Why Chinese Tech Giants Are Dropping Nvidia GPUs for Domestic Chips

Recent U.S. export restrictions have forced Nvidia to create "cut‑down" AI GPUs for the Chinese market. Although these special‑edition chips comply with the rules, their performance is significantly reduced, leading major Chinese cloud providers—including Tencent, Alibaba, Baidu and ByteDance—to scale back orders and explore domestic solutions.

Market Impact

Chinese technology firms account for roughly one‑fifth of Nvidia's revenue. An Omidia report shows that in Q3 2023, the top twelve customers for Nvidia's H100 GPU were dominated by Tencent, Baidu, Alibaba and ByteDance. This shift reflects a broader trend of decreasing reliance on Nvidia's high‑end products.

Key Reasons for the Shift

Regulatory uncertainty: Ongoing U.S. restrictions make it unclear whether Nvidia can continue supplying chips to China.

Performance degradation: The special‑edition GPUs have been deliberately throttled. For example, the H20 chip delivers a total processing performance of 2368 with a density of 2.9 (below the 3.2 regulatory limit); the L20 scores 1912 with a density of 3.1; and the L2 records a density of 5.2 but only 1544 total performance, falling short of the 1600 limit.

Supply stability: Domestic manufacturers can offer a more reliable supply, prompting Chinese cloud operators to consider a "Plan B" based on home‑grown silicon.

Performance Comparison

In practical terms, the H20’s peak theoretical compute is 296 TFLOPs, far below the consumer‑grade RTX 4090 (661 TFLOPs) and the flagship H100 (1 979 TFLOPs). This gap diminishes the appeal of Nvidia’s downgraded offerings for high‑end AI workloads.

Domestic Chip Opportunities

Chinese vendors are gaining traction. Huawei, confirmed by CEO Jensen Huang, has secured three major customer channels: internet companies (e.g., 5 000 Ascend 910B chips ordered in 2023), state‑owned telecom operators (China Telecom’s $390 million AI server purchase, China Unicom’s $20 million spend), and government‑backed AI computing centers that have been buying Huawei AI chips since the 2022 graphics ban.

Despite these advances, Nvidia’s CUDA ecosystem and its extensive hardware‑software stack remain unmatched, and even AMD struggles to compete directly.

Outlook

Analysts from TrendForce predict that Nvidia’s share of China’s high‑end AI chip market could fall from about 80 % to 50‑60 % within five years as U.S. restrictions tighten. However, many Chinese customers still view Nvidia as the primary supplier for the next 12 months, especially for workloads that demand the full CUDA stack.

In the long term, continued export controls are likely to accelerate domestic chip development, but widespread substitution of Nvidia GPUs with local alternatives will remain a gradual process.

Overall, the evolving regulatory environment presents both challenges and opportunities for China’s AI chip industry, with domestic solutions poised to grow but still facing significant ecosystem hurdles.