AI Capex Cycle Hardens as Geopolitics and Infrastructure Friction Mount

Three deals this week collectively redefine how AI compute is acquired and financed. Jane Street's $6 billion CoreWeave offtake plus $1 billion equity stake mirrors energy infrastructure financing models, treating GPU capacity as a scarce long-duration asset. OpenAI's reported $20 billion Cerebras commitment with an equity component would make a frontier lab a strategic co-investor in an alternative chip architecture — a direct hedge against Nvidia supply queue exposure. TSMC's 58% profit growth, raised guidance, and Taiwan's resulting overtaking of the UK in stock market capitalisation confirm that these commitments reflect real committed demand, not announcement-phase intent.

The downstream consequences are significant. Long-duration equity-linked procurement concentrates GPU cloud market share among the well-capitalised early movers, squeezing spot availability for smaller operators. It also validates undercapitalised entrants — shoe company Allbirds and Bitcoin miner Cango both announced GPU-as-a-Service pivots this week — whose business plans assume margins and hardware access that committed offtake arrangements are now foreclosing. Goldman Sachs projecting $1 trillion in AI investment over three to four years sets the macro frame, but the micro-level reality is that the most strategic compute positions are being locked up now, before that capital fully deploys.

The UK's week encapsulates the tension at the heart of sovereign AI strategy. OpenAI paused its Stargate project citing energy costs and regulatory friction, drawing a public ministerial rebuke that revealed how little enforcement leverage governments hold once announcement-phase commitments are made without binding penalty structures. Simultaneously, Anthropic — cooling on Washington over Pentagon tensions and federal lawsuits around its Mythos model — announced a fourfold London headcount expansion. The UK government matched this with a £500 million sovereign AI fund, creating a self-reinforcing loop where public capital attracts private lab investment which justifies further public commitment. The fund is modest relative to compute and talent costs, but the directional signal is clear.

The infrastructure siting crisis adds a second dimension. Maine's legislative ban on new data centre construction — the first at state level — creates durable legal barriers that community relations programmes cannot resolve. Combined with grid interconnection queues, water rights constraints, and the engineering reality that next-generation AI racks require 120kW or more of liquid-cooled power density, viable large-scale AI training site geography is compressing. The winners are jurisdictions offering cheap grid power, streamlined permitting, and — increasingly — political alignment with the labs' regulatory preferences. That dynamic is precisely what the UK is attempting to monetise, and what the Middle East and Sun Belt US are already exploiting.

ASML's declining China revenues confirm that the US-Netherlands-Japan export control coalition is materially compressing China's access to advanced lithography — the controls are working at their intended target. But the strategic frame emerging from multiple signals this week is that China has accepted this constraint and pivoted to the layer where controls are weakest: inference deployment at scale. Open-source model releases from Alibaba and others seed global developer ecosystems at near-zero marginal cost, building infrastructure dependency across Global South markets where US export controls create a vacuum. Lightelligence's Hong Kong IPO clearance and the broader photonics computing push represent the longer-term hedge — alternative silicon architectures designed to reduce dependence on EUV-produced conventional chips entirely.

Congressional testimony this week confirmed China operates a dual-track acquisition strategy — purchasing through licit channels and extracting through theft — while simultaneously flagging that US immigration and research funding restrictions are eroding the domestic innovation base those controls are designed to protect. Anthropic's experience is illustrative: black-market Claude workarounds emerged within days of tightened access controls, demonstrating that Chinese demand for frontier US models is significant enough to sustain circumvention infrastructure. The structural policy dilemma is acute: extending controls to inference hardware would be economically disruptive to allied relationships given its commercial ubiquity, but failing to do so risks ceding the deployment layer that generates the revenue and influence that training-tier controls were meant to prevent China from capturing.