Jensen Huang, Nvidia’s founder and chief executive, made a deliberately low-key entrance to China this month: a stroll through a Shanghai fruit market, a signed red envelope for a stallholder and a visit to Nvidia’s new Shanghai office in Zhangjiang. The small gestures were bookends to a more conventional schedule that included a company year‑end event and meetings at three of Nvidia’s main Chinese hubs — Shanghai, Beijing and Shenzhen — underscoring that China remains central to the company’s strategy even as geopolitics complicates access to cutting‑edge chips.
The commercial stakes are material. Nvidia reported roughly $17 billion of revenue from China in fiscal 2024, about 13% of group sales, and Huang has repeatedly warned Washington that excluding US firms from China’s AI market would be costly. His public comments have been matched by private attentiveness: the Shanghai visit came after a run of PR and partner events and follows meetings with municipal officials last year, signalling a concerted effort to reassure customers and suppliers on the ground.
Yet Huang’s market-side diplomacy conceals a tougher strategic truth: Chinese innovation and industrial mobilisation are reducing their reliance on imported accelerators. Despite US export controls on high-end Nvidia parts such as the H200 family, local researchers and developers have adapted by innovating across software, model design and system architecture. Huang himself has highlighted several Chinese open-source models — DeepSeek‑V3.2, Kimi K2 and Alibaba’s Qwen — as evidence that high‑performance AI can be achieved without unrestricted chip access.
That trend is visible in the hardware supply chain. IDC data show domestic compute accelerators’ share of China’s data‑centre market climbing from 14% in 2023 to 34.6% in 2024, a shift driven in part by import curbs and in part by heavy local investment. A wave of Chinese chipmakers — including Moore Threads, Biren Technology, Tianshu Zhixin and others preparing IPOs — are racing to close the performance gap and capture cloud and enterprise demand within China and beyond.
Beyond silicon, China’s response includes experiments in alternative architectures. Photonic computing, which uses light rather than electricity to move data between and inside chips, has attracted significant capital and academic attention. Startups such as Lightelligence, founded by MIT alumnus Shen Yichen, have raised large rounds and published results in top journals; their “LightSphere X” optical GPU nodes aim at large‑scale deployment and could sidestep some limitations of traditional semiconductor scaling.
For Nvidia, these developments present a familiar paradox: the Chinese market is too big and innovation too fast to ignore, but domestic advances make future competition more likely. Huang’s public acknowledgement that Chinese companies — notably Huawei — could serve the market in Nvidia’s absence is both conciliatory and candid. It recognises that technological ecosystems are resilient and that the displacement of one dominant supplier does not halt AI progress.
The broader implication is geopolitical as well as commercial. Export controls and a bifurcating tech order have accelerated China’s push for self‑reliance, while open‑source models and alternative compute approaches lower the barrier to entry for local players. Western firms that can operate within China’s regulatory and commercial constraints stand to benefit in the near term, but the longer arc points toward a more diversified global AI infrastructure in which multiple architectures and suppliers coexist.
Huang’s market visit was therefore theatrical but telling: a public signal of commitment and a reminder that the contest over AI infrastructure will be fought on multiple fronts — software ecosystems, domestic chip design, supply‑chain finance and, possibly, emerging photonics platforms. For global players, the question is no longer whether China will develop alternatives, but how quickly those alternatives will reach parity and how multinational companies will adapt their business models in response.