Friday Chatter (20250829)

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

Meta TPU

S: OK, our last recording was back in late April. Summer is almost gone, and a lot has happened. The latest news: Meta and Google signed a $10 billion GCP deal. I’m not entirely sure, but it seems related to TPU.

T: Well, I’m not sure Meta really plans to move a lot of training and inference to TPU. If they do, that seems like a very stupid decision.

S: One thing is clear: Google won’t serve the latest and greatest TPU to external customers, no matter what GCP promises. GDM still lacks TPUs for internal research and traffic. Inside Google, GDM’s voice carries more weight than GCP’s, no matter how much GCP wants to be “customer first.”

T: Sure, from Google’s perspective, I agree—they’re capacity-constrained, GPU or TPU alike. But from Meta’s perspective? It makes no sense. If you don’t want 100% dependence on NVIDIA to avoid vendor lock-in, I get it. But switching to TPU is just another, maybe worse, lock-in—with a provider who competes directly with you. NVIDIA doesn’t compete with Meta on models, ads, or social platforms. But if you tie yourself to Google TPUs, you’re literally funding Google’s efforts in areas that compete head-on with Meta.

S: From what I’ve heard, that $10 billion is exploratory—Meta experimenting with TPU. If it were truly meant to cover the next three years, $10B wouldn’t be nearly enough. But it’s big enough for Google to hopefully give Meta preferential TPU access. Still, it’s puzzling why Meta feels the need. They have no TPU expertise. All the PyTorch community collaborates with NVIDIA or AMD; Google is a distant third—or even fourth. Honestly, Apple probably has better PyTorch support now than TPU. From the outside, it looks baffling. I also worry Meta might waste effort again trying to design ASICs for GPGPU workloads. But even then, I don’t think those (old) efforts would fold into MSL. If Meta goes deeper into ASICs, MSL would likely staff up independently. Outsiders can watch that—whether Meta really intends to build their own GPU-like chips.

T: People completely underestimate how hard ASICs are. Especially if you want them competitive with NVIDIA on total cost of ownership. Look at Amazon’s junk—Inferentia, Trainium—nobody wants it. Then there’s connectivity, thermal issues, even basics like heat dissipation. None of it is trivial. It takes multiple painful iterations. Google spent almost a decade and seven TPU generations before they got something halfway decent. Sure, maybe you hire ex-Google engineers and cut some corners, but even then—how do you secure foundry allocation? I don’t know.

S: Maybe as investors we overestimate NVIDIA’s moat. Sure, they dominate on chips, but they don’t monopolize the technology itself. At least four companies—NVIDIA, Google, AMD, Apple—can design and release competitive chips within a generation.

T: On paper, maybe. AMD looks competitive on paper. As for Apple—I don’t really know. You tell me, is it real or just on paper?

S: I’m pretty sure Apple and AMD’s next gen will be competitive. Both are releasing within three months. As for Google, I expected TPU v7 in July or August.

T: They just shared specs at Hot Chips.

S: Okay, that matches. Then v7 should launch this month, though I haven’t seen it. We’ll check. So yes, all these companies can do competitive single-node chips. As investors, we might overestimate NVIDIA’s moat—they’re just in a comfortable spot now.

T: True, but now it’s not just about chips—it’s rack-level solutions: scaling up, out, across. AMD and Apple definitely aren’t there yet. Google maybe, but not the others.

S: Ecosystem matters, sure. But zoom in—NVIDIA can only afford one misstep, maybe one generation (~18 months). That’s luxurious. In this cut-throat industry, one generation is survivable, two is not. So yes, they dominate now, but they’re not unbreakable.

T: Let me push back. What about the B200? Honestly, its ramp was a flop. Supposed to mass-produce last fall, but delayed again and again. Only this quarter did CSPs start ramping. Definitely a misstep. But it also shows how insanely hard rack-level solutions are. NVIDIA figured it out—so from GB200 to GB300 it’s smooth. But for others with no rack-level experience, the learning curve will be brutal. They’re not just “a generation away” from NVIDIA. No way.

NVIDIA

S: From NVIDIA—that’s not what I meant. I meant NVIDIA can afford one generation of misstep. In chip world, that’s ~18 months. So one and a half years, give or take. Anyway, let’s move on to NVIDIA’s story. They just released their quarterly earnings.

T: Yeah. Headlines say results were slightly below buy-side expectations, but I think it was a very strong report—clean beat and clear re-acceleration. Strip out China noise, and revenue was up ~15% QoQ. Guidance is for 17–19% QoQ next quarter. Look at networking: two quarters ago, people panicked that Ethernet ASICs were taking share. But this quarter, NVIDIA’s networking revenue blew out expectations. Margins too—two or three quarters ago gross margin fell from mid-70s to low-70s, sparking fears. Now it’s climbing back toward mid-70s, exactly as management guided: biggest product ramp ever, margins dip, then recover. I’m not the biggest fan of NVIDIA’s CFO—she could explain better—but at least management is being transparent.

S: But NVIDIA is even offering discounts on B300 and GB300 in promo emails—what’s going on? And I’m suddenly getting cold emails and calls from Supermicro sales. Last year, early on, you couldn’t even get through to them—they were overwhelmed with demand. Now they’re chasing customers. That’s a big shift.

