China's Push to Wean Tech Firms Off Nvidia Chips Faces Uphill Battle Despite Huawei Gains
Summary
- • China's tech firms still depend on Nvidia chips for AI training despite Beijing's self-sufficiency push
- • Huawei says Ascend outsells Nvidia in China; one tracker puts Nvidia at 34%, Huawei at 28%
- • DeepSeek reportedly plans 160,000+ Huawei Ascend chips for inference, keeping Nvidia for training
- • Software lock-in and performance gaps reportedly slow the shift; gap may last until 2030
Details
China is struggling to move tech firms off Nvidia chips despite Beijing's self-sufficiency drive
Government policy favors domestic AI hardware, but major Chinese companies continue to rely on Nvidia, so substitution is slower than policy goals imply.
Huawei claims its Ascend chips outsell Nvidia's in China
This is Huawei's own claim and conflicts with third-party market share estimates, so it is company framing rather than established fact.
H1 2026 China data-center AI accelerator share reportedly: Nvidia 34%, Huawei 28%
Per one market-research estimate (IDC). Nvidia still leads, though Huawei is close behind; counting differences may explain the gap with Huawei's claim.
Chinese labs still rely on Nvidia for training because of CUDA vs CANN software lock-in and performance gaps
CUDA is Nvidia's mature developer ecosystem and CANN is Huawei's equivalent. Porting training code is difficult and Ascend reportedly trails in performance.
Tech giants are reportedly repurposing over 1 million Nvidia gaming and workstation chips for AI this year
This workaround reflects shortages of data-center-grade accelerators and shows how much demand for Nvidia silicon remains.
DeepSeek reportedly plans at least 160,000 Huawei Ascend 950DT chips for inference, keeping Nvidia for training
A split approach: domestic chips take deployment workloads while performance-critical training stays on Nvidia hardware.
Transitioning from Nvidia to domestic chips reportedly costs about 50% more in money and time
This discourages rapid migration, since firms pay a premium in engineering effort and delay. The basis for the figure is unclear.
The performance and supply gap with Nvidia is expected to persist until at least 2030
This is a reported expectation, not a certainty. If it holds, China's dependence on Nvidia hardware or workarounds continues through the decade.
Industry Update = sector development, Market Impact = competitive position, Stat = market data, Tech Info = technical background, Infrastructure = hardware supply, Strategy = company plans, Financials = cost, Insight = outlook
What This Means
China's effort to replace Nvidia with homegrown chips is advancing but incomplete, particularly for training frontier AI models. Huawei is gaining ground mainly in inference, though its share claims are contested. Chinese firms pay extra in money and time, or use workarounds, which could slow their AI development. Much of this comes from secondary reporting, so figures are best read as estimates.