DeepSeek Turns to Huawei, Not Nvidia, for Final Tuning of Its Newest Model

A shift has occurred in the AI industry, and its significance goes well beyond a single corporate decision. DeepSeek, the Chinese company whose language models have repeatedly rattled Silicon Valley over the past year, has entrusted the final optimization of its newest model not to Nvidia but to Huawei and other Chinese makers of compute accelerators. Reuters reports this, citing sources familiar with the matter.

To grasp the scale of the decision, it helps to understand how large language models are built. Development has several stages: the model is first trained on enormous datasets, then goes through additional tuning, and finally reaches a stage known as final tuning. At this last step, accelerator vendors optimize the model for their own hardware architecture, aiming for maximum performance and efficiency. Until now, this stage has almost unchallenged belonged to Nvidia, whose GPUs became the gold standard for training and deploying neural networks. DeepSeek's decision breaks that unspoken monopoly.

The context is critical here. Since 2022, Washington has steadily tightened export restrictions on shipments of advanced chips to China. Nvidia's flagship accelerators were banned, first the A100 and H100, then their "cut-down" versions designed specifically for the Chinese market. Each new round of sanctions narrowed Chinese companies' ability to use American hardware. But it also created a powerful incentive to build a domestic ecosystem. Despite its own sanctions problems, Huawei is actively developing its Ascend line of accelerators, positioned as a direct alternative to Nvidia's products for machine learning workloads.

That DeepSeek bet on Huawei specifically for the final tuning stage says a lot. This is not merely a symbolic gesture, nor a forced move because of a lack of access to Nvidia chips. Final optimization requires deep integration between software and hardware, and close collaboration between the model's developers and the engineers building the accelerators. In effect, DeepSeek is declaring that the Chinese hardware platform has matured enough to be relied on at the most critical moment: right before a product ships.

For Nvidia, this is a worrying signal, though not a catastrophic one. Jensen Huang's company still dominates the global AI accelerator market, and its data center revenue keeps setting records. But China is not just another market. It is the world's second-largest economy and a country with an ambitious state program for developing artificial intelligence. Every time a major Chinese player shows it can function without American technology, Nvidia's position in that market weakens irreversibly. Even if sanctions are eased someday, the habit and the ecosystem built around Ascend will not go away.

For Huawei, this is a moment of triumph. The company that Washington tried to cut off from global semiconductor supply chains is now a key partner to one of the most promising AI startups on the planet. This strengthens Huawei's standing not only as a maker of smartphones and telecom equipment but as a full-fledged player in AI infrastructure. The Ascend line gets a powerful reference case that will help attract other customers in China and beyond.

The broader picture also deserves attention. DeepSeek has already shown it can build competitive models with substantially lower compute costs than its American rivals. If it now turns out that final optimization on Chinese hardware is no worse in quality than optimization on Nvidia, that will upend assumptions about how critical dependence on American chips is for developing cutting-edge AI.

We are watching a parallel technological universe take shape. Just as powerful, just as ambitious, only built on a different foundation. And DeepSeek's decision to entrust Huawei with the final tuning of its model is not just corporate news. It marks the point at which the US-China technological decoupling has passed the point of no return in yet another critically important area.