T: Well, Supermicro is its own story. After their accounting fiasco, they’ve been suffering on both the supply and demand side. So I’d bet this is more of a Supermicro issue than an NVIDIA issue. Also, I haven’t heard of anyone—besides CoreWeave—getting real B300 allocation yet. No idea what’s going on. Are you on some secret pre-release list or something?

Unemployment

S: What about tariffs and unemployment? And now rates are going lower. California unemployment was 5.5% in August, I think. What’s happening there?

T: First, summer unemployment numbers are noisy—tourism, summer jobs, volatility. I wouldn’t put much weight on them. Second, honestly no one knows what’s going on with policy. On one hand, the OBBB stimulus is still running into next year. On the other, tariffs are clearly inflationary. Back in April we said “nobody knows,” but now it’s obvious: tariffs are inflationary. The only real question is whether it’s transitory or lasting. I hate the word “transitory”—Powell in 2021 scarred it—but no one really knows. The bigger question: can growth accelerate enough to outrun the inflation?

S: I mean, we've been 2% inflation for extended time before COVID. So that's what it's sort of conditioned us to that thought, like 2% inflation year over year, but that's not normal, right?

S: We had ~2% inflation for years pre-COVID, so people got conditioned to think that’s normal. But it’s not, right?

T: Yeah. I asked Gemini to do deep dives comparing 1990s vs 2000s inflation. Results were volatile. On average: early 2000s ~2.5–3%, 1990s ~2.7–3%. Definitely higher than 2%. What matters more is real GDP growth. That’s why Powell is leaning growth-oriented, not just inflation-targeting. If we can hold real GDP at 2.5–3%, we’re fine.

S: I think we’re fine, and markets agree. Housing’s moving. I even saw mortgage rates back in the high-2% range—with some strings attached of course. Markets are anticipating Fed cuts.

T: How is that possible?

S: Honestly no idea. My worry is the Fed chickens out and cuts rates, but the 10-year doesn’t follow.

T: Well, since Jackson Hole, the 10-year’s down ~10bps, from ~4.3% to ~4.2%. Not huge, but not nothing. Still, it’s hard to read. Lots of noise: courts ruling Trump’s tariffs illegal, Trump pushing appeals, etc. I think tariffs will stay.

S: I think so.

T: Yeah.

S: Exactly. It’s a cornerstone of Project 2025, alongside federal overreach. Tariffs will stay—high tariffs for a long time. Inflationary, yes. But will it be one-time or slow burn? My thesis: slow burn. It doesn’t feel elastic. Just vibes, but that’s my read.

T: Talking about vibes, right? Yeah.

GPT-5

S: We haven’t talked about the GPT-5 release yet.

T: Oh, I haven’t even tried it. I’m happy with Gemini. Every query I send it, it does deep research. Overkill sometimes, but it’s cheap, fast, and not rate-limited. I honestly haven’t used OpenAI or ChatGPT in half a year.

S: Actually, GPT-5 is a good release. You no longer pick models—it just works. It has fresher built-in knowledge, so recent questions are answered directly without search. Search always made ChatGPT’s worldview conflicted—what it “knew” vs what it just scraped. GPT-5 has neural knowledge, so answers are cleaner. And ChatGPT’s app is better than Gemini’s—Gemini still has bugs, like losing your entire conversation if you background the app. Basic stuff. People dismiss GPT-5 too quickly. It’s going to serve close to a billion people. People underestimate OpenAI.

T: Who said that? No one I know. Everyone I know has been trying to invest in OpenAI through SPVs, even paying $500B valuations while they’re officially raising at $350B. Who exactly are you talking to?

S: I don’t know. Maybe the people I talk to are too forward—living in the future. But Google’s doing great too. Gemini 2.5 Pro is a very good model. They’re not releasing Kingfall, but 2.5 Flash Image—nicknamed Nano Banana—is excellent at image gen and editing. Lots of strong models out there.

T: We’re living in different worlds. Every investor I know worries Google has structural issues—losing search moat, not aggressive enough against OpenAI, risk of becoming the next Yahoo. Totally different worldview from yours.

S: Maybe some truth. But little things matter. I stopped using Gemini for quick questions—I just use Google Search + ChatGPT. Search has AI mode now, which helps, but sometimes it flips back to classic search randomly. Gemini always insists on searching first. GPT-5 is better for direct answers—it just gives them, fast. Those little things matter. And inside Google, no one builds careers by fixing little annoyances.

T: By making customers actually love the product.

S: Exactly. In big companies, people don’t build careers by making the product nicer for users. They build careers by moving metrics.

T: My experience is the opposite. One day I uploaded a math problem photo to Gemini—it solved it fast, correctly, though the method was clunky. I tried the same with ChatGPT: 4.0 gave a wrong answer, 0.3 froze for ten minutes then gave another wrong one. That experience was so bad I stuck with Gemini. But maybe you’re right, maybe I should give GPT-5 a try.

S: Yeah—and if you use Nano Banana, it’ll even write the answer directly on the picture.

T: On the picture? Oh man. And it wasn’t even a difficult math problem—just a normal one.

S: Yep, surprising. OK, let’s end on that note.