well positioned. China supply chain is estimated to go back to full capacity by end of March. HPC and cloud capex should be counter cyclical. Gaming/PC demand might weigh on the revenues and earnings a bit.

AMD - add around $40, ~x25 normalized earning with some downside protection NVDA - add around $200-220, again ~x25 TSM - add around $48-50, ~x18 Cut SPY and PM gradually to raise cash Tencent and DG - hold Uber - wait a bit to see if the support range $25-$28 holds Redfin and HD, do nothing now, don’t go against the cyclical headwind

Watch list (more details on valuations to follow, ordered by priority):

AAPL - add around $220-$250, x18 GOOGL - add around $1000-$1100 VISA: - add around $140

Intel’s 10nm Canon Lake debacle taught us that not every shrinkage are good in terms of raw performance. Shrinkage in manufacture process can certainly help power consumption, but as we go down to 7nm and 5nm territory, it is far from given performance will be better with each lithography process improvement. Particularly, the widgets provided by TSMC would certainly favor simpler units as the process continue to shrink. When we look at improvements on IPC, we should take these as improvements over architecture tweaks, rather than natural benefits from the shrinkage. With that in mind, let’s examine what Zen 2 gives us from the specs.

Floating-point

When comparing Zen with more recent Coffee Lake, people often complained the lack support of AVX512, and the neutered support for AVX2. Limited by the floating-point register width in Zen / Zen+, AVX2 support was emulated by two 128bit registers. With the extra transistors available in 7nm Zen 2, AMD finally added 256bit width registers for AVX2. This is a welcoming change for productivity apps such as ffmpeg, Adobe Premiere or DaVinci Resolve.

Cache

It is no coincidence that in 7nm, AMD managed to cram 32MiB per chiplet L3 cache into Zen 2. Cache is expensive when you can use these transistors for many important use cases (more ALUs, more FPUs etc.). Given the 7nm trade-offs we discussed earlier, it seems far more economical now to use these transistors as memory instead. AMD is not alone. A12 Bionic from Apple doubled the system cache size as well when moving to 7nm. For today’s application, more cache meant less memory stall and better real-world performance.

Chiplet

Zen architecture was known for its chiplet design. Chiplet design lets AMD to ship the same Zen chip from server all the way down to Ultrabook laptop. The new Zen 2 design separated IO / Infinity Fabric into another die manufactured at 14nm. It is a cost saving design, but rumor on the street is that the new Infinity Fabric can be operated at different frequency from the main memory. It would still be interesting to benchmark how the separate IO die going to impact more multi-thread cooperative programs (i.e. complex gaming logic).

Conclusion

We left with very positive impression after AMD’s Computex keynote. From technical standpoint, it made all the right decisions in Zen 2. With the correct leverage over 7nm process, it is hard to find how Zen 2 cannot deliver the performance boost it promised across server, HEDT, desktop and laptop offerings.

MAU metrics is not without faults. Facebook traditionally under-reports its MAU numbers (their milestone announcements such as 100M MAU and 1B MAU both comes a few days, even weeks after stabilized around that number). It can also be affected by seasonality. Holiday seasons and New Year can see positive changes to MAU that are temporary.

While MAU is a reasonable metrics for many investors to look at. New generation of social media companies such as Snap are advertising DAU (Daily Active User) metrics. Mark Zuckerberg, was reportedly looking more at DAU and WAU (Weekly Active User) after the company's shift to mobile in 2012. On mobile, if a user only opens the app once every month, it is not a meaningful interaction and not so much monetizable at all.

The recent changes for Twitter's quarterly reporting to mDAU (monetizable Daily Active User) is an interesting move. This cuts to the core of how different companies report DAUs and why an accurate reporting of DAU is difficult. Thus, prompts me write this article to give a quick primer on DAUs.

Monetizable in DAU?

There are several things interesting in Twitter's announcement of this change. Many other social media companies don't put the emphasis on monetizable because it is implied. It simply doesn't make sense for Instagram having DAU that is not monetizable. Users can only interact with Instagram on the Instagram app, and Instagram sells ads globally. The only exception, for Facebook / Instagram probably comes from less than a million people in the holdout group that have never seen ads from the very beginning. Facebook holds this tiny fraction of people to continuously gauge how much advertising load will start to hurt its user engagement.

