AI adoption: why Chinese companies are walking away from Nvidia

AI adoption at Chinese companies is accelerating: according to Bloomberg, large companies intend to direct up to 46% of their budgets to AI accelerators from domestic manufacturers instead of Nvidia over the next 12 months.

Why almost half of the budget is at stake

Just a few years ago, Chinese technology companies — cloud providers, internet conglomerates, AI startups — relied mainly on American hardware. Nvidia's H100 and H200 accelerators set the industry standard both for performance and for the richness of their software ecosystem: CUDA, cuDNN and TensorRT were the de facto infrastructure for training large language models.

Successive rounds of American export restrictions changed the picture. First the A100 and H100 chips were banned, then the "cut-down" versions designed specifically for the Chinese market, the A800 and H800. When the restrictions extended to the flagship H200 in 2023–2024, China was left practically cut off from Nvidia's top accelerators. It was this external pressure that turned import substitution from a political slogan into a real procurement priority.

Who will fill the vacated niche

The main contenders for the reallocated budgets include several Chinese AI accelerator developers, each of which has carved out its own niche.

• Huawei — the Ascend line (910B, 910C) is positioned as an alternative to the Nvidia H100 for large training clusters
• Cambricon — specialized neural processing units (NPUs) for cloud inference
• Moore Threads — GPU accelerators focused on compatibility with the CUDA ecosystem
• Biren Technology and Enflame — players scaling up capacity with the backing of state funds

None of them has yet reached full performance parity with Nvidia. But in corporate procurement, performance is not the only criterion: guaranteed supply without sanctions risk, price subsidies and compatibility with the local software stack make domestic manufacturers increasingly competitive at industrial purchasing volumes.

What the 46% figure means

The 46% figure does not mean Nvidia is losing the Chinese market entirely. The company continues to sell chips in China that current restrictions allow, notably the H20, which was designed with deliberately reduced specifications. Even so, the intention to send almost half of chip budgets to local manufacturers signals a systemic shift: import substitution is ceasing to be a forced measure and becoming a deliberate long-term strategy.

Notably, Bloomberg obtained the data by surveying companies directly, which makes the numbers reflect real business intentions rather than public rhetoric. At the same time, the shift is speeding up the development of the entire surrounding ecosystem: accelerator makers are increasing investment in compilers, frameworks and cloud platforms that can compete with CUDA on domestic hardware.

"This is a structural shift, not a situational reaction to a single quarterly report," is how

Bloomberg relays the view of analysts familiar with market dynamics.

What this means

If Bloomberg's forecasts materialize, by mid-2027 China's AI accelerator market will be far less dependent on a single American supplier. For Nvidia, that means pressure on its third-largest sales market. For the global industry, it is a signal that a competitive alternative to the established GPU monopoly may emerge faster than Western analysts expected. The success of the Chinese chip stack could also open the way to exporting the technology to other markets looking for alternatives to Western semiconductor infrastructure.

What percentage of their budgets are

Chinese companies shifting from Nvidia to local chips? According to Bloomberg, up to 46% over the next 12 months, a record rate of replacing American AI accelerators with domestic ones.

Which companies in China are moving to local AI solutions?

Cloud providers, internet conglomerates and AI startups.

By what percentage are

Chinese companies cutting Nvidia purchases? Up to 46% of AI accelerator budgets is being shifted to domestic manufacturers, a record rate of substitution.