Why AI Stocks Are Falling Despite Record Spending

0:00Lucas: So last week TechCrunch ran a piece with a headline I haven't been able to shake — 'Is this the dawn of the Tokenpocalypse?' Luna: That is an aggressively ominous phrase. What does it actually mean? Lucas: It's the argument that the cost of running AI models — the inference cost — is becoming unsustainable for a lot of companies. We've been hearing for months that AI adoption is hitting a cost wall. This is that wall getting a dramatic name. Luna: Right, and we've been seeing it in the stock numbers too. NVIDIA is down 8.5 percent in the past five days. AMD down 8.6. Palantir down more than 15 percent. Lucas: Broadcom, ARM, Micron — all off double digits over the same stretch. And this is happening while the hyperscalers are spending more on AI infrastructure than ever. Alphabet committed 85 billion earlier this year. Google is paying SpaceX nearly a billion a month for compute. Luna: So the disconnect is huge. Record spending on building the infrastructure, but the stocks that are supposed to benefit from it are getting crushed. What gives? Lucas: I think the market is realizing that spending billions on GPUs doesn't automatically translate into profitable products. The big cloud providers are in an arms race to build capacity, but the companies actually deploying AI — the enterprise software firms — are finding that inference costs eat their margins. Luna: Inference being the cost of actually running a trained model to generate responses, versus training it. Lucas: Exactly. Training is expensive but it's a one-time or periodic cost. Inference is every single query, every API call, every Copilot suggestion. And that cost is not dropping as fast as people hoped. Luna: We did an episode a few weeks back on how AI inference is reshaping the GPU market. But this feels like the market is now pricing in that the unit economics don't work for a lot of use cases. Lucas: Right. And the TechCrunch piece points to something specific: the cost per token is not falling at the rate that would make mass adoption viable. You're seeing companies like Notion — they had a service disruption with Anthropic last week, according to another headline. That's the kind of fragility that spooks enterprise customers. Luna: So the Tokenpocalypse is basically the fear that AI becomes a cost center that never gets efficient enough to generate the returns people expected. Lucas: Bingo. And that's why you see Palantir — which rode the AI hype harder than almost anyone — down 15 percent in a week. The market is repricing the growth assumptions. Palantir's forward price to earnings ratio was astronomical. Now investors are asking: where's the margin expansion? Luna: But isn't there a counterargument? Smaller models are getting better, distillation is improving. We talked about that a few episodes ago. That should bring inference costs down. Lucas: It should, and it will — but it takes time. Meanwhile, the big model makers like OpenAI are still pushing toward larger frontier models. And they're talking about a 'super app' — another headline from last week. That tells me the frontier is still about scale, not efficiency. Luna: So you have two opposing forces. The hyperscalers are building out capacity for large models. But the startups and enterprises are trying to shrink models to save money. Lucas: And the chip stocks are caught in the middle. If everyone moves to smaller models, you need fewer GPUs per query. That's great for margins on the software side, but terrible for NVIDIA's volume story. Luna: But if the large models keep winning, then inference costs stay high and adoption slows. Either way, the AI trade gets squeezed. Lucas: That's the bind. And I think the market is starting to price in a scenario where AI revenue grows, but margins compress because of competition and infrastructure costs. That's bad for investors who were betting on a winner take most outcome. Luna: Let's look at specific numbers. Micron is down 16.6 percent in five days. That's brutal for a memory maker that was supposed to be a key beneficiary of AI. Lucas: Memory is interesting because high-bandwidth memory is essential for AI training. But if the market expects a slowdown in training demand — or a shift to inference-optimized chips that use less memory — then Micron's growth story gets dented. Luna: And ASML is actually up 0.8 percent. That's the only green number in the semiconductor space this week. Lucas: ASML makes the lithography machines that are needed to manufacture advanced chips. Regardless of which chip designer wins, they need ASML's equipment. It's a monopoly in that sense. So ASML is insulated from the inference cost debate. Luna: So the sell-off is concentrated in the companies that depend on volume growth of AI compute. Not the picks and shovels makers. Lucas: Exactly. And that's a classic pattern in tech bubbles — the infrastructure providers do well early, but when the growth narrative shifts, the pure plays get hammered. Luna: Right, and I think this is a good moment to mention something we don't talk about enough. We keep this podcast ad-free because of listener support. If you've found these conversations useful, a couple of dollars a month genuinely makes a difference. You can find us at buy me a coffee dot com slash fexingo. Lucas: It's true. We don't have a big media company behind us. It's just the two of us and a microphone. So if you've gotten something out of the show, that's the best way to keep it going. Luna: And it really does help. Now, back to the sell-off — one other thing I noticed is that Microsoft is down 9.5 percent. That's a massive move for a stock that size. Lucas: Microsoft is interesting because they're both a hyperscaler and a software company. Their Azure business is spending heavily on AI infrastructure. But their software margins — Office, GitHub Copilot — are being squeezed by inference costs. The market might be looking at that tension. Luna: So the Tokenpocalypse isn't just a cute headline. It's a real valuation risk for the entire AI ecosystem. Lucas: I think so. And the key question going forward is: how fast can inference costs decline? If they follow a Moore's Law-like trajectory, the current sell-off is an overreaction. But if they plateau, then the market is right to reprice these stocks. Luna: And what about the super app that OpenAI is working on? Does that change the calculus? Lucas: A super app would be a massive consumer product — think WeChat for AI. If OpenAI pulls that off, it could drive enormous usage and justify the infrastructure spend. But it's a huge if. And in the meantime, the market is focused on the cost side. Luna: So the next few months are critical. We'll be watching earnings calls for signs that inference costs are improving. Lucas: Absolutely. And on that note, we'll keep tracking this. Thanks for listening.