China's AI Stack Closes In as Western Controls Fray

Three developments this week collectively constitute the most substantive empirical challenge yet to the US export control regime's strategic premise. The Taiwan-Japan transshipment arrests confirm that organised networks are reliably supplying restricted Nvidia chips through allied jurisdictions, exploiting enforcement gaps at the systems integration layer that hardware-focused controls were never designed to close. Huawei's Tau Scaling Law advances a parallel architectural trajectory — prioritising data throughput over transistor density — that is explicitly designed to render ASML lithography denial irrelevant by 2031. And China's AI startup funding tripled to $16.2 billion in Q1 2026 alone, concentrated in LLMs and embodied AI, demonstrating that capital formation is accelerating rather than being suppressed.

Taken together with DeepSeek's permanently discounted frontier model and nine domestically certified AI chips entering government procurement, the picture is of a bifurcated Chinese strategy: acquire restricted Western hardware through grey-market channels in the short term while building sovereign alternatives to eliminate long-term dependency. The export control architecture achieves delay, but the substitution pathways — architectural innovation, capital intensity, efficiency research, and open-source leverage — are multiplying faster than controls are being tightened. Policymakers need to begin planning for a scenario in which the constraint becomes non-binding before the end of the decade.

The binding constraints on AI infrastructure expansion are consistently proving to be non-financial. Electrician shortages mean that data centres are being built faster than they can be energised, with multiple hyperscale projects sitting complete but unconnected pending licensed electrical commissioning. SK Hynix's iHBM thermal architecture announcement — embedding cooling directly into the HBM interface layer to cut thermal resistance by 30% — signals that heat management inside densely stacked memory is a hard physical ceiling on inference throughput that external cooling cannot resolve. Meanwhile, Micron and Argonne research confirms that reasoning-model inference workloads are memory-bandwidth-bound rather than compute-bound, meaning current cluster designs optimised for training-era workloads will require retrofitting as chain-of-thought architectures become the inference norm.

The security layer compounds these physical constraints. The confirmed prompt injection attack via the jqwik open-source library demonstrates that agentic coding pipelines face an unmitigated attack vector at the dependency level — natural language instructions embedded in code context can direct AI agents to take destructive actions with no exploit required. Uber's public admission that it exhausted its annual AI budget in four months without demonstrable productivity return adds a measurement gap to the trust problem: enterprises are running agentic workflows they cannot audit, against dependencies they cannot fully trust, with outcomes they cannot reliably quantify. The infrastructure stack has physical, thermal, and security layers that financial capital alone cannot accelerate.

Anthropic's $965 billion valuation — the first time it has surpassed OpenAI — marks a new phase in frontier AI capital concentration where crossover and growth-stage funds are making multi-billion-dollar conviction bets rather than defensive positions. The round's timing is inseparable from the governance moment: Illinois has passed the most demanding AI safety legislation in the US, mandating third-party audits of major frontier developers, while the EU has already opened diplomatic talks with Washington over Anthropic's forthcoming Mythos model before its public release. Anthropic's Project Glasswing and Exploit Evals framework are not coincidental — they represent safety infrastructure being built to regulatory specification ahead of enforcement, converting compliance readiness into commercial differentiation in enterprise and government procurement.

The parallel IPO preparations of OpenAI and China's MiniMax create an unusual moment of simultaneous public price discovery for US and Chinese frontier AI, with Anthropic's private valuation now functioning as a direct comparable that will complicate OpenAI's pricing narrative. The SaaS rehabilitation — software stocks posting their best month since 2001 — and the concurrent VC rotation into hardware and inference infrastructure reflect a market distinguishing between application-layer software with embedded AI workflows and generic software facing commoditisation. The value capture question remains open: whether durable margin accrues at the frontier model layer, the infrastructure layer, or the enterprise integration layer will determine where the most defensible investment positions lie across the next capital cycle.