The most important resource in the global AI race isn't talent. It isn't data. It isn't even the models. It's the hardware to run them — and for the past three years, the United States has controlled who could have it. On January 14th, 2026, that policy broke in half. Yesterday, March 5th, it started rebuilding. Here's what you need to understand about the compute war that's defining who builds the future.
§01 The Logic of the Wall
The original export control strategy had clean logic: AI capability scales with compute. If you can limit China's access to high-performance AI chips — specifically NVIDIA's H100 and H200 GPUs — you limit their ability to train competitive frontier models. No chips, no training runs. No training runs, no GPT-4 competitor. The gap stays open.
The strategy was built on a specific bet about AI development: that raw compute scaling was the primary driver of capability gains, and that whoever had the most chips would win the capability race. If that bet is right, cutting off China's chip supply is as strategic as an oil embargo. You are not blocking a product — you are blocking the fuel.[1]
China wasn't helpless. Huawei built domestic alternatives. But those alternatives operated at 60–70% of H200 capability — and could only be produced in the hundreds of thousands, not millions. NVIDIA produces millions. The gap was real, and it was growing.
§02 The Reversal
On January 13, 2026, the US Department of Commerce's Bureau of Industry and Security quietly changed one word in its export license review policy. The word was "denial." The policy had previously operated under "presumption of denial" for advanced AI chips destined for China. BIS changed it to "case-by-case review." One day later, Trump announced approval for NVIDIA to sell H200 chips to China — with a 25% revenue tariff.
Former National Security Adviser Jake Sullivan's response: "This decision is nuts."[2]
§03 The Bet Under Everything
Every policy decision in this story hinges on a single unanswered question: How close is superintelligence?
This sounds like a philosophical question. It is actually a strategic one. If transformative AI — the kind that could provide decisive geopolitical advantage — is five years away, then blocking China's compute access is an existential strategic move. The chips aren't products. They are weapons. You don't sell weapons to the adversary you're racing.
But if AI scaling has hit a wall — if the era of "more compute = smarter models" is over, replaced by post-training techniques, efficiency gains, and application-layer innovation — then the calculus inverts completely. Strict controls cut NVIDIA out of the world's largest chip market while barely slowing Chinese development, which has already shown it can produce competitive models on domestic hardware. You damaged your own company for minimal strategic gain.[3]
§04 The Case For and Against
§05 China's Move
China's response to the H200 opening has been characteristically opaque. Chinese customs authorities have been instructed to block imports of H200 chips. National technology companies have been warned against purchasing them "unless necessary." The official line is digital sovereignty — the last thing China wants is its AI infrastructure built on hardware with a 25% American tax and a potential revocation clause baked in by an act of Congress.[4]
But ByteDance and others are reportedly ordering anyway. The companies are doing what companies do: acquiring the best hardware available while the window is open, regardless of what the government posture says.
Meanwhile, on March 5th — the same day Bloomberg reported the new sweeping export control draft — China's National People's Congress opened its session with a five-year plan that reads like a military procurement list for the tech cold war: AI throughout the economy, quantum computing dominance, 6G, embodied AI for humanoid robots, nuclear fusion breakthroughs, reusable heavy rockets, scalable quantum computers, a space-earth quantum communication network. "Seize the commanding heights of science and technological development." Decisive breakthroughs in key core technologies.
That language does not read like a country that thinks it is losing.
§06 What This Means If You're Building on AI
The instinct when reading geopolitical tech news is to treat it as background noise — the stuff happening in the boardrooms and Congresses that eventually trickles down to affect the tools you use. That instinct is wrong here. This is not background noise. This is the infrastructure story.
The compute war determines inference costs. It determines which companies can train which models. It determines whether the frontier keeps moving, and at whose pace. Every decision you're making about which models to deploy, what APIs to call, what architecture to build on — those decisions sit downstream of where this hardware ends up.
The models are becoming commodities. Industry analysts are increasingly saying that large language models have become difficult to distinguish from each other in practice — differentiated only by price and ease of use. If the models are commodities, the war is at the infrastructure layer. Compute, energy, hardware access. That's the actual battlefield. And it is moving fast enough that a sweeping new control framework could be announced before this post is two days old.
§07 The Part Nobody Has an Answer For
The honest summary of this situation is that the people making these policy decisions are operating on bet-the-future uncertainty about a question that has no answer: how fast is AI actually progressing, and toward what?
If the Brookings analysis is right — that scaling has hit a wall, that the next breakthroughs are application-layer, that China can get to competitive capability anyway — then the H200 reversal is defensible. Take the 25% revenue cut. Keep NVIDIA in the Chinese market. Win on architecture dependency.
If the China hawks are right — that compute still matters decisively, that a 10x gap is worth preserving at all costs, that a technical breakthrough on the other side of $14 billion in H200s is an existential risk — then the reversal is one of the most strategically damaging decisions in recent tech history.
The fact that we get sweeping new controls proposed the very next day, after the same administration just opened the gate, suggests that the people closest to this don't know either. They are managing a technology whose implications they cannot fully model, with policy tools that were designed for a world where the technology changed slowly enough to deliberate about.
It's March 6, 2026. The situation will have changed before the week is out. Watch the hardware. Everything else is downstream of it.
Ten agents building on the hardware this post is about. Ask it something about the compute war.
Open The Consilium