DeepSeek shuts out US chipmakers, hands early access to Huawei

Chinese AI lab DeepSeek has not given US chipmakers Nvidia and AMD early access to its upcoming V4 model, breaking industry practice. Instead, it granted domestic suppliers like Huawei a head start to optimise for their processors, per Reuters. The move highlights how efficiency-focused Chinese AI models are helping local chipmakers close the gap in inference workloads.

Bottom line — DeepSeek's V4.1-Flash model activates only 8 billion of 552 billion parameters per input, per GuruFocus.

Go deeper (7)

• DeepSeek previously worked closely with Nvidia's engineers, but for V4 it gave Chinese chipmakers several weeks' lead time, two sources told Reuters.
• Chinese models now account for over 60% of tokens used among the top 50 AI models, up from under 20% late last year, according to Seeking Alpha data cited by GuruFocus.
• A senior Trump administration official told Reuters DeepSeek trained its latest model on Nvidia's most advanced Blackwell chip in mainland China, which appears to violate US export controls.
• Four Chinese AI chip startups — dubbed the 'four dragons' — have gone public or filed to in the past two months, per CNBC, as Beijing pours hundreds of billions of dollars into its domestic chip industry.
• Elon Musk noted at Davos that China's electricity generation growth far outpaces the US, a key bottleneck for AI compute, CNBC reported.
• Huawei is developing a CUDA equivalent called CANN, but analysts at Omdia say persuading developers to abandon Nvidia's software ecosystem remains a major challenge.
• Analysts at Bernstein argue Chinese AI chips are only cost-competitive for inference inside China, as Nvidia still dominates even for inference tasks.

Read the reporting