Friday Chatter (20250328)

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

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.