Friday Chatter (20250228)

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

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?