AI Power Grabs: Governance Fractures, Infrastructure Strains, Market Nerves

Three distinct governance failures crystallised this week, each at a different jurisdictional level but sharing the same structural diagnosis. In the US, the DOJ's intervention against Colorado's AI law advances a deregulatory federal pre-emption strategy that dismantles state frameworks without providing a federal replacement — leaving the most substantive enacted AI law in America vulnerable to invalidation in a regulatory void. In China, the NDRC's one-sentence Manus blocking order demonstrates the opposite pathology: governance by opaque executive fiat, asserting extraterritorial authority over corporate structures without articulating a legal basis. In London, the Metropolitan Police's Palantir deployment against its own officers proceeded without disclosed legal basis, public consultation, or enforceable regulatory oversight — the fallout managed through mayoral politics rather than institutional accountability.

The EU AI Omnibus negotiations, still unresolved on core scope questions, and the UK government's inter-departmental incoherence on AI infrastructure energy planning add further evidence that Western democratic governance systems are struggling to match the operational tempo of AI deployment in both public and private institutions. Google's classified Pentagon agreement — whose terms are shielded from standard procurement accountability — and frontier labs' self-managed model restriction decisions (Mythos, GPT-Rosalind) both illustrate that consequential AI governance choices are being made outside formal oversight structures, with the character of foreign policy decisions but none of the interagency process.

This week delivered three simultaneous stress tests for the AI infrastructure investment thesis. OpenAI's reported revenue and user growth miss triggered a market-wide selloff spanning semiconductors, cloud platforms, and AI-native infrastructure — confirming that hundreds of billions in committed capital are effectively co-insured by the commercial performance of fewer than five frontier model companies. Energy infrastructure emerged as the new binding constraint: Wood Mackenzie's projection of $65 billion in US power equipment spending by 2030 and Oracle's 2.45 gigawatt fuel cell plan in New Mexico reflect an industry in which grid interconnection queues, not chip availability, now set the pace for capacity expansion. And the Strait of Hormuz closure introduced a third vector: a conventional military conflict degrading photoresist and helium supply chains across Asian chipmakers, exposing a chemical input blind spot that onshoring strategies and export control frameworks were never designed to address.

Supply chain depth data adds nuance: PCB makers, passive component suppliers, and high-bandwidth memory producers are all reporting strong demand, with some Asian supply chain companies posting triple-digit revenue growth. SK Hynix and Samsung are pushing customers toward long-term contracts as HBM shortages persist. This breadth of demand is a sign of infrastructure cycle maturity — but also of synchronised fragility. A disruption at the commodity chemical, power equipment, or second-tier component layer would propagate through the entire stack with less warning and slower remedy than an advanced-node chip shortage. Geopolitical fragility compounds the picture: a confirmed pause on Middle East data centre investment by an Oaktree-owned operator illustrates that the Gulf buildout — positioned as a major growth market — is directly exposed to regional conflict risk.

The US-China AI rivalry moved from rhetorical to structural this week across three simultaneous developments. Beijing's forced unwinding of Meta's Manus acquisition establishes that Chinese-origin AI agent technology is a controlled outbound asset regardless of corporate domicile — a symmetric mirror to CFIUS that now makes cross-border China AI M&A structurally impaired at both ends. Concurrently, the hardening norm of frontier model restriction among US labs — with Claude Mythos and GPT-Rosalind being held back on dual-use grounds — is creating a de facto access regime for advanced AI that maps onto geopolitical alignment: allied institutions may receive access, adversaries will not, and neutral states face pressure to choose. China's cybersecurity establishment is already treating Mythos as an inbound offensive threat vector, reflecting that both sides now understand frontier models as instruments of national power.

China's Lingshen exascale supercomputer announcement — reportedly the first exascale-class system built entirely from domestic Huawei Kunpeng CPUs with no foreign components — adds a compute sovereignty dimension that the US export control strategy did not anticipate. Rather than freezing Chinese AI capability, controls appear to have accelerated architectural divergence: a CPU-only exascale path, Huawei's Ascend accelerator ecosystem, and RISC-V adoption across Chinese chip design collectively suggest a parallel compute stack maturing faster than Western analysts projected. The strategic risk is not Chinese parity on GPU benchmarks — it is sufficient compute sovereignty to sustain frontier AI research and military applications independently, at which point export controls lose most of their coercive leverage.