Huawei chips scramble hits firms

China’s AI race just got a lot more physical. The immediate news is not another flashy model benchmark — it’s a scramble for chips. After DeepSeek released a preview of V4 on April 24 and tied the model more closely to Huawei hardware, big Chinese internet companies including ByteDance, Tencent, and Alibaba moved to secure more of Huawei’s Ascend 950 chips. That matters because compute, not just code, is now the bottleneck inside China’s AI push. (msn.com) ### What actually changed? DeepSeek’s earlier headline models were still associated with Nvidia hardware for training, but V4 marks a clearer shift toward Huawei’s stack. Huawei said last week that its Ascend SuperNode systems based on the Ascend 950 series fully s(msn.com)for chips rises fast because every prompt needs hardware behind it. (msn.com) ### Why are firms scrambling now? Because V4 is not just a lab demo. It is being deployed, and deployment creates immediate demand for inference chips across cloud providers, GPU rental firms, and internal AI teams. Reuters’ reporting says ByteDance, Tencent, and A(msn.com)rt on compute starts calling the same supplier. (money.usnews.com) ### Why is the 950PR the key part? The 950PR looks like Huawei’s first chip that customers see as a real volume alternative, not a symbolic substitute. Reuters-based reports say it outperforms Nvidia’s H20 — the most advanced Nvidi(money.usnews.com)ll place the 950PR behind Nvidia’s more advanced H200 class. (invezz.com) ### Why does Nvidia matter so much here? Because the whole Chinese market has been shaped around what Nvidia could and could not legally sell there. The H20 existed as a compliance chip — a cut-down product designed to fit earlier U.S. export rules. Then Washington tightened licensing in April 2025, whi(invezz.com)e and fewer foreign options. (techcrunch.com) ### Can Huawei actually supply enough chips? That is the hard part. Reuters-based reports say Huawei aims to ship about 750,000 units of the 950PR this year, with mass production starting in April and larger shipments expected in the second half of 2026. That sounds big, but demand may outrun(techcrunch.com)h is one thing; sustained volume is another. (invezz.com) ### Why is this bigger than one model launch? Because it signals a hardware fork in the AI world. China is no longer just trying to copy the software layer while waiting for foreign chips. It is pairing local models with local accelerators and local server systems. Think of it less like swapping one com(invezz.com)t, switching back gets harder. (money.usnews.com) ### So what is the real constraint? Manufacturing depth. Huawei can win design slots and still get stuck on packaging, yields, memory, and system integration. That is why this story feels important: demand is now real, but supply is still fragile. The next phase of China’s AI buildout may be decided less by model cleverness than by who can physically deliver enough working machines. (businesstimes.com.sg) ### Bottom line? This is a chip story wearing an AI headline. DeepSeek gave Huawei proof that a serious frontier-style model can lean on domestic hardware, and China’s biggest tech firms reacted immediately. If Huawei can ship at scale, China’s AI stack gets more self-contained. If it cannot, the shortage becomes the story. (msn.com)