AI's Fault Lines Deepen: Sovereignty, Spending, and Silicon Scarcity

This week saw the most explicit mutual assertion of AI sovereignty yet. China's NDRC blocked a completed cross-border acquisition on the grounds that AI capabilities developed by Chinese nationals are strategic national property — regardless of where the entity is incorporated. Simultaneously, DeepSeek's V4 models running natively on Huawei Ascend chips at prices 97% below OpenAI's demonstrate that US export controls have incentivised, rather than prevented, the construction of a fully autonomous Chinese AI stack. The two developments are mirror images: Washington restricts hardware outflows; Beijing restricts talent and capability outflows. The net result is accelerating bifurcation.

The United Front analysis underscores that hardware-focused export controls may be optimising against the wrong attack surface, while algorithmic know-how and research methodology transfer through human networks at high bandwidth beneath enforcement thresholds. Meanwhile, Chinese physical AI systems — robotaxis priced at $33,700 all-in — are being deployed commercially in third-country markets, creating infrastructure dependencies before Western policy frameworks have developed responses. The developer pricing war, with DeepSeek undercutting US API prices by 97%, is now extending this sovereign competition into the global developer ecosystem, where lock-in is the durable source of platform power.

Three concurrent stress signals at OpenAI — missed user and revenue targets, board-level concern about data centre commitments, and a Microsoft restructuring that trades exclusivity for cash flow certainty — suggest the frontier AI funding model is approaching a reckoning. The Microsoft deal's removal of the AGI clause is simultaneously pre-IPO hygiene and a signal that the philosophical scaffolding of the lab era is being replaced by conventional commercial logic. The Musk trial, running in parallel, creates an additional legal overhang on the for-profit conversion. If OpenAI, the market's bellwether, is visibly straining against growth assumptions, the pressure will propagate to LP scrutiny of AI infrastructure funds and hyperscaler capex timelines.

Against this, Accenture's confirmed deployment of Microsoft Copilot to all 743,000 employees and OpenAI's FedRAMP Moderate authorization represent the enterprise demand signals that the funding model depends on. The question is timing: infrastructure commitments are sized for a market that is forming more slowly than modelled at the application layer. David Silver's $1.1 billion Ineffable Intelligence raise — at a $5.1 billion seed valuation for a months-old lab — shows capital is still flowing at extreme multiples to elite research talent, compounding the talent extraction pressure on incumbents already missing their targets. The simultaneous Advantest capacity warning and Utah's 9GW power approval gap illustrate that the infrastructure constraints are real upstream and downstream of the capital cycle.

Google's decision to split its TPU v8 into separate training and inference variants is the clearest architectural signal yet that AI hardware demand is not homogeneous, and that hyperscaler custom silicon is maturing fast enough to compete at frontier scale. A million-TPU training cluster is the most credible proof point that alternatives to NVIDIA's monolithic GPU approach are viable — not in cost-optimised inference tiers, but at the highest-intensity training workloads. TSMC's CoWoS roadmap projecting 48x compute gains by 2029 through packaging rather than node shrinks confirms that advanced packaging is now as strategically significant a chokepoint as wafer fabrication, with TSMC holding dominant control of leading-edge capacity through the decade.

The supply chain stress is not only at the packaging layer. Advantest's near-7% share drop after flagging tight test equipment capacity is a hard upstream constraint: every chip must pass through test infrastructure before shipping, and this bottleneck cannot be resolved by packaging or fabrication investment alone. Core Scientific's conversion of a 300MW bitcoin mining site into a 1.5GW AI campus and Utah's approved 9GW campus — against a state grid that produces less than half that — illustrate the same structural gap at the power layer. Stranded industrial assets are being repriced around AI demand faster than grid infrastructure can support them, creating a pipeline of approved capacity that may not be deliverable on the timescales implied.