DeepSeek and Huawei Chip Co-Design Threatens Nvidia's AI Training Grip

By Priya Raman

DIGITIMES reports that a chip co-design effort between DeepSeek and Huawei could shift AI training workloads away from Nvidia hardware. The claim rests on a single headline. The outlet published no figures, no dates, and no named sources in the summary distributed through its feed.

That absence matters. Trade claims about Nvidia displacement arrive weekly. Most collapse under scrutiny. This one, at minimum, names two specific companies and a specific mechanism: co-design (the joint engineering of a model architecture and the silicon that runs it).

Why would co-design matter?

DeepSeek builds large language models. Huawei builds AI accelerators, including the Ascend series, which competes with Nvidia's data-center GPUs. Co-design means DeepSeek would tailor its training pipeline to Huawei's chip architecture rather than to Nvidia's CUDA software stack (the proprietary toolkit most AI labs currently depend on).

The logic is plausible. Huawei cannot yet match Nvidia on raw performance. A model designed around Ascend's specific strengths could partially close that gap. China's export controls already push domestic labs in this direction.

But plausible is not proven. The report offers no benchmark data, no training-run results, and no timeline. DIGITIMES frequently cites supply-chain sources it does not name. Readers should treat the claim as informed speculation, not verified fact.

What is not said?

Three things are missing from the report.

• No silicon generation is specified. Ascend chips span several capability tiers.
• No workload scale is given. Training a frontier model differs from fine-tuning a smaller one.
• No commercial agreement between the two firms is confirmed.

Each gap leaves room for a weaker interpretation. A partnership this consequential would normally arrive with press releases and specifications. None have appeared.

How large is the potential impact?

Nvidia dominates AI training hardware today. Its GPUs power the majority of large-model training runs outside China. US export restrictions already bar its top-tier chips from the Chinese market, which makes domestic alternatives structurally attractive rather than merely competitive.

If the co-design claim holds, the effect would concentrate in China first. DeepSeek previously demonstrated that efficient training methods can reduce hardware demands. Pairing that discipline with Huawei silicon would extend the strategy from software into hardware.

If the claim fails, the status quo persists. Nvidia retains its position, and the headline joins the long list of displacement stories that never materialized. Both outcomes remain live. The evidence published so far does not distinguish between them.

What should watchers track?

Verification will come through observable signals, not statements. A training run on Ascend hardware at frontier scale would constitute evidence. Public benchmarks comparing throughput and cost per token would constitute evidence. Statements from either company, with numbers attached, would constitute evidence.

Until then, the report stands as a signal of intent at most. The market should discount accordingly.