US Government AI Gate Tightens as Capital Rotates to Physical Stack

The Trump administration's intervention in Anthropic and OpenAI release schedules has moved from political signal to operational reality this week. Mythos 5 is back online for over 100 vetted US organisations; GPT-5.6 is limited to a government-approved preview tier. Both labs are complying without statutory obligation, under a 30-day advance access window tied to an executive order. Treasury Secretary Bessent's engagement — triggered by bank warnings about frontier models threatening Federal Reserve payment rails — reveals that the security concern driving this regime is concentrated in financial infrastructure, not solely defence, which explains why the cleared list includes companies as well as agencies.

The downstream consequences extend well beyond the labs themselves. Data centre operators and enterprise IT teams that plan capacity around model deployment dates now face a government-controlled variable with no defined criteria, no statutory basis, and no predictable timeline. For buyers outside the cleared list — non-US multinationals, international governments, European enterprises — the effect is an immediate procurement constraint. China's Zhipu AI is filling that gap with GLM 5.2, which benchmarks near US frontier models at lower cost and without access restrictions. Europe, meanwhile, is hedging: Italy joined the US-led Pax Silica framework while domestic AI investment narratives are strengthened by the case that US-controlled access is an unreliable foundation for sovereign digital infrastructure.

Three converging pressures are straining the economics of AI infrastructure investment. First, AI distillation allows competitors to build capable systems at a fraction of frontier training cost, compressing the return on the hundreds of billions being committed to training compute. Second, Apple and Microsoft price increases driven by component cost inflation are rippling through Asian memory chip stocks, signalling that input-cost pressure may dampen the device demand cycle that has underwritten semiconductor capex expansion. Third, enterprise CFOs are imposing ROI discipline on AI spend that was absent during the 2023-25 deployment wave, creating a revenue headwind for frontier labs at the moment their infrastructure costs are highest.

OpenAI's Jalapeño custom inference chip and its reported IPO delay are both responses to this squeeze. Jalapeño is the mechanism for converting model capability into sustainable unit economics at inference scale — without it, serving a mass consumer and enterprise market at volume perpetually widens losses. The IPO delay reflects that OpenAI's private-market valuation requires a clearer profitability narrative before facing public scrutiny. The physical confirmation of Vantage's second Lighthouse campus building in Wisconsin is a needed reality check in the other direction: some announced hyperscale commitments are genuinely materialising. But Oracle's debt load and negative free cash flow attracting serious analyst scepticism suggests the infrastructure overbuild thesis is no longer fringe.

A consistent pattern is emerging across multiple capital allocation signals this week. Goldman Sachs is internally flagging the physical economy as the next AI investment wave. Onsemi's $7 billion acquisition of Synaptics explicitly targets a $30 billion addressable market in edge inference and power management for robotics and physical AI. Government-backed geothermal investment in the US and a 15-year vPPA in Japan both position firm baseload power as the critical enabling infrastructure for AI data centre siting. A Dogecoin-origin cryptominer is acquiring a 150MW natural-gas-powered AI/HPC site in Arkansas. These are not isolated deals — they reflect a shared thesis that the next margin pool in AI is in the physical layer, not the software layer.

The common logic is that physical AI infrastructure — power, cooling, edge compute, and industrial deployment hardware — offers more predictable returns than frontier model access fees, which are exposed to distillation, commoditisation, and now government access controls. Geothermal's near-90% capacity utilisation versus 25-35% for intermittent renewables makes it structurally attractive for co-located data centre power, and both the US drilling investment and the Japan vPPA reflect 15-year conviction horizons rather than speculative positioning. The Onsemi-Synaptics deal represents the first consolidation move explicitly designed to integrate the power and inference layers for physical AI platforms — a sign the edge market is transitioning from fragmented components to integrated platform competition.