Infrastructure Under Fire, Safety Under Siege, Supply Chains Locked In

Iran's Islamic Revolutionary Guard Corps published detailed threats against OpenAI's $30 billion Abu Dhabi facility, including satellite imagery of the 1GW Stargate data center, marking a watershed in how adversaries view AI infrastructure. This represents the first publicly documented military threat against a named AI facility and demonstrates systematic intelligence collection on compute asset locations. The threat comes as multiple frontier labs concentrate capacity in the Middle East to access abundant energy, creating geographic vulnerabilities that insurance markets and national security planners are only beginning to model.

The implications extend beyond this single facility. Adversaries have clearly mapped the concentration of global AI capability in specific physical locations, many in politically unstable regions chosen for energy access rather than security. Companies that pursued Middle Eastern partnerships for power availability now face questions about whether compute resilience requires geographic distribution similar to financial system infrastructure, even at the cost of higher energy prices and longer buildout timelines.

The General Services Administration's draft AI procurement terms would allow agencies to compel contractors to disable trust and safety mechanisms, whilst California's executive order mandates safety standards for state purchasing within four months. Civil society groups warn the federal approach weaponises procurement to eliminate safeguards, whilst California uses the same lever to enforce them. This procurement-focused strategy reflects a pragmatic shift: governments can implement purchasing requirements through executive action without legislation, whilst companies must comply to access lucrative contracts.

The Anthropic-DoD dispute over dual-use restrictions is spawning systematic attempts to reshape procurement rules government-wide, moving safety commitment conflicts from isolated incidents to structural policy battles. Companies that built trust and safety infrastructure now face contractual prohibitions on using it in government work, creating perverse incentives that penalise prior safety investment. This dynamic is establishing a bifurcated market where some firms serve only government whilst others exit that sector entirely, with profound implications for which capabilities flow into public sector applications.

Anthropic's multi-hundred-billion-dollar agreement with Google and Broadcom represents a fundamental shift in how frontier labs secure compute: they are no longer simply buying GPU time but co-designing custom silicon and guaranteeing years of capacity before chips even exist. The structure gives Anthropic assured infrastructure access whilst locking competitors out of the same supply, and it positions Google as both infrastructure provider and indirect investor. Separately, Nvidia's SchedMD acquisition extends vertical control beyond hardware into orchestration software, potentially making it harder for competing chips to function in multi-vendor clusters.

This vertical integration pattern raises capital requirements for competition and tilts the market toward entities that can secure long-term chip partnerships. Samsung's eight-fold profit increase on AI memory sales and Firmus Technologies' $505 million raise for data center buildout demonstrate that infrastructure capital flows even as broader venture markets contract. The divergence suggests picks-and-shovels infrastructure attracts massive investment whilst application-layer startups face higher bars to prove unit economics, potentially creating a shakeout year for AI applications that have not demonstrated clear monetisation paths.