Military AI Doctrine Hardens as China Chips Close the Gap

The Pentagon's 'any lawful use' contracting terms with seven AI firms represent an enacted governance standard, not a temporary procurement decision. The operative constraint on military AI use is now existing law rather than any bespoke ethical framework, and the exclusion of Anthropic — the one lab that objected to broad terms — creates a structural incentive for competitors to accept permissive conditions to preserve government contract access. The concurrent appointment of the former Pentagon think-tank head to Anthropic's leadership, and the deeper revolving door between national security establishments and frontier labs, is hardening the alignment between commercial AI development priorities and US defence doctrine in ways that will shape capability trajectories for years.

For allied governments, the 'AI-first' doctrine creates an interoperability imperative and an architectural dependency risk simultaneously. For strategists tracking escalation dynamics, AI-accelerated military decision cycles compress crisis windows in scenarios like Taiwan or the South China Sea. No current international arms control mechanism addresses AI-enabled military capabilities, and the DoD's own responsible AI principles remain non-binding guidance. The precedent being set through procurement is therefore the operative governance standard by default — and it is one that other advanced economies, particularly through NATO and Five Eyes coordination, may feel pressure to replicate.

The convergence of signals this week — Huawei's 60% revenue growth forecast, Cambricon's 160% quarterly surge, memory shortages projected through 2027, and Apple raising Mac Mini prices due to AI-driven processor scarcity — confirms that semiconductor supply chain stress has broadened from GPU accelerators into every layer of the stack. US export controls, designed to constrain Chinese AI capability, are producing a documented second-order effect: a protected domestic Chinese market channelling AI hardware demand toward Huawei Ascend, Cambricon, and MetaX, accelerating the indigenous capability timeline that controls were intended to slow. Chatham House, CFR, and War on the Rocks commentary is converging on the same assessment — hardware controls alone are insufficient, and the acceleration paradox is becoming conventional wisdom among strategists.

For the US-aligned infrastructure buildout, the binding constraint is not capital — $725 billion in announced hyperscaler capex confirms financial commitment — but physical supply. Samsung and SK Hynix's record-low fulfilment rates and multi-year forward booking by customers signals that money cannot simply manufacture HBM memory faster. Nvidia's parallel move from chip vendor toward AI infrastructure systems integrator, through its Invenergy-Emerald partnership on flexible AI factories, reflects a strategic response to this environment: controlling more of the stack reduces exposure to any single supply chokepoint. But it also concentrates architectural influence in ways that create new dependencies for hyperscalers and sovereign programmes alike.

Two distinct but structurally related accountability gaps surfaced this week. The CDT-MIT finding that fine-tuning foundation models produces unpredictable safety degradation — with no current compliance framework able to assign liability for the resulting harm — mirrors the IAPS finding that frontier models run internally for weeks before public release with no reporting obligations under the EU AI Act, US state bills, or the federal RAISE Act. Both gaps share the same architecture: governance instruments are calibrated to the foundation model as publicly released, and have no effective reach either upstream into internal deployment or downstream into the adaptation chain. As fine-tuning becomes the dominant mode of commercial AI deployment, this gap will generate harms that fall outside every existing compliance mandate. The EU AI Act's AI Omnibus process is the most active vehicle for closing these gaps, but the multi-actor accountability disputes playing out over deepfake liability — developer versus platform versus perpetrator — illustrate how technically and politically complex the design questions are.

The AI revenue accounting concern adds a financial governance dimension to the same pattern. If frontier AI companies are booking joint venture revenue in ways that do not reflect genuine third-party demand, the growth rates underpinning both private valuations and public market narratives are systematically overstated. The absence of audited public financials for most frontier labs makes this nearly impossible to assess from outside. Together, these governance gaps — in safety accountability, regulatory reach, and financial transparency — suggest that the architecture of AI oversight is lagging the actual risk landscape at multiple layers simultaneously.