AI Cold War Hardens: IP Controls, Sovereign Stacks, and Capex Discipline

Three developments this week collectively confirm that the US-China AI competition has entered a qualitatively new phase. The White House distillation memo attempts to close the arbitrage that allowed Chinese labs to compress the capability gap without matching US compute investment. Simultaneously, Huawei's day-zero Ascend chip integration with DeepSeek V4 demonstrates that China is no longer playing catch-up — it is building a vertically integrated domestic stack explicitly designed to reduce foreign chokepoints. And China's State Grid Corporation deploying 6.8 billion yuan of AI robotics into sovereign power infrastructure converts that stack into durable physical capability that cannot be reached by software-layer controls.

The policy logic of distillation restrictions faces a fundamental enforcement problem: model distillation is technically difficult to attribute, Chinese labs can claim independent development, and the most capable Chinese open-weight models are themselves now a distillation source for third-country actors. If the US formalises these restrictions, Beijing's rational countermove is to accelerate open-weight releases, seeding a global ecosystem built on Chinese intellectual foundations rather than American ones. The Cohere-Aleph Alpha merger adds a third dimension: European sovereign AI is consolidating around a non-US alternative specifically to serve regulated markets and governments that cannot or will not use American hyperscaler infrastructure. The technology perimeter is fracturing in multiple directions simultaneously.

Meta and Microsoft's combined 23,000 redundancies, explicitly framed as AI capex funding, are the most visible expression of a dynamic visible across the sector: AI productivity gains in code generation — with Google at 75% AI-written code and Meta targeting similar levels — are reducing the marginal return on software engineering headcount and making labour a funding source for the next wave of AI investment. ServiceNow's 16% single-day collapse, despite beating earnings, reflects markets pricing this logic into enterprise software valuations: if AI reduces seat counts and per-user licence revenue, the SaaS growth model that justified a decade of premium multiples is structurally impaired. The EQT warning about stalled PE exits from software portfolios confirms the repricing is not confined to public markets.

The Gartner IT spending forecast revision — up nearly three percentage points despite the IEA declaring the worst energy crisis in history — reinforces that AI infrastructure has been re-categorised as non-discretionary capex by hyperscalers. Intel's exit from consumer GPUs to redirect resources toward data centre accelerators follows the same logic at the hardware layer. The risk embedded across all these decisions is a concentration of AI productivity gains at the top of the income and capability distribution: FT-Focaldata polling shows high earners adopting AI at dramatically faster rates than lower-income workers, meaning aggregate enterprise productivity uplifts may be materially lower than headline projections, even as the displacement of mid-skill labour accelerates.

The TSV packaging bottleneck analysis, Samsung's planned May strike action, and the surge in Chinese optical component investment are not independent stories — they collectively describe an AI hardware supply chain that is both deeply oligopolistic and actively bifurcating. TSV yield is a hidden leverage point: HBM supply can be constrained even when wafer starts are adequate, and the process maturity required is a genuine barrier to entry that reinforces the SK Hynix-Samsung-Micron oligopoly. Samsung's 18-day strike, arriving in the same quarter as peak Blackwell system deployment demand, creates a supply cliff in Q3 that current market pricing has not fully reflected. Intel's retreat from consumer GPUs narrows the GPU competitive landscape further, compounding Nvidia's pricing power.

The bifurcation dynamic operates at the component level below the compute chip layer. Chinese optical transceiver suppliers — Innolight, Eoptolink, HG Genuine — already hold significant share in 400G and 800G data centre transceivers globally, and investor flows are accelerating into the sector on AI demand projections. Unlike compute chips, optical interconnects are not subject to current export control frameworks, meaning Chinese suppliers' growing role in AI data centre infrastructure expands regardless of chip restriction regimes. The Blaize-NeoTensr edge AI partnership for Asia-Pacific similarly illustrates a parallel stack being built for markets that will not or cannot use the Western hyperscaler supply chain. The methodological point from NextPlatform reinforces the opacity: infrastructure buyers using GPU-hours as a primary cost benchmark are making procurement decisions that systematically underweight the architectural divergence now baked into the global AI hardware landscape.