Skip to content
LIVE
The Executives BriefThe Executives Brief

Alibaba is training its next AI model on homegrown chips as US curbs tighten

Nikkei Asia reports the next-generation model is being trained on Alibaba's own silicon, the clearest sign yet that China's biggest cloud player is designing Nvidia out of its critical path.

ByTurki Al-MutairiBusiness Desk, The Executives Brief
·4 min read
Alibaba is training its next AI model on homegrown chips as US curbs tighten
Executive summary

Alibaba is training its next-generation AI model on its own chips, according to Nikkei Asia, as US export controls on advanced semiconductors continue to tighten. The move signals that China's largest cloud provider is treating custom silicon as a supply-chain hedge, a shift with direct consequences for Nvidia's China business, Huawei's role as the domestic default, and every rival lab renting compute.

Alibaba is training its next-generation AI model on its own chips, Nikkei Asia reports, a detail that lands harder than it sounds. Training, not just running, is the step in the AI stack where American silicon has been hardest to replace, and it is the step the US export controls are built to bite. If China's largest cloud company is genuinely pushing a next-gen model through homegrown accelerators, then the most expensive and technically unforgiving part of its AI roadmap no longer sits entirely behind a switch Washington controls.

Here is the distinction that makes this matter. Inference, the work of answering a prompt once a model is already built, is comparatively forgiving: you can run it on less powerful silicon, and Chinese firms have done exactly that for years. Training is the opposite. It means wiring thousands of chips together for weeks or months, moving data between them at brutal bandwidth, and surviving a single failure without torching the entire run. That is why Nvidia's data-center GPUs became the default everywhere. Any credible move away from them says something about how far the domestic alternative has come, and how much cost and risk Alibaba is willing to absorb to get there.

The backdrop is a ratchet, not a single event. Washington began restricting exports of advanced AI chips, and the tools used to make them, to China in 2022, and has tightened the rules repeatedly since. Each round narrowed what Nvidia could legally sell into the market and pushed Chinese buyers toward stockpiles, workarounds, and homegrown designs. Later rounds extended beyond the processor itself into memory and related components, squeezing the plumbing that large-scale training depends on. For Alibaba, the logic is not sentimental. If the supply of top-end accelerators can be severed by policy, designing your own silicon stops being a research project and becomes a supply-chain hedge.

Alibaba is not starting from zero, which is part of why this is credible. Its chip design arm, T-Head, has previously shipped AI silicon aimed at inference workloads, and the company has spent years building Alibaba Cloud into one of Asia's largest providers, a business with enough internal scale to justify custom parts. It also ships the Qwen family of models, released in open weights and widely downloaded, fine-tuned, and deployed by developers well beyond China. A company with its own models, its own cloud, and its own chip team can, in theory, optimize all three together in ways a pure buyer of Nvidia hardware cannot. That vertical stack is the real asset here, not any single benchmark.

The hard part is not the design. It is the manufacturing. Export controls cover not only chips but the advanced equipment needed to fabricate them, which is why China's domestic foundries still trail the leading edge. A clever accelerator architecture is worth little if nobody can print it at volume and at a competitive yield. That constraint is the real test of the Nikkei Asia report: whether Alibaba's in-house silicon is a genuine training workhorse or a carefully framed hedge that still leans on imported parts, stockpiled inventory, or older process nodes. The answer will not show up in a press release. It will show up in how fast the next model ships and how much Alibaba spends to train it.

The second-order effects run in several directions. Nvidia's China revenue has already been reshaped by successive rounds of restrictions, and every credible domestic substitute makes that market harder to win back. Huawei has positioned itself as China's default AI chip supplier, so Alibaba choosing to self-supply rather than buy from a domestic rival is a statement about where the strategic value sits. Self-supply also changes the unit economics: a chip you own amortizes across years of internal demand, while a chip you rent carries someone else's margin on every token. For Tencent, Baidu, ByteDance, and every other Chinese lab renting compute, it resets the build-versus-buy question at the exact moment the cloud price war is intensifying.

Zoom out and the pattern is bigger than one company. Sovereign AI, the push by governments and large enterprises to control their own compute, models, and data, is now a global theme, visible in the Gulf, in Europe, and in Washington's own industrial policy. Alibaba training a next-gen model on its own chips is that theme showing up on a single balance sheet, and it is the version that matters most commercially, because cloud capacity is where AI ambition either becomes a product or stays a slide. The strategic read for any executive or board: identify which parts of your AI stack depend on suppliers you cannot control, and price the hedge before you need it. Alibaba just showed what paying for that hedge looks like. It is not cheap, and it is not fast, and it is increasingly the price of staying in the race.

Executive ActionsLocked

This story's Key Insights and Take-aways are locked.

Create a free account to unlock Executive Actions for one credit.

Register to Unlock

Always free for Executives Club members. Join the Club

More in Business