Friday Chatter (20250221)

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

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.