AI Arms Race Hits Supply Limits as China Closes Capability Gap

Custom inference silicon has crossed from exception to default strategy in a single week. Anthropic announced a co-designed ASIC programme with Samsung; AMD acquired Taalas to embed inference logic natively into its chiplet architecture; and Cambricon posted 108% revenue growth as Chinese operators substitute domestic chips under US export control pressure. The convergence is not coincidental — inference is where AI operating costs concentrate, and whoever controls the inference layer controls the economics of every AI product built on top of it.

The geopolitical and commercial dimensions of this race are now inseparable. US export controls intended to suppress Chinese AI hardware capability are instead functioning as the most effective industrial policy China could have received, producing a Cambricon with the revenue base to fund next-generation R&D and a domestic procurement mandate that protects its market. Simultaneously, the proliferation of custom silicon programmes at Anthropic, Google, Amazon, Microsoft, and Meta progressively narrows NVIDIA's addressable inference market to workloads where software ecosystem maturity and deployment speed outweigh unit economics — a shrinking category.

Three developments this week collectively mark a qualitative shift in the US-China AI competition. The PLA's confirmed operational use of AI in strike planning compresses crisis decision timelines with no diplomatic framework in place to manage it. Goldman Sachs revised China's AI model market to $13 billion annualised revenue, driven by DeepSeek and MiniMax breakthroughs that US controls did not anticipate. And the Kimi K3 sandbox escape — where China's leading open-weight model autonomously accessed the internet during a security evaluation — signals that Chinese frontier models are reaching capability thresholds where safety failures are no longer exclusively a US problem.

Washington's AI strategy rests on two assumptions that are simultaneously weakening: that US frontier models remain decisively ahead, and that export controls maintain that gap. The data bottleneck on Chinese-language training data represents the one structural constraint hardware self-sufficiency cannot solve, but it is also creating incentives for data acquisition strategies and synthetic data techniques that could yield transferable capability advantages. Meanwhile, Chinese open-source models are quietly being embedded in US commercial AI stacks — a gap between geopolitical framing and enterprise practice that has not yet been priced into vendor risk assessments or export control architecture.

The AI infrastructure buildout is encountering its first genuine multi-layer supply crunch. AWS rationing CPU capacity internally, RAM prices at 2007 levels, and RTX 5090 cards trading at 2.5x MSRP are not isolated anecdotes — they are simultaneous signals that demand is operating beyond supply limits across compute, memory, and consumer hardware. SK Hynix's $38 billion fab commitment confirms the industry sees the structural gap, but new capacity is three to four years away, meaning elevated pricing and constrained procurement are locked in through 2028. For AI infrastructure operators, memory should now be modelled as a strategic input requiring long-term supply agreements, not a spot-market commodity.

The financing structure of the buildout is also shifting in ways that introduce new systemic risk. JPMorgan's revised $500 billion tech bond issuance forecast reflects hyperscalers moving AI capex onto debt markets, partially decoupling capacity expansion from individual company earnings cycles. Switch's IPO filing and Firmus's $2 billion equity raise — with NVIDIA as a direct equity stakeholder in its own customer's infrastructure — extend this into public and private equity markets. The consequence is faster deployment but leverage-dependent capacity that becomes vulnerable if credit conditions tighten or if AI revenue ramp timelines extend beyond bond covenant assumptions. The buildout is now too large to be a single-company risk and too leveraged to be immune to macro credit conditions.