The Gist: Executive Overview

The week crystallised a fundamental truth: voluntary AI safety commitments and responsible use policies have no enforcement mechanism once models reach customers, especially government ones. OpenAI's admission it cannot control Pentagon usage, Anthropic's forced reversal after attempting to maintain distance from military applications, and the absence of any regulatory framework governing AI in combat operations reveal that the entire edifice of corporate AI ethics dissolves under strategic pressure. Companies can decide whether to sell to defence departments, but once they do, no oversight exists. The White House data centre energy pledge follows the same pattern — symbolic commitments designed for optics, with no compliance mechanisms or penalties.

This pattern extends beyond military applications: Google's Gemini faces the first wrongful death lawsuit with unclear liability frameworks, Meta's privacy violations expose gaps between stated policies and actual data handling by subcontractors, and intellectual property protections prove meaningless when AI systems synthesise copyrighted work. The governance vacuum exists because existing regulatory authorities were designed for different technological paradigms, and new frameworks remain theoretical while deployment accelerates. The result is a widening gap between the pace of capability deployment and the speed of accountability structures.

The binding constraint on AI capability is shifting from algorithms to physical infrastructure — energy, semiconductors, and data centre locations. Seven tech giants signed a political pledge on electricity because bipartisan voter backlash over rate increases threatens permitting for new facilities. Nvidia is reallocating TSMC production capacity in response to export controls, revealing that chip manufacturing allocation matters more than model weights for determining who can deploy at scale. Physical data centres have become military targets in the Iran conflict, forcing reconsideration of geographic concentration strategies. Memory prices are surging as data centre buildout competes with consumer electronics for supply.

This infrastructure-centric competition is reshaping the semiconductor industry: Broadcom projects $100 billion in AI chip sales by 2027 as hyperscalers pursue custom silicon to reduce Nvidia dependence, while Meta announces plans for proprietary training chips despite recent supplier deals. The competitive moat is no longer who has the best model, but who controls access to fabrication capacity, electricity supply, and geographically secure facility locations. Europe's push for technological sovereignty through deep tech investment reflects understanding that dependency on US infrastructure creates strategic vulnerability.

The Anthropic case demonstrates that AI companies in the US cannot sustainably maintain distance from defence applications without facing regulatory designation as security risks and commercial penalties through client flight. The Trump administration's willingness to label Anthropic a supply chain threat for briefly resisting military integration establishes a precedent: strategic alignment with government priorities overrides corporate preferences when officials view technology as militarily relevant. This dynamic is not limited to the US — European governments are directing capital toward indigenous defence and AI capacity, India is tasking startups with satellite defence systems, and middle powers like Canada are articulating hedging strategies against great power dominance.

The operational military deployment of AI in Iran operations, with US Central Command openly confirming AI tools are managing targeting data at scale, marks a watershed in capability diffusion. This is distinct from experimental systems or intelligence analysis — it represents AI as a core dimension of operational tempo in active combat. Countries without comparable access face structural disadvantages in information processing speed during high-intensity conflict. Export controls are proving effective not just at limiting sales but at forcing production allocation shifts at TSMC, creating a more durable form of technological leverage than licensing regimes alone.