Export Controls, Capital Lock-Ins, and the Fracturing of AI Governance

This week exposed a striking asymmetry in the direction of AI governance travel. Australia moved from aspiration to enacted regulation by attaching binding constraints to government use of AI in administrative decisions — a concrete legislative step that distinguishes it from most G20 peers. The EU's AI Act implementation is advancing through formal consultation on high-risk classification, even if civil society bodies like CDT Europe are identifying ambiguities that could allow systemic misclassification. Meanwhile, both English-speaking AI powers are moving in the opposite direction. The US lost the key architect of its voluntary frontier model review process with Chris Fall's abrupt departure, leaving federal AI governance structurally dependent on non-binding pledges — instruments that create political cover without enforcement. In the UK, reported plans to abolish DSIT, home of the AI Safety Institute, would dismantle the country's primary technical AI safety capacity and its main interlocutor with the EU on regulatory alignment.

The industry response is calibrated to this governance vacuum. The coordinated letter from over 20 companies defending open-weight AI models — signed by NVIDIA, Microsoft, Meta, Palantir, and a16z, but notably not OpenAI or Anthropic — is a pre-emptive lobbying effort timed precisely to a Congressional markup where no binding federal framework yet exists. The coalition's composition reveals the fault line: infrastructure and compute players who benefit from model proliferation aligned against proprietary model developers who are commercially threatened by it. If the US fails to establish binding federal standards, it cedes the field to a patchwork of accelerating sub-national legislation that creates compliance complexity without coherent safety outcomes.

Three developments in a single week reveal that the AI compute market is being restructured through capital commitments rather than product competition. NVIDIA's expanded SK Group accord locks in preferential HBM supply at a scale that structurally disadvantages AMD and Chinese GPU makers dependent on the same memory. Anthropic's reported approach to SK Hynix for chip-making materials signals that frontier labs are following hyperscalers into vertical silicon integration, compressing the timeline for a bifurcated chip market of general-purpose GPUs and workload-specific inference ASICs. And SK Hynix itself is emerging as the strategic kingmaker — simultaneously supplying NVIDIA, taking equity in Etched's transformer ASIC startup, and potentially supplying Anthropic's custom programme — converting memory supply leverage into equity upside across every plausible inference architecture outcome.

Bond markets are beginning to price the risk that this capital intensity cannot be sustained. Moody's explicit warning that AI capex threatens credit quality at the world's most cash-generative corporations, combined with widening spreads ahead of hyperscaler earnings, introduces a demand-side constraint that equity market optimism has obscured. TSMC's confirmed price hike for advanced packaging propagates through every AI hardware vendor's cost structure simultaneously, and power availability — not chip supply — is emerging as the binding 2027-2028 constraint in US markets, insulated from no federal pledge mechanism. Infrastructure planners face a rack-level planning environment that will look materially different within 24-36 months: a mix of training GPUs, inference ASICs, and potentially lab-custom silicon rather than a homogeneous GPU estate.

This week's evidence presents a picture that cannot be resolved into a single US policy posture. China's RedNote achieved a perfect IMO score with a lightweight model, demonstrating frontier mathematical reasoning capability extracted from compute-efficient architectures — directly challenging the assumption that chip export controls will bottleneck Chinese AI development on the timelines that matter to current policy. Simultaneously, the joint UK-US government assessment of Kimi K3 found it significantly below US frontier models in offensive cyber capability — the dual-use domain most central to export control justification. These two findings are not contradictory; they reflect a domain-specific capability picture that the existing unified restriction regime cannot cleanly address. Allies who see the cyber assessment as evidence that controls are over-scoped will be harder to hold in the coalition.

Beneath the capability debate, China's AI sector is completing a structural financial decoupling from the West that renders economic coercion tools progressively less effective. State capital is the common investor across DeepSeek, Zhipu, Unitree, and CXMT; MetaX has confidentially filed a Hong Kong IPO to fund domestic GPU development; and Western VC has retreated under investment screening regimes that are accelerating the very state-directed consolidation Washington sought to forestall. Meanwhile, China is converting its WAIC platform into Global South AI diplomacy, offering infrastructure access and governance inclusion to markets that will account for the majority of new AI users over the next decade. Hong Kong's concrete experience of US compliance frameworks bifurcating AI tool access previews what swing-state economies across Africa, Southeast Asia, and Latin America will face as decoupling deepens.