If the US Wins the AI Race, It Gets More Dangerous: China's Optimal Move Is to Blow Up TSMC
If the US achieves overwhelming AI dominance, China's game-theoretic optimal move is to blow up TSMC. Silicon Valley won't admit that America's dependence on Chinese supply chains is badly underestimated — and that dependence won't be fixed outside of a conflict.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
If the US wins the AI race, it gets more dangerous
Ben Thompson's answer runs opposite to the mainstream Silicon Valley narrative: if the US achieves overwhelming AI dominance, China's game-theoretic optimal move is to blow up TSMC. He thinks the reasoning itself isn't complicated; what's complicated is that Silicon Valley won't admit how badly America's dependence on Chinese supply chains is underestimated — not just fabs, but actuators and precursor materials. That dependence won't be fixed outside of a conflict, because as long as your competitors are still buying from China, moving back to the US on your own puts you at a relative disadvantage. The rational choice is to wait until "literally no choice" before moving.
— Ben ThompsonChina being six to nine months behind is a decent equilibrium
The current picture: OpenAI and Anthropic are clearly at the frontier, Google's situation is unclear, and Grok and Meta are chasing. China's capabilities are strong, and distillation keeps it roughly six to nine months behind. Ben thinks this equilibrium is broadly favorable to the US; the question is only how long it holds. He doubts China can truly overtake, because the last six to nine months are extremely hard to cross, and the acceleration of AI using AI to improve AI is already showing up at OpenAI and Anthropic. He also pushes back on the claim that open-source models are free: the inference costs of GLM and Kimi are real, Kimi's serving costs are high, and the cost per answer is significantly higher.
— Ben ThompsonWhere the money comes from is the nearest bottleneck
What worries Ben is a timing mismatch: whether the revenue from investment returns arrives in time to meet the investment. The capital curve is moving downward — first free cash flow, then tech companies burned through the debt market in a single year, then Google started issuing stock, and Nvidia is putting together a $500 billion thing to tap pensions and insurance float. He asks: what comes after that? Ideally you return to cash-flow funding, but if there's a gap in the middle and it can't be bridged, there could be a big bang. Even if it blows up, though, AI won't disappear and won't stop improving — just like the railroad and internet bubbles, on a human scale it ultimately doesn't matter.
— Ben ThompsonGoogle is becoming a Berkshire-style absolute-numbers player
Ben looks at the Nvidia deal and Google's stock issuance together, and Google's issuance shocked him: why would a company that trusts its own judgment dilute its upside? The Berkshire analogy is See's Candies, with high margins but a ceiling on absolute profit and no reinvestment runway, versus BNSF, with far worse margins but a far larger absolute amount — a year of free cash flow exceeding See's entire lifetime. Google Search is the most perfect business model in history, the purest aggregator, zero marginal cost, while AI burns money but its TAM is all white-collar work, eventually plus robots. Using all your free cash flow, issuing debt, then issuing stock, you get a smaller slice of an astronomically larger pie.
— Ben ThompsonMemory makers turned themselves into the Strait of Hormuz
Ben compares memory makers to Iran: the Strait of Hormuz's deterrent value lies in not using it — once you actually use it, the UAE and Saudi Arabia will build pipelines and new ports, and a blockade in 2035 would be useless. His worry is that memory makers have already used up that deterrence — it hurts now, but no one will let themselves get into that position again, and the first priority on the algorithm side becomes "how do we use less memory." Likewise TSMC has only one leader, and customers will route around it.
— Ben ThompsonThe shortage saved Intel, not technology
Ben says that in an unchanging world TSMC would win forever, but precisely because TSMC didn't invest in recent years, the shortage will become extreme enough that big tech companies will tolerate the pain of propping up Intel's and Samsung's logic processes — "scarcity ultimately saved Intel." He expects Intel to announce a major partner for the first time. The solution to the geopolitical problem isn't reasoning with people, but building a compute use case big enough that everyone has an economic incentive to prop everyone else up, so that geopolitical insurance comes free.
