Friday Chatter (20250214)

Friday Chatter is a new format we're experimenting with. It features anonymous conversations between two or three people discussing industry rumors and providing both forward- and backward-looking insights into the market.

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.