Combat AI Proven, China Surges, and Capital Floods Every Layer

Three developments this cycle crystallise a deepening strategic irony: SMIC's record $3 billion quarter — funded by captive demand that US sanctions created — is generating the margin to reinvest in the very capability gap controls are meant to preserve. CXMT's public listing gives China's primary domestic DRAM producer capital markets access at precisely the moment HBM demand is accelerating. And China's supernode architecture — aggregating domestic chips to approximate banned Nvidia-class compute — has transitioned from workaround to mainstream deployment model, displayed openly at the World Artificial Intelligence Conference in Shanghai. Controls are fragmenting rather than halting Chinese AI development, producing a bifurcated compute landscape: supernodes for volume workloads, rationed Nvidia for precision-demanding tasks like coding where domestic chips still fall short.

The financial dimension reinforces the trajectory. Alibaba's willingness to absorb a 75% net income collapse to fund $10 billion in quarterly AI capex — mirroring the AWS and Azure buildout playbooks — signals that Chinese hyperscaler investment has entered a self-reinforcing cycle. Combined with Alibaba's Qwen model matching OpenAI's cost-efficiency flagship on benchmarks at consumer-hardware scale, and Shanghai's new five-year digital economy plan adding municipal policy weight to corporate investment, China's AI infrastructure capacity is on a trajectory to expand significantly through 2028 regardless of export control headwinds. The aggregate picture is not one of a constrained competitor but of a competitor whose constraint environment is accelerating certain dimensions of its capability.

Capital is flowing into AI infrastructure with a simultaneity not seen since the cloud buildout of the 2010s. Broadcom's $60-billion-plus debt raise for custom silicon, Micron's $10 billion AI memory research lab, Alibaba's near-$10 billion quarterly capex, Danfoss forecasting its cooling segment to at least double its group revenue share, and Zayo locking in 15,000 route miles of fiber supply through 2030 are not isolated bets — they are a coordinated multi-layer buildout in which every supplier from wafer to cooling fluid is making long-duration commitments. The fixed-income market's appetite for AI infrastructure paper, most visibly in Broadcom's financing, is functioning as a real-time confidence gauge on institutional belief in the duration and scale of the buildout. California's record $366 billion in venture capital — more than all other states combined — with AI as the dominant driver, confirms that equity markets are equally committed.

The application layer presents a more complex picture. Anthropic's expectation of an IPO matching SpaceX's record listing, and the broader frontier lab valuation landscape, rests implicitly on enterprise AI spend becoming sticky at scale. But new data showing businesses switching between OpenAI and Anthropic as model rankings shift — and the Open Machine CEO's framing of two parallel enterprise tracks, neither resembling committed SaaS-era spend — exposes a durability gap between infrastructure confidence and application monetisation. Stripe's acquisition of OpenRouter and Ramp's launch of its own model router in the same week signal that incumbents see control of the model-routing layer as the durable position: whoever owns routing owns the data on which models enterprises prefer, at what price, and for which tasks — an information advantage that persists regardless of which frontier lab leads at any given moment.

Operation Epic Fury has moved the military AI debate from theoretical to empirical: 13,000 strikes over 38 days, with the opening 24 hours generating roughly 1,000 targets via Claude through Palantir's Maven Smart System, represents enacted capability with documented battlefield outcomes. Adversaries — principally China and Russia — now have a demonstrated US operational standard to calibrate against, not a projected one. The strategic response is already visible in China's accelerating PLA AI integration, where the gap between official doctrine — human commanders retain final authority — and operational ambition is widening under Xi Jinping's direct pressure. War on the Rocks analysis surfaces the structural instability: PLA official media cannot reconcile its human-control reassurances with the tempo demands that AI-enabled operations actually impose.

The governance architecture on both sides is lagging operational deployment by years. A Council on Foreign Relations proposal for a FINRA-style domestic AI regulator is reportedly under White House review — while the US is already deploying combat AI at scale. Allies seeking to anchor interoperability standards to US frameworks cannot do so when the framework does not exist. The arms control literature from nuclear history offers three warnings that apply directly: verification of AI military systems is technically intractable, definitional disputes precede any binding limits, and the window for meaningful controls narrows as deployment accelerates. That window is demonstrably narrowing. The absence of a multilateral framework to govern the interaction between US and Chinese AI-military systems is not a future risk — it is a present asymmetry both powers are exploiting while nominally signalling openness to dialogue.