AI Power Grab: Regulators, Rivals, and Critical Minerals Reshape the Stack

Two developments this week reinforce the same underlying dynamic: the AI infrastructure buildout is becoming structurally harder for anyone without deep capital and vertical integration. FERC's new interconnection framework rewards facilities that generate their own power, disadvantaging mid-tier colocation operators and conversion projects — exemplified by the North Tonawanda moratorium blocking Digi Power X's cryptomine repurposing. Meanwhile Amazon's reported move to sell Trainium externally, if confirmed, would pit a hyperscaler's proprietary silicon directly against NVIDIA in the merchant market, a competitive threat that only a company already monetising its own chip programme at AWS scale could credibly mount.

The infrastructure consolidation thesis extends into the semiconductor supply chain. Intel's 18A progress at VLSI matters precisely because a credible second leading-edge foundry would reduce the AI industry's single-point dependency on TSMC — but the gap between technical validation and production readiness remains real. Advanced packaging is emerging as the next chokepoint, with Amkor's Intel win underscoring that chiplet integration and high-bandwidth memory stacking now concentrate manufacturing risk in a similarly small number of OSAT providers. Slovenia's FRIDA launch and the broader European sovereign compute wave represent a deliberate counterweight: distributed national investments designed to operate outside this consolidation dynamic entirely.

The Trump administration's use of export control law to restrict Anthropic's international model distribution is the week's most consequential regulatory event — not primarily because of its immediate commercial impact on Anthropic, but because of the legal precedent it asserts. Treating AI model access as equivalent to a controlled physical export good gives the executive branch a lever over the commercial distribution strategy of every US-jurisdiction frontier lab. The incoherence of the posture — restricting distribution while simultaneously clearing the company of the threat designation that would justify restriction — suggests the action is as much about establishing the legal instrument as it is about any specific security concern. Constitutional lawyers are already contesting the theory.

The Anthropic situation sits alongside a broader pattern: FERC reshaping data centre economics through grid access rules, China weaponising indium controls explicitly against the AI hardware buildout, and Reliance's state-adjacent AI rollout to 500 million users reshaping the emerging market competitive map. In each case, state actors are using regulatory, infrastructure, or resource leverage to shape AI's commercial trajectory. For frontier AI companies, this week confirms that operating environment risk — executive intervention, export restriction, grid access conditions — belongs in capital allocation models alongside the more familiar competitive and technical risks.

UnitedHealth's $3 billion commitment is the most visible marker of an enterprise AI inflection that is visible across sectors this week. The healthcare giant's AI agents are not being tested — they are calling doctors' offices, parsing millions of patient calls, and summarising medical charts at operational scale. The dual framing of cost reduction and reputational rehabilitation in the wake of intense public scrutiny is analytically important: it shows that enterprise AI deployment is now embedded in corporate strategy responses to political and social pressure, not just efficiency programmes. For AI vendors selling agentic workflow automation and voice AI, UnitedHealth is simultaneously a reference customer and a template for how regulated industries justify transformation-scale investment.

Jane Street's scaling to 3,500 staff — with 500 more planned — illustrates the complementary dynamic in financial services: the largest AI monetisation is occurring inside closed organisations that treat AI as proprietary operational advantage rather than product. This segment of enterprise adoption is systematically undercounted in market analyses focused on model builders and cloud providers. The implication for AI vendors is that the most durable demand signal is not from companies evaluating AI but from companies that have already concluded they cannot compete without it — and are now hiring and deploying at infrastructure scale.