DeepSeek's V4 preview reportedly runs partly on Huawei chips

AI models need two kinds of hardware muscle. One is inference — serving answers after the model already exists. The other is training — the much harder, more compute-hungry phase where the model gets built in the first place. That distinction is why DeepSeek’s V4 preview matters. The interesting part is not just that the model can run on Huawei chips, but that Huawei says its chips were used for part of the training too. ### What actually launched? DeepSeek released a preview of V4 on April 24, with two variants: V4-Pro and V4-Flash. The model is open-source, aimed at long-context and agent-style tasks, and marks DeepSeek’s biggest release since the models that made it a breakout name in China’s AI race. Reuters coverage tied the launch directly to Huawei hardware support, which is the real news hook here. (techwireasia.com) ### Why is Huawei the important part? Because DeepSeek’s earlier headline models, including V3 and R1, were trained on Nvidia chips. China’s AI firms have spent the last few years trying to work around U.S. export controls while still relying heavily on Nvidia’s software stack and accelerators. V4 looks like a step away from that dependency — not a clean break, but a real step. (tech.yahoo.com) ### Why does training matter more than inference? Inference is the easier claim. Lots of companies can say a model “runs” on alternative chips after the heavy lifting is done somewhere else. Training is different — it is where the software stack, memory bandwidth, interconnects, and cluster reliability all get stress-tested at once. If Huawei silicon can handle even part of that job for a frontier Chinese model, the story stops being a demo and starts looking like ecosystem progress. (techwireasia.com) ### So was V4 fully built on Huawei chips? No clear evidence says fully. The careful wording matters. Reuters-linked coverage says Huawei chips were used for part of V4-Flash’s training, and that V4 is supported on Huawei’s Ascend 950-based supernode clusters. That is a meaningful milestone, but it is not the same as proving Huawei has already replaced Nvidia across the whole pipeline. (techwireasia.com) ### What changed after the launch? Demand moved fast. Within days of the V4 release, Reuters reported that Chinese tech firms were scrambling to secure Huawei Ascend 950 chips, with the launch helping trigger a surge in orders. That is the market test that matters — not press release language, but whether buyers change behavior. (techwireasia.com) ### Why does this hit Nvidia’s position? Nvidia’s moat has never been just raw chip performance. It is the whole package — CUDA, developer habits, tools, and the assumption that serious AI training means Nvidia by default. China has been trying to chip away at that stack from below. DeepSeek V4 suggests the effort is getting more credible, at least inside China’s own market where export controls make domestic substitution more urgent. (money.usnews.com) ### Why should non-AI buyers care? Because this spills beyond chatbots. If Chinese hardware and software stacks get more self-contained, buyers of connected equipment — including scientific and lab systems with embedded compute — may face more region-specific dependencies. The catch is not that every device suddenly becomes unusable. It is that compatibility, update paths, and export exposure could diverge more sharply between U.S.-aligned and China-aligned tech stacks. (money.usnews.com) That is the kind of supply-chain split that shows up late, after procurement decisions are already locked in. ### Bottom line? DeepSeek’s V4 preview does not prove Huawei has caught Nvidia outright. But it does show something narrower and important — China’s AI stack is getting better at doing the hard part on domestic chips, not just the easy part. That makes the competition more structural, and more durable, than a one-off model launch. (techwireasia.com) (money.usnews.com)