Export Controls Unravel as AI Infrastructure Race Enters Physical Limits

Two developments this week, taken together, materially weaken the strategic premise of US chip export controls. C4ADS has documented that billions of dollars of restricted Nvidia accelerators are reaching China through structured transshipment networks — not opportunistic smuggling but a high-volume, organised supply operation. Simultaneously, Huawei has accelerated its Ascend 960DT launch to Q1 2027, driven not by engineering breakthroughs but by market urgency from Chinese hyperscalers who need a credible domestic alternative at scale. The controls are functioning as a price and friction mechanism, not a hard capability ceiling.

The downstream consequences are compounding. If restricted chips are available in China at scale and a domestic alternative is maturing faster than scheduled, the competitive moat assumed by US AI labs is materially narrower than policy frameworks suggest. For capital markets, this reprices Nvidia's China total addressable market and the valuation of US AI infrastructure plays predicated on Chinese compute constraint. For regulators, it raises the question of whether further control tightening has diminishing returns, or whether the policy instrument itself needs redesigning. The upcoming US-China summit — with AI explicitly on the agenda and AI executives at the table — is the nearest event that could binary-shift this calculus.

The formation of the AI Energy Management Alliance by Google, Nvidia, and Anthropic alongside utility RWE, Crusoe's $3.9 billion infrastructure round, and rising cloud prices from Nebius all point to the same structural reality: the AI buildout's binding constraint has decisively shifted from model capability to physical infrastructure. The 100 GW grid target being pursued by the major labs is not aspirational positioning — it is a prerequisite for the next generation of data centre campuses that regulators and grid operators must approve. Institutionalised demand response coordination between AI labs, chip makers, and utilities is now a gating factor on that approval process.

Advanced semiconductor packaging is emerging as a parallel chokepoint that receives a fraction of the policy attention directed at front-end fabs. Manufacturing gains in AI hardware are increasingly limited not by transistor density but by the integration of compute, memory, and storage into thermally manageable systems — and that packaging capacity is highly concentrated at TSMC's CoWoS lines. India's twin-track approach — $12 billion in semiconductor investment pledges and over $1.2 billion in state-backed data centre loans — represents the most sophisticated emerging-market attempt to insert a new node into this constrained supply chain, though pledge conversion to operational capacity remains the critical test.

Three developments this week illustrate different failure modes in the current AI governance architecture. The EU KIDS Act formal proposal adds a third vertical regulatory layer on top of the AI Act and DSA, raising immediate questions about enforcement fragmentation between national data protection authorities, Digital Services Coordinators, and AI Act market surveillance bodies. Meta's Oversight Board issued a binding deepfake removal order and found Meta's own safeguards 'inadequate' — a finding that simultaneously strengthens Ofcom's and the EU's enforcement basis while demonstrating that platform quasi-judicial bodies are making substantive governance decisions ahead of state regulators who are still building capacity. And confirmed research has shown that Google's SynthID watermarking scheme — deployed as an accountability and provenance tool — inadvertently degrades the safety alignment it was meant to complement, creating a new attack surface from within the governance toolkit itself.

The legislative environment in the US is being pulled in conflicting directions simultaneously. The EFF is pushing Congress toward specific, auditable cybersecurity mandates grounded in the OpenAI-Hugging Face breach. AI Now is routing governance concerns through antitrust frameworks that would engage existing FTC and DOJ powers. And the 'pacing the frontier' proposal backed by Altman and Amodei is drawing formal antitrust scrutiny for potentially functioning as a coordinated market-restriction mechanism. None of these vectors are aligned, and the drafting battles — not the passage votes — are now the decisive regulatory moment. The window for shaping scope is narrowing as political momentum builds faster than technical consensus.