DeepSeek, other Chinese AI firms pivot to Huawei chips to bypass Nvidia constraints

- DeepSeek’s V4 launch pushed Chinese cloud giants toward Huawei’s Ascend 950 chips, while Jensen Huang joined Donald Trump’s Beijing trip on May 13. - ByteDance, Tencent, and Alibaba reportedly rushed to place Ascend orders after V4 showed Huawei silicon could handle part of frontier-model training. - The real shift is structural — China now has a usable domestic AI stack, even if Nvidia still leads at the top end.

AI chips are the bottleneck here. Not models, not apps, not even data centers in the abstract — the actual accelerators that train and run large models. That bottleneck is why this story matters. The news is that DeepSeek’s latest model gave Huawei a live demonstration that its chips are no longer just a backup plan, and on May 13 Jensen Huang boarded Donald Trump’s China trip just as Chinese buyers started hoping Nvidia’s H200 might get unstuck again. ### What changed this week? The immediate trigger was DeepSeek V4. After the model’s late-April debut, demand for Huawei’s Ascend 950 jumped, with ByteDance, Tencent, and Alibaba all reaching out for new orders. That matters because V4 was not just “compatible” with Huawei hardware — Huawei said its chips were used for part of training, which is the hard part of the stack, not just inference. (businesstimes.com.sg) ### Why is training the key detail? Inference is what happens after a model is built — answering prompts, generating code, spitting out images. Training is the expensive phase where the model learns in the first place. If Huawei chips can help with training, then Chinese labs are not merely finding a cheaper server for deployment. They are building a path around Nvidia’s most important choke point. That is why this landed as more than a procurement story. (businesstimes.com.sg) ### Does this mean Huawei has caught Nvidia? Not exactly. Nvidia still leads on the frontier — especially on software maturity, developer tooling, networking, and the sheer efficiency of training giant models at scale. The catch is that export controls changed the question. Chinese firms do not need Huawei to beat Nvidia everywhere. They need Huawei to be good enough, available enough, and integrated enough that domestic model development can keep moving. (techwireasia.com) DeepSeek just made that case more believable. ### Why does DeepSeek matter so much here? Because DeepSeek is not a lab demo. It has become a proof point for efficient Chinese model building under constraint. Earlier DeepSeek systems were trained on Nvidia hardware, but V4 tightened the technical link with Huawei. Once a visible model maker shows the stack works, the next buyers do not have to make a patriotic bet — they can make a practical one. Basically, DeepSeek de-risked Huawei for the rest of the market. (economictimes.indiatimes.com) ### So why is Jensen Huang in Beijing? Because Nvidia does not want to lose China by default. Huang joined Trump’s trip after a late invitation, and the signal was obvious — Chinese customers are hoping the visit might revive stalled efforts to sell H200 chips locally. That does not mean a deal is done. But it shows Washington’s restrictions and Nvidia’s commercial interests are now pulling against an increasingly credible Chinese substitute. (techwireasia.com) ### What is Huawei’s opening? Scale and inevitability. Huawei expects AI chip revenue to jump at least 60% this year, helped by domestic demand created by the same export rules that were supposed to slow China down. That does not erase Huawei’s own manufacturing constraints — supply is still tight — but it does mean local customers now have a reason to optimize around Ascend instead of waiting for policy relief from Washington. (money.usnews.com) ### Is the market splitting in two? More like three ways. The top frontier labs still want the best U.S. chips. Chinese firms are investing in domestic substitution with Huawei. And everyone in between is building optionality — using whatever hardware, software, and policy window is available. That is the bigger shift. The AI supply chain is no longer converging on one global standard. It is fragmenting into parallel stacks shaped by geopolitics as much as by performance. (msn.com) ### Bottom line? The important thing is not that Huawei suddenly “won.” It is that the fallback is turning into a real platform. Once that happens, export controls stop being a clean off-switch and start looking more like a subsidy for local substitution. (techwireasia.com) (businesstimes.com.sg)

Get your own daily briefing

Scout delivers personalized news, insights, and conversations tailored to your role and industry.

Download on the App Store

Shared from Scout - Be the smartest in the room.