AI Autonomy Surges as Governance Gaps Widen and Tech War Deepens

Three simultaneous governance failures crystallised this week across distinct domains. The CDT audit documents federal agencies deploying AI in consequential decisions — immigration adjudication, law enforcement support, benefits processing — with no binding enforcement mechanism to compel the safeguards the OMB guidance nominally required. The Pentagon's $54 billion autonomous warfare request proceeds without updating its 2020 AI ethics principles or Directive 3000.09, despite operational context that bears no resemblance to the doctrine in force. And the Trump administration's retreat from its own Anthropic legal dispute confirms that the 'national champion' logic is now overriding enforcement calculus for frontier AI companies with perceived strategic value — effectively creating a two-tier system where scale and government proximity determine accountability exposure.

The enforcement vacuum is being partially filled from unexpected directions. Florida's criminal probe of OpenAI marks state attorneys general as the most aggressive active accountability actors in the US system — a pattern that mirrors how state-level enforcement drove tobacco and opioid accountability when federal agencies were passive. Meanwhile, the Metropolitan Police's Palantir discussions test whether UK data protection and procurement law are adequate governance instruments for high-risk AI adoption absent specific AI legislation. The common thread across all these cases is that the gap between stated policy and enforceable accountability is widening faster than any single jurisdiction is moving to close it.

The hardware-centric US export control thesis absorbed its most damaging week of evidence. DeepSeek's V4, backed by an explicit Huawei chip commitment, demonstrates that training efficiency innovation has reduced the compute requirements for frontier-class models enough to operate within a constrained hardware environment. Simultaneously, Chinese models now account for four of the top ten by token consumption on OpenRouter — a software-layer expansion that existing controls cannot touch. The White House memo framing distillation as strategic theft signals awareness of this gap, but enforcement faces technical obstacles that make it closer to a narrative framework than an actionable control mechanism. On the legislative side, the House Foreign Affairs Committee's 20-measure package targeting semiconductor manufacturing equipment access is still at committee stage, giving Chinese firms further runway to accelerate domestic substitution.

The Anthropic Mythos consortium model points toward the logical endpoint of this dynamic: frontier AI access stratified along geopolitical alliance lines, with a controlled US-aligned tier and a freely available Chinese-aligned open-source ecosystem. Nations in the Global South face the same structural choice they faced with Huawei 5G infrastructure — lower-cost Chinese capability or higher-cost Western alternatives — and the Chinese open-source strategy is deliberately engineered to be the rational economic choice for price-sensitive markets. Talent movement is the under-addressed transfer vector: Tencent's Hy3 model being led by a former OpenAI researcher illustrates that tacit knowledge — architectural intuitions, training methodology, research culture — flows through researcher mobility in ways that technology control lists cannot capture.

The vertical integration signal has moved from hyperscalers to a broader class of AI-intensive organisations. SpaceX's reported plan to manufacture GPUs in-house, Tesla's commitment to Intel's not-yet-production 14A node for custom AI silicon, and SoftBank converting Osaka factory space to produce batteries for its own data centres collectively represent a structural conclusion: merchant silicon and energy markets are too constrained, too price-volatile, and too capacity-constrained to support planning horizons beyond 12-18 months. Intel's 20% share price surge on data centre CPU demand and Applied Digital's $7.5 billion hyperscaler lease confirm that AI infrastructure spend is genuinely broad-based — but the supply chain stress flagged by multiple vendors in the same week, with shortages cascading from GPUs into power management controllers and memory, confirms that financial commitment cannot simply override physical constraints.

The TSMC dimension reinforces this picture from the supply side. Taiwan's regulatory move lifting single-stock concentration limits for domestic funds — directly strengthening TSMC's capital position ahead of its most capital-intensive multi-geography expansion — is an underappreciated form of industrial policy that parallels what the US government is attempting through CHIPS Act incentives. Advanced packaging capacity, particularly CoWoS and HBM stacking concentrated at TSMC, is emerging as an independent chokepoint as chiplet architectures become the dominant AI accelerator design. The implication for enterprises is that the addressable market for third-party infrastructure services may be smaller than projected as hyperscalers and near-hyperscalers bring critical layers in-house, while proprietary energy and cooling solutions acquire unexpected strategic value.