Friday Chatter (20250418)

Friday Chatter features anonymous conversations between two or three people discussing industry rumors and providing both forward- and backward-looking insights into the market.

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?