The Fortress and the Gap

Fifty-four nations have signed on to FORGE, a US-led bloc to diversify critical mineral supply chains away from China ^[1]. Thirty-four more have joined Pax Silica, a State Department logistics platform designed to track and expedite shipments of semiconductors and critical minerals from trusted suppliers, explicitly to reduce reliance on China ^[2]. Project Vault, a strategic minerals reserve, is taking shape. Texas data centers have requested 466,497 megawatts of power — more than five times the state grid's historical peak ^[3] — and Meta is building ten natural-gas plants generating 7.5 gigawatts for a single Louisiana facility, with GE's turbine orders sold out through 2029 ^[4]. This is not a policy debate. It is concrete, steel, and gas. Washington has made a deliberate pivot from containing Chinese AI at the border — blocking models and chips — to fortifying the physical supply chain that powers it. The scale is real, the funding is bipartisan, and the momentum is undeniable. The problem is that the government's own advisory commission has said, in plain language, that none of this addresses where the contest has actually moved. In March, the US-China Economic and Security Review Commission issued a report that drew a distinction Washington's strategy had not acknowledged.

US export controls primarily target the digital loop, restricting access to advanced chips used for frontier model training — but are not well suited to addressing the physical loop of deployment-driven data creation and accumulation across China’s manufacturing base. — U.S.-China Economic and Security Review Commission

The same report found that approximately 80 percent of US AI startups now use Chinese open-source base models.

Open model proliferation creates alternative pathways to AI leadership. — U.S.-China Economic and Security Review Commission

The commission's vice chairman, Michael Kuiken, described the dynamic in starker terms.

There’s a bit of a deployment gap in the embodied AI space between the US and China. That’s something that over time compounds itself ... We’re starting to see that compounding now. — Michael Kuiken

The fortress, in other words, was built for one war. The other side is fighting a different one. Look at what has accumulated since the commission's report. DeepSeek, the Hangzhou-based lab, has established a dedicated team and updated its V4 Pro model specifically to build AI agents that compete with Anthropic's Claude Code — the tool that currently leads US business AI adoption at 43.5 percent of companies ^[5][6]. In May, DeepSeek permanently cut V4-Pro prices by 75 percent, leveraging Huawei Ascend processors to achieve costs 12 to 19 times lower than OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7. The company described the cut as a structural efficiency gain, not a promotion. It is now developing its own custom inference chips, reducing its dependence on both Nvidia and Huawei. Zhipu AI, placed on the US entity list, trained its GLM-5 and GLM-Image models entirely on domestic Huawei Ascend chips and released them as open weights ^[7]. Chinese open-weight models surpassed American models in global downloads on Hugging Face in January. The deployment layer is not a future battlefield. It is the present one, and Chinese models are cheaper, open, and increasingly hardware-independent. The US response at this layer is a different kind of inventory. The House has launched investigations into individual American firms — Cursor and Airbnb — for using low-cost Chinese AI models. Individual agencies have issued bans that experts describe as often ineffective, because AI is embedded in modern software and employees routinely bypass restrictions. The White House's own AI adviser, David Sacks, has dismissed deployment-layer restrictions in language that makes the administration's internal divide explicit.

the leading closed labs, already a duopoly in terms of AI model revenue, want the government to eliminate their open-source competition. — David Sacks

Legal experts have concluded that a blanket ban on Chinese AI models is likely unconstitutional under the 1965 Supreme Court ruling in Lamont v. Postmaster General, which protects the right to receive foreign materials. Even if a ban were legally feasible, open-weight models can be downloaded and hosted on private servers, making enforcement impossible. So the deployment-layer response is reactive, fragmented, and politically contested — investigations after adoption, bans that leak, a White House adviser calling the whole project corrupt, and a constitutional barrier to the one tool that might match the scale of the problem. Then there is the dynamic inside the American camp itself. Anthropic, whose Claude Code dominates US enterprise AI, quietly reduced session limits during peak hours in March, causing developers to exhaust quotas in minutes. The company also blocked third-party open-source tools from accessing Claude under flat-rate subscriptions, pushing enterprises toward proprietary Anthropic products. The effect is not subtle. At the precise moment Washington is debating whether and how to restrict Chinese AI models, the leading US lab is making its own tool more expensive, more restricted, and harder to access — while DeepSeek offers a cheaper, unguarded alternative at one-nineteenth the cost, with no session limits, and is building agents to compete directly with Claude Code. The US's dominant enterprise AI provider is pushing developers toward the Chinese tools Washington cannot govern. The commission saw this coming. Its report was not a warning about Chinese capabilities — it was a warning about American architecture. The controls were designed for the digital loop: chips, training compute, model weights. The competition has moved to the physical loop: deployment, data creation, the layer where 80 percent of US startups already run on Chinese open-source models. The fortress is real, funded, and bipartisan. The commission has said, from inside the house, that it does not address the layer where the contest is now being decided. The compounding Kuiken described is no longer a forecast. It is in the record.