The US government has moved from theoretical concern about frontier AI to operational control over its deployment. Two weeks apart, the Trump administration ordered Anthropic to take Mythos offline for foreign users, then asked OpenAI to delay GPT-5.6's general release. Both companies complied. The stated rationale is national security, but the mechanism is pure regulatory discretion: no law, no formal process, just government requests that major AI labs treat as directives. OpenAI's public pushback, claiming restrictions "shouldn't be the norm", rings hollow when the company is simultaneously releasing GPT-5.6 to a government-vetted subset of users. The administration has granted Anthropic permission to distribute its model to over 100 US companies and agencies, including their non-American employees, which suggests the concern is less about capability leakage and more about who gets to decide who uses what. This is regulatory capture dressed as security policy, and it works because the companies have no real alternative. They need government approval to operate at scale. OpenAI's announcement that it's building Jalapeño, its custom inference chip with Broadcom, looks less like technical independence and more like insurance against future restrictions, a way to own the supply chain when the government controls the distribution chain.
The real shift is territorial. South Korea is training half a million soldiers as drone operators. China's Tencent is embedding DeepSeek's models into WeCom, its enterprise collaboration tool. Europe wants its own AI, explicitly leveraging Trump's protectionist posture as an advantage. The EU has flagged AWS and Azure as gatekeepers under the Digital Markets Act, which will force Microsoft and Amazon to open their cloud platforms to competitors. Meanwhile, SpaceX's bonds are selling off, yields moving toward junk territory, despite a $25 billion debt deal, suggesting markets are pricing in risk around Elon Musk's sprawling operations. The pattern is fragmentation. Instead of a unified global AI infrastructure centered on US companies and Nvidia chips, you're seeing nation-states and blocs building their own stacks, training their own workforces, and using regulatory leverage to carve out protected markets. This isn't competition; it's de facto sanctions wrapped in procurement policy.
The enterprise layer is quietly reorganizing around two premises: that public cloud AI is too expensive and too exposed, and that custom infrastructure beats generic platforms. Enterprises are merging their OLTP and OLAP storage to feed AI agents real-time operational data. Microsoft is turning Windows into an AI operating system, promising unmetered local inference so companies can run models for free on their own hardware. Apple is raising prices up to 25% to cover memory costs, which means AI is now a line item in hardware budgets, not a service. The companies building their own chips, OpenAI, Google, Apple, SpaceX, are not trying to compete with Nvidia in general-purpose compute. They're optimizing for inference at scale, which is where the margin and control live. This is the economics of vertical integration in a world where government can shut down your API access with a phone call.
Sloane Duvall