Beijing's AI Build-Out Still Runs on American Silicon
Despite export controls and a push for self-sufficiency, US-designed chips are set to power nearly two-thirds of China's expanding AI compute capacity in 2027.
The Scale of Dependence
China's artificial intelligence infrastructure is expanding at breakneck pace, with total available compute projected to leap from roughly 4.6 million to 8.8 million "H100-equivalents" in 2027. The metric, benchmarked against Nvidia's data-centre GPUs, has become the industry standard for comparing AI training and inference capacity across heterogeneous hardware. Yet beneath that headline growth lies a structural vulnerability: approximately 65 per cent of that capacity will still rely on chips designed by US firms, according to new research from a Washington-based think tank.
The figure is striking not because it reveals Chinese reliance on foreign technology - that much has been documented - but because it quantifies the gap between policy ambition and technical reality. Since 2022, Beijing has poured subsidies into domestic semiconductor champions and Washington has tightened export rules on advanced logic chips. The persistence of US-designed silicon at the heart of China's AI stack suggests that neither policy lever has yet achieved its intended effect.
How the Compute Flows In
The 65 per cent share comprises two channels. The first is direct: chips that clear current export-control thresholds - whether older-generation Nvidia parts, downgraded SKUs designed specifically for the Chinese market, or processors from AMD and Intel that fall below performance cut-offs. These devices land in data centres operated by Chinese cloud providers, research labs, and AI start-ups, forming the substrate for training runs and inference workloads.
The second channel is indirect and harder to police. Chinese entities lease compute from offshore cloud providers - primarily hyperscalers with data centres in Singapore, Tokyo, and jurisdictions beyond the reach of US export rules. The workloads are Chinese; the hardware is American; the jurisdiction is ambiguous. This arrangement has grown in sophistication over the past eighteen months, according to supply-chain analysts we have spoken to across the region. Contracts are structured to avoid explicit breach of end-user restrictions, yet the practical effect is to extend access to frontier-class chips that cannot be shipped to the mainland.
At Opentechwire, we have tracked the financing and buildout of several such offshore GPU clusters since mid-2025. The capital often originates with Chinese venture arms or state-backed funds; the servers sit in leased racks in Singapore or Kuala Lumpur; the inference API endpoints are consumed by applications serving users in Shenzhen and Hangzhou. The regulatory surface area is vast, and enforcement remains inconsistent.
Domestic Alternatives Lag in Performance and Scale
China's semiconductor industry has made measurable progress. Huawei's Ascend 910B, for instance, is deployed in several large-scale training clusters, and start-ups such as Biren and Moore Threads have shipped products to select customers. Yet the performance gap remains wide. The 910B delivers competitive throughput on certain workloads, but its software ecosystem - compilers, libraries, framework integration - still trails Nvidia's CUDA by years. Model developers report longer debugging cycles and higher engineering overhead when porting code to domestic accelerators.
Scale is the other constraint. Yield rates at Chinese fabs producing advanced logic remain below those at TSMC, and access to extreme ultraviolet lithography tools is effectively cut off. As a result, domestic production cannot yet meet the volume demanded by the country's AI sector, even if performance parity were achieved tomorrow. The doubling of compute capacity in 2027 will be absorbed by foundation-model pre-training, fine-tuning for vertical applications, and the inference load from hundreds of millions of consumer-facing chatbots and agents. Domestic chips will supply a portion of that capacity; American-designed silicon will supply the majority.
Policy Implications and the Export-Control Dilemma
For Washington, the data presents a dilemma. Export controls have slowed China's access to the most advanced chips - Nvidia's H100 and successor architectures are largely absent from the mainland - but they have not severed the link. The offshore-cloud loophole and the continued export of lower-tier devices mean that Chinese AI labs can still train models at scale, albeit with longer timelines and higher costs.
Tightening the rules further carries risk. Expanding the performance threshold to capture more chip SKUs would affect sales to allied markets and complicate supply chains for multinational cloud providers. Restricting access to offshore compute would require extraterritorial enforcement mechanisms that few governments are prepared to implement. The result is a control regime that imposes friction without achieving containment.
For Beijing, the reliance on US-designed chips is both a technical and a strategic problem. Technical, because it constrains the pace at which Chinese labs can iterate on large models and compete with OpenAI, Google, and Anthropic. Strategic, because it leaves critical infrastructure vulnerable to policy shifts in Washington. The 65 per cent figure is a measure of that vulnerability. It also explains the scale of state support flowing into the domestic chip sector, even as returns on that investment remain uncertain.
What the Trajectory Suggests
The trajectory over the next twenty-four months will depend on three variables: the pace of domestic chip improvement, the durability of offshore-cloud workarounds, and the willingness of US policymakers to accept leakage in exchange for preserving market access elsewhere. None of these variables is moving in a single direction.
Chinese labs are training competitive models today, and they are doing so on a hybrid stack that mixes Ascend, Nvidia, and offshore capacity. That pragmatism is likely to persist. The goal is not ideological purity; it is functional AI systems that can be deployed at scale. As long as American silicon remains the most efficient path to that goal, it will continue to flow into China's AI infrastructure, whether through front doors or side channels.
The 65 per cent share in 2027 is not a static endpoint. It is a snapshot of a system under stress - one in which policy, technology, and market incentives are pulling in different directions, and in which no single actor has the leverage to resolve the tension cleanly.
