The Export Control Paradox: How Americas Chip Ban Made China Faster and Left the Pentagon Scrambling

The Pentagon is negotiating to acquire a frontier AI model it successfully argued should be kept off defense contracts.

The contradiction captures the bilateral AI arms race more precisely than any policy paper. Anthropic built a model called Mythos capable of exploiting security flaws in every major operating system and web browser, a capability the company itself deemed too dangerous for public release. The Defense Department designated Anthropic a supply chain risk after the company declined to give the Pentagon full unrestricted access to its models, citing restrictions on mass domestic surveillance and autonomous weapons use. A federal appeals court declined to temporarily lift the designation. The Pentagon won in court. And then the talks resumed. Neither side is walking away.

The Anthropic contradiction is one half of a two-part problem. The other half is China, and it is the harder one. American export controls were supposed to slow China's AI. Instead, they may have accelerated exactly what they were designed to prevent. Three US defense and intelligence officials told The New York Times in April that the Pentagon has concluded America's program for unmanned combat drones is lagging China. The finding, coming five months after China showed autonomous drones at a Beijing military parade, landed inside a US defense establishment already reassessing how it competes. Companies like DeepSeek developed techniques to extract more performance from limited hardware, including sparse activation, memory optimization, and lower-precision training. These were survival strategies for a lab under sanctions. They also made Chinese AI more globally competitive. A model that works efficiently on modest hardware can be deployed by anyone, anywhere, without Nvidia chips or hyperscale data centers.

Export controls designed to constrain China ended up making Chinese AI harder to contain globally. And the military implications followed. The drones China displayed in September were not theoretical. The efficiency that DeepSeek proved could produce capable AI models also produced capable autonomous systems. The United States, which spent decades building a defense industry optimized for reliability and incremental improvement, suddenly needed to respond to a competitor that had compressed development timelines by necessity.

This is the bilateral AI arms race as it actually operates: not a single contest but two simultaneous pressures that work against each other. The US is trying to match China manufacturing efficiency while also negotiating through an internal conflict about what kinds of AI it will allow its own military to use.

Anduril, the defense startup, moved production of its Fury high-speed combat drone up by three months, starting in March instead of July, at its new Ohio facility. Fury uses commercial supply chain components, including aluminum rather than titanium and a business jet engine chosen partly because its maintenance network is already established. The company calls this baked-in manufacturability. The Pentagon calls it a way to close the gap before a capability deficit becomes a tactical problem. But speed on manufacturing does not resolve the governance problem. A drone that cannot be shipped because of a contract dispute is no use in closing a gap that was demonstrated five months ago.

Alibaba has committed to investing more than $50 billion in AI over the next three years. Microsoft spent roughly $80 billion on AI capital expenditures in 2025 alone. The four major American hyperscalers have plans to spend a combined $650 billion in 2026. The efficiency frontier has moved from pure compute to the ratio of capability to cost. Those numbers are not the constraint. The constraint is whether the technology can be built, governed, and deployed in a world where efficiency and safety are now in direct tension.

China did not face that tension in the same way. Its AI labs were constrained by hardware access, which forced efficiency. The resulting systems do not come with corporate restrictions on how they can be used. When Anduril looks at the competitive landscape, it is not just looking at Chinese drones. It is looking at a Chinese AI ecosystem that was shaped by American policy into something that is simultaneously more efficient and less constrained.

The Pentagon has acknowledged the gap. Anduril has moved to close it. The question is whether the United States can solve for efficiency and governance at the same time, in the same weapons systems, before the next parade.