It is certainly true in 2012. At that time, Facebook's newsfeed didn't have any ads and its mobile apps only support newsfeed, not the ad units. You probably could, in theory, exclude these people from mDAU metrics as Twitter defines it.

For mDAU in Twitter's case, they specifically called out "people who saw embedded Tweets on other websites" or "viewing Twitter content without login" are not mDAUs. It can be more than that sometimes.

What Action?

It is no secret that social media companies track every interaction you have with their websites and apps. What constitutes as active sometimes can be a contentious topic internally. For example, if every time you receive a push notification counted towards an action, the DAU number can be inflated dramatically (Twitter's DAU to MAU ratio is around 39%, push notification enable rate is around 50% to 90% depending on apps).

Facebook defines the action as logged in and visited the site, that include viewing, commenting, liking or posting contents. Other social media apps could use tap to open as the indication of active. However, this can be tricky. A program bug or crash could artificially inflate tap to open metrics, causing the counterintuitive effect to inflate DAU metrics.

From When to When is a Day?

DAU number is susceptible to cut off times. If a social media's peak traffic time is around 1900 in Eastern Standard Time (United States) and if their cut off time is around the peak traffic time, a user who uses the social media app every the other day can be counted as a whole DAU everyday because their interactions are around the cut off time, therefore, counted towards both the previous day and the day after. Depending on the cut off times, DAU metrics can be inflated intentionally. Many social media apps use its low traffic time as the cut off time, which could change over time as their international presence expands.

Ramadan holiday is famous inside social media company circles because it shifts the traffic pattern considerably in Middle-East countries. Their peak traffic will be pushed a few hours later to after the sunset. For some social media companies whose cut off time is global, this may deflate their Middle-East metrics a bit.

When to Conclude the Day?

We cannot manufacture another hour in a 24-hour day. However, the data pipeline, especially the ones with mobile client involved can be delayed. If a user opens the app at 11:55 p.m. with limited network connectivity, should we wait until that data uploaded to the backend or should we cut loose? If we are going to wait, for how long then? Some companies take 6-hour cut off (i.e. wait until 6 a.m. next day to run the batch job to compute DAU). Some companies wait a whole day. Sometimes it depends on at which point, the extra wait time doesn't generate meaningfully more DAUs; sometimes it depends on the data retention policy in your Term of Services.

What is the DAU Reported Quarterly?

The above analysis ignores the glaring fact that fake accounts and bots exist. Sometimes, these will be counted towards DAU by mistake or simply the inability to identify and filter out.

Because DAU is a daily metrics, the reports in many quarterly earnings from social media companies are actually less rigorous than many thought. Facebook reports the DAU as the averaged DAU everyday over last month of the quarter. Since Facebook traditionally under-reports its engagement numbers, they also subtract unspecified amount to make the number looks more stable. Twitter and Snap report the DAU as the averaged DAU everyday over the whole quarter with unspecified rounding mechanism. All in attempts to smooth the DAU number to be stable quarter-over-quarter while quantitatively meaningful year-over-year. Facebook's calculation are more current while Twitter / Snap's numbers are laggy.

What That Means to Investors?

mDAU / DAU could be a reasonable number to glance over every morning as a routine for management. But it is only one of the many metrics (W6/7, MAU, Core Interaction etc.) competent management team tracks to guide decision-making. For serious investors, the resolution of quarterly reported DAU is simply too crude to base any conclusions alone. Do user studies, gather engagement patterns, understand demographics and geographics, the peak traffic patterns and then make sense of these DAU metrics. The monetization potential lies within the engagement pattern, not the absolute number of DAU.

1. Application service providers (SaaS) such as Workday, Salesforce, or smaller ones (Slack, Okta, Atlassian);

2. Infrastructure providers such as AWS, Google Cloud, Microsoft, or smaller ones (DigitalOcean, Paperspace, Backblaze);

3. Chips and network equipment providers such as Mellanox, Cisco, SuperMicro, AMD, Intel, NVIDIA.

Other players, such as Oracle or IBM, have businesses awkwardly positioned in both 1. or 2. or 3. and neither are dominate market players. Some, such as Microsoft, have significant presence in SaaS. These companies may fit into more than one type.

