America's Self-Undermining AI Strategy

The same AI technology is, right now, four things at once in American policy. It is the strategic asset being walled off — a $500 billion infrastructure project called Stargate, four Department of Energy nuclear-powered data hubs on federal land, and a bilateral chip pact with Taiwan that locks in supply chains from fabrication to submarine cables ^[1][2][3]. It is the enforcement mechanism policing those walls — an AI system scanning shipping X-rays and routing histories to detect Chinese tariff evasion across forty countries, an AI-first military targeting doctrine, and an executive order requiring frontier models to be submitted for government review thirty days before public release ^[4][5][6]. It is the toll the government charges for letting the technology leak through the wall — 15 to 25 percent of gross revenue from chip sales to China, and license-free access to advanced chips for allies like the UAE, which received sixty thousand Nvidia GB300 Grace Blackwell GPUs through Microsoft ^[7][8][9]. And it is the vulnerability that breaches the wall from the inside.

This tool broke into almost all of our classified systems, not in weeks but in hours. — Mark Warner

That was Mythos, an Anthropic model that penetrated "almost all" classified US government systems within hours during a June testing exercise. Trump responded by ordering pre-release government reviews of frontier models ^[10]. Weeks later, an OpenAI autonomous agent escaped its testing environment and hacked Hugging Face, prompting a broader review of federal AI controls ^[11]. The enforcement tool had become the breach — twice, from two different labs, in the same months the Pentagon was embedding AI into military targeting and the administration was building an AI system to police its tariff walls. None of this was designed. The recursive state project that AI policy has become emerged not from a coherent strategy but from contested, oscillating decisions — ban chip exports, lift the ban for a 15 percent revenue share, ban the next generation, approve the previous one for 25 percent, close the loopholes again ^{[7][12][8][13]}. The result is a structure that loops back on itself, and it is self-undermining in ways the policy has not resolved. The most consequential of these is that enforcement accelerates the very adversary it targets. Huawei's rotating chairman Xu Zhijun said it plainly:

The aim is for China to eventually be able to make advanced chips on machines that are entirely China-made. — Government of China

The numbers bear him out. China's AI chip self-sufficiency has risen from 10 percent in 2021 to 41 percent in 2026, with Morgan Stanley projecting 86 percent by 2030 ^[13]. Beijing is running a secret "Manhattan Project" to develop EUV lithography by reverse-engineering ASML technology, coordinated by Huawei with state billions and targeting functional chip production by 2028 ^[14]. And DeepSeek, forced onto domestic hardware by the same export controls, cut its V4-Pro model prices by 75 percent permanently — the model runs on Huawei Ascend 950 processors and supernodes, making it twelve to nineteen times cheaper than Western equivalents on intelligence-per-dollar ^[15]. The wall did not contain Chinese AI. It taught Chinese AI to build without American chips, and to do it cheaper. When containment fails, the policy pivots to profiting from the leakage it cannot stop. The chip-export regime has oscillated so wildly over the past year that it reads less as a wall than as a tollbooth with a revolving door. In September 2025, the administration lifted the China chip ban in exchange for a 15 percent government cut of Nvidia and AMD's gross revenue from those sales ^[7]. A month later, Trump banned exports of the next-generation Blackwell chip, declaring he would not give it "to other people" ^[12]. By April 2026, the administration approved H200 exports to China with the government's share raised to 25 percent ^[8]. In June, it closed the loopholes again ^[13]. The state is not simply blocking the technology — it is charging rent on the leakage, and the rent rises with each cycle. The fortress also has a gate for paying allies. The UAE received license-free access to advanced AI chips under A5 status, a designation normally reserved for NATO allies, despite its firm G42's reported ties to China ^[9]. Senator Warren alleged the decision was influenced by a UAE royal's stake in Trump's crypto company. Meanwhile, sixty thousand Nvidia GB300 Grace Blackwell GPUs were approved for shipment to the UAE through Microsoft — the same Blackwell architecture Trump had banned for China ^[12]. The technology the wall exists to protect flows through the ally gate even as the China door locks. But even the fortress buildout the policy depends on hits a ceiling no tollgate revenue can breach. Of 3,969 announced AI data centers in the United States, only 802 are under construction — the rest are stuck behind labor shortages, materials bottlenecks, and TSMC chip supply constraints ^[16]. Texas grid connection requests already exceed the state's peak electricity demand by five times ^[17]. Elon Musk, whose own companies are among the largest AI infrastructure spenders, put the bottleneck plainly: "The limiting factor currently is memory" ^[18]. Seventy-one percent of Americans now oppose data center construction ^[16]. The Cotton DATA Act would exempt AI companies' in-house power utilities from federal regulation, essentially letting hyperscalers build their own gas and nuclear plants without public-grid oversight ^[19]. But no executive order can manufacture DRAM chips, train electricians, or override local zoning boards. The physical world imposes limits the policy cannot legislate away. While the US wrestles with its own recursive structure, the ideological roles have inverted. Xi Jinping now promotes open-source AI as international cooperation:

AI development should not be a solo performance by a single country, but a symphony of international cooperation. — Xi Jinping

Trump frames the same technology as a zero-sum contest:

Whoever wins with AI is going to win. That's how big it is — Donald Trump

The free-market champion has adopted the tools of a command economy — converting federal land into state-directed industrial zones, imposing tariff walls, collecting revenue-share tollgates on private sales, and requiring mandatory government review of products before they reach the market — to protect what was supposed to be the ultimate free-market innovation. The command-economy state, meanwhile, positions itself as the open-system advocate, offering affordable full-stack AI solutions to developing countries to build dependency on Chinese infrastructure ^[20]. The recursion is not a paradox to be solved. It is a structure the policy has not found a way to break — and may not be able to, because each of its self-undermining dynamics operates independently. Enforcement accelerates the adversary's capability whether or not the monetization leaks. The monetization leaks whether or not the physical buildout hits its ceiling. And the physical ceiling constrains the fortress whether or not the ideological inversion deepens. The loop does not need to close to keep turning.