1/12 I spent yesterday re-reading the TPUv7 deep-dive from SemiAnalysis and had to put my phone down twice. One sentence in particular broke my brain: Anthropic just signed a $52B deal to buy TPUv7 directly from Google. That’s not a cloud contract. That’s a lab buying its own nuclear weapons. 2/12 Everyone (me included) has been saying “NVIDIA owns AI hardware” for two years like it’s gravity. Turns out gravity only works until the biggest customers decide they’re tired of paying 75% margins to one company. 3/12 Here’s the part people still don’t get: Google didn’t beat NVIDIA at raw performance. TPUv7 is “only” 20–30% faster on paper. They beat them on price-per-token by ~50% at system level. That’s not incremental. That’s the kind of gap that ends empires. 4/12 Think about it this way: In 2023, if you wanted to train a frontier model you basically sent a blank check to Jensen. In late 2025, Anthropic just sent Google a $52B purchase order and cut their inference bill by two-thirds overnight. That’s not competition. That’s defection. 5/12 And it’s spreading. Meta is running TPU pilots. xAI is suspiciously quiet about their next cluster. Every lab is now doing the same math: “Why am I paying NVIDIA’s 75% margin when I can pay Google’s 30% margin and still get better uptime?” 6/12 The craziest detail: Google is now selling the chips directly. Not just renting hours on GCP. Actual metal. Ship it to your own colo. They turned the thing they built for Search/Docs/YouTube into a product anyone with a billion dollars can buy. That’s like if AWS started selling their custom Graviton CPUs on DigiKey. 7/12 NVIDIA’s only real defense left is CUDA lock-in. But the labs have figured out that rewriting in JAX/PyTorch-XLA costs a few engineering quarters once—and saves hundreds of millions forever. At that scale, the math is brutal. 8/12 So here’s where we are today: The frontier labs have become the new sovereigns. Google and NVIDIA are now just uranium enrichers bidding for their business. The customer rebellion already happened. We just didn’t notice because it was quiet and happened in purchase-order form. 9/12 If you’re a startup founder still defaulting to CUDA “because ecosystem” in 2025—stop. You’re voluntarily paying a 100% tax that no longer exists. Port to JAX this quarter and your series B can train a model twice the size. 10/12 If you’re an investor still modeling NVIDIA at 70%+ margins through 2030—update the spreadsheet. The new equilibrium is ~55% and falling. Still absurdly profitable. Just not immortal. 11/12 The simplest way to say it: AI compute just went from “scarce luxury good” to “commodity you can buy by the pallet from Google.” Everything else (price, roadmap, moats) follows from that single fact. 12/12 Next time someone drops a new SOTA benchmark, don’t ask “how big was the model.” Ask “whose silicon made that cheap enough to run?” The answer is increasingly “not NVIDIA.” And once you start asking that question, you can’t unsee what’s happened.