Application Service Providers (SaaS)

SaaS companies may have a different software delivery mechanism comparing with traditional application software companies. They also have a subscription-based revenue model. However, fundamentally, SaaS companies are competing in the software business. Their competitive advantages lie in how well they serve their customers (how competitive their pricing is, how feature complete to a given customer etc.). They dream to have vendor lock-in but legally they cannot (GDPR).

There are switching costs as any existing software companies have. There will be monopolies on specific technologies (thinking about Adobe). But that is nothing new. SaaS business is not interesting to us because while they may command a premium, the premium is not inherited from the cloud movement, but from other competitive advantages they possess. That is why no wonder Oracle, Microsoft, Intuit (and to certain degree, Adobe) continue to thrive with this new business model.

Cloud Infrastructure

This is the part YZ and I were most fascinated with. The cloud infrastructure companies eventually will be like PG&E, they provides computing as PG&E provides electricity. While sounds dull, electricity is not a bad business! Most infamously, without regulation, Enron made tons of money on manipulating electricity pricing. As if this is not dramatic enough, all major players in this market have deep pocket.

For now, they are not going to command a premium on the infrastructure they provide. It needs to be a healthy business, yes. But AWS, Google and Microsoft can afford losing money for a very long time. (Probably not Amazon, they need to make their e-commerce balance first. But that is not challenging for someone like Bezos.) At the adoption and expansion stage we are in, they are competing with existing self-managed data centers and individual IT departments. It is not their focus to extract more money. Conquer and assimilation take precedence. They will provide cheap and ops-free infrastructure that are miles better than existing self-managed data centers.

As time goes by, how these cloud infrastructure providers operate will diverge greatly from how to run a self-managed data centers. For example, infrastructure providers will be more aggressive at evaluating the components they use. Unlike self-managed data centers where purchase decisions are driving by relationships, the purchase decisions from these infrastructure providers are more or less merit-based. They have much better statistics to evaluate cost of their components. If financially make sense, switch will be swift.

With a deep pocket, the infrastructure providers are not afraid to go down to the component level. They will invest in chips and hardwares that make sense to them. They are also eager to go up. Middleware services such as database management will be much more integrated with specific hardware. The cost of self-managed database on commodity hardware will stop making sense at that time. This is the specialization and vendor lock-in stage.

In specialization and vendor lock-in stage, infrastructure providers will command significant premium because the price collusion they had with each other. Depending on the political environment at that time, it can be extremely lucrative or morally-corrupt.

Silicon

We want to participate the above infrastructure provider business. However, it is challenging. Big players (Amazon, Google, Microsoft, AliCloud) in public market traded as a package of complex business conglomerates. If we buy Amazon today, we are buying into cloud infrastructure, e-commerce (both retail and marketplace), entertainment (both original content and streaming), portable hardware (Kindle), even brick-and-mortar (Amazon 4-Star, Books). Same with Google, and to certain extent, Microsoft. YZ and I don't feel comfortable to make investment decisions with bunch of other risks we don't want.

Silicon, so here things become interesting. Let's take an example of Intel. During the adoption and expansion stage, cloud infrastructure providers will purchase more CPUs so that they have enough computation power to compete with self-managed data centers. In shorter timeframe, it drives un-nature growth on CPU manufacturers. But in longer timeframe, this is a displacement effect. Self-managed data centers will die, along with their demand for CPUs. The growth will fallback with the actual demand of computing power from the society. Maybe with the efficiency improvements the cloud infrastructure providers made, the growth will even dampen a bit.

Why we buy silicon if the future is as gloomy as I described?

Infrastructure providers differ from the self-managed data centers in many ways. First, the purchase decisions are driving by the rigorous cost model from statistic data. If Lisa Su was able to deliver a chip as powerful as Intel's, and as power-efficient as Intel's. If the price makes sense, infrastructure providers will deploy these at fast pace. Looking at AWS today, if we launch an instance, the only difference in the first page is on whether x86_64 or Arm_64. AMD or Intel are only differed in the second page with the little 'a' suffix of the instance type.

Unlike usual skeptics from self-managed data centers, if a software-hardware combination can deliver significant efficiency improvements, the cloud infrastructure providers will adopt quickly. That is the story of NVIDIA in the past three years, rising from a niche gaming chip provider to today's giant in machine learning computing.