— Ben ThompsonApple is the king of determinism, AI is a probabilistic business
Ben says Apple's moat is controlling the customer entry point, so suppliers come to it, which means AI suppliers can be bought in; in the future it might even rely on on-device inference using the user's electricity, without paying inference costs. But he questions why anyone expects Apple to be good at AI: AI is fundamentally probabilistic, and Apple is the king of deterministic products — an iPhone has to be good the first time it ships, and Apple has never recalled an iPhone. He'd rather Apple do what it's good at; it's fine if it doesn't touch AI.
— Ben ThompsonNvidia's moat is being discounted in ways you can't see
Ben says Nvidia's position is "unnatural": it provides 25% backstops to NeoClouds, takes equity, and commits to buying through 2030 — essentially lowering the other side's cost of capital, meaning it has taken on the risk itself. Risk doesn't disappear, it just shows up somewhere else. If compute is oversupplied and nobody wants those GPUs, Nvidia loses money — discounting that back at expected value amounts to a price cut, just one that doesn't show up in gross margin. The real long-term rival is the hyperscaler: Google has already sold about 20% of its TPUs to Anthropic, and Amazon has hinted it will sell Trainium externally too, and as a commodity rather than a differentiator.
— Ben ThompsonA power shortage is actually good for Nvidia
Nvidia won't say it out loud, but if the world really runs out of power, that's probably good for it: when power is the binding constraint, the only way forward is maximum token efficiency, and Nvidia is still the most token-efficient. Conversely, Nvidia's biggest problem over the past two years has been that the US brought far more power online than expected — whether behind-the-meter generation, West Texas natural gas, or restarted nuclear plants. The more abundant the power, the more time Amazon has to get Trainium right, and the more time Google has to make TPU competitive on efficiency.
A bubble has to leave behind something that pays for a long time
If this really is a bubble, what you want is a long-term return. What the internet bubble left behind was fiber buried in the ground — Google's business is largely built on dark fiber bought for almost nothing after the bubble, and the core internet still runs on WorldCom's fiber; what the railroad bubble left behind was assets like BNSF that are still paying Google today. So what will AI leave behind? GPUs won't last that long, data centers barely count, and what really lasts is power. If all this blows up and leaves excess power behind, that's a wonderful world — humanity has always been constrained by energy scarcity, and a world of energy abundance is hard to imagine.
In their own words · checked verbatim
I think it would be very problematic for the U.S. to win.
Ben Thompson2:07
We're working our way down the capital curve. We started with free cash flow. The speed with which the tech companies blew through the debt markets is kind of incredible. It took like a year. And now Google's issuing equity.
Ben Thompson9:21
Risk doesn't disappear. It just moves.
Ben Thompson39:42
Risk doesn't disappear. It just gets handed off. And sometimes that risk doesn't manifest in losing money. It manifests in not making money.
Ben Thompson40:44
The scarcity is what ultimately saved Intel.
Ben Thompson43:47
The hyperscalers have always been the threat to NVIDIA for that reason. They're actually bigger.
if we get to a world where we actually run out of power, that's probably good for NVIDIA because in a world where we're totally constrained on power, we have to get the best efficiency, the best token efficiency. And I think NVIDIA is still the most token efficient, so that is a good world for them.
You want a bubble that produces something that lasts.
Figures
| Nvidia financing scale | $500 billion | 9:21 |
| Gap between Chinese models and the frontier | six to nine months | 6:11 |
| Nvidia's backstop ratio for NeoClouds | 25% | 1:08:36 |
| Share of TPUs Google sold to Anthropic | about 20% | 1:09:36 |
Glossary
- reluctant accelerationist
- Someone who is extremely bullish on AI while holding reservations about capabilities migrating into unverifiable domains.
- NeoCloud
- Emerging cloud providers built specifically for AI compute, distinct from the traditional hyperscalers.
- hyperscaler
- Cloud providers like Google, Amazon and Microsoft — enormous in scale, with extremely low cost of capital.
- token efficiency
- How many tokens you can produce per unit of power or cost; becomes the core metric when power is constrained.
How to listen
Founders, investors and engineers watching the AI capital cycle, semiconductor supply chains and geopolitical risk — especially anyone trying to judge compute and power constraints.
If you already know the railroad and internet bubble analogies, skip the stretch at 10:23.