Looking forward, in the next two years, we believe the dominance of NVIDIA in cloud machine learning, especially in machine learning training will continue. Inference is easy, and many existing chip makers or startups can make fix pipeline inference chips. For training, you need more flexible configuration of either the loss function, or innovative layers design. Because moving data between different computation devices are expensive, the programmable pipeline support in devices is a significant advantage to drive innovative neural network designs. Unfortunately, only few players have the experience of designing a compatible (to CUDA, OpenCL or OpenMP) programmable pipeline (Intel, AMD, NVIDIA). We fail to see how NVIDIA's lead in that area can be challenged in that short timeframe.

In the next 8 to 10 years, cloud infrastructure providers will start to replace their fleet with custom designed chips (around 30% to 50%). Like what AWS already did for their testing water ARM chips. These chips will start their life as massive parallel chips for web servers, or controller chips for massive storage system (See the recent Huawei’s Kunpeng 920 feature list). Slowly moving up to replace chips for database and other computation-intensive tasks. Looking at how Apple handles its mobile chip lead today, if these chips make sense financially, it will be done in a reasonable timeframe.

However, during that stage, even for deep pocket cloud infrastructure players, they cannot simply afford to (or financially make sense to) build their own fab. That leaves them to collaborate with TSMC, Samsung or Global Foundry for design and manufacture. This is a riskier bet because depending on political climate (and geopolitics), today's seemingly dominate fab could not be a viable choice for American companies. On AMD, we are optimistic that Zen 2 architecture will have the right ingredients to be a sensible alternative to cloud infrastructure providers. It seems they made all right decisions on moving to 7nm, having a separate IO die and double the floating-point vector width to 256-bit. It remains to be seen how it performs in real-world, and what's their launch date like.

Financials

NVIDIA

NVDA sold off massively after Q3 earning and guidance disappointment. While we think in the near term (following quarter or two) the stock might remain weak due to channel and ASP pressure caused by crypto crash, the 3-year thesis remains intact - NVDA can and will continue to grow its data center revenue by at least mid 30 - mid 40 percent annually. Furthermore, with virtually no credible competition on this medium term horizon, we don’t see any problem for them to continue commanding the current gross margin around 70%. Finally, to keep the math simple (and conservative), we give no credit to crypto going forward and assume only modest growth rate for segments other than data center. Modeling it all out, our base case target price for Nvidia comes in at $230, or 70% upside. As for the bear case, the stock will find some support around $110, or 20% downside, if ASP and margin surprise to the downside in the next earnings announcement. That’s not to say this isn’t a real risk, however we feel the risk/reward ratio is really compelling for long term investors who can withstand short term volatility.

AMD

AMD was our best idea early last year, and despite it has gone up more than 80% since then, we remain constructive on the name. One conclusion from our NVDA discussion above is that AMD will not pose any real threat to NVDA in high end GPU or data center. However, we liked what we see on the CPU side. Powered by process advantage from TSMC and Zen 2 architecture, for the first time since 2006 it’s not inconceivable that AMD will become a meaningful second player in server CPU market and can take up to 20% of the market share. Given total addressable market is about 20bn, our base case is that AMD will gain 10% market share over the next 12 to 18 months. Assuming 30% net margin and net-net zero growth rate for everything else, this gets you to a target price of $25-$30. But we do see greater downside risk for AMD if either EPYC’s ramp disappoints or current crypto overhang deteriorates. The stock tends to find some support around $15, which according to our calculation corresponds to a low single digit server market share. The best strategy with AMD in our opinion is to be patient and opportunistically accumulate shares when risk/reward ratio makes sense.

TSMC

In the long run, TSMC will not only benefit from the partnership formed with AMD on 7nm chips but more (we will elaborate in a separate thesis). Though based on our estimation, in near term the contribution from AMD’s 7nm products to TSMC’s top line growth will only be marginal. At 4% yield and with proven pricing power, enviable operating efficiency and best in class management, we continue to view TSMC as a steady compounder even though current valuation is not particularly cheap. Moreover, we are impressed by the accumulative know-how of running a high quantity production state-of-the-art fab and the market seems to under appreciate the process advantage TSMC possesses over Intel right now. Granted, at current stage, it is still something too early to quantify.