The news, reported by Reuters citing sources close to the company, goes straight to the heart of the US-China tech fracture. DeepSeek, the Hangzhou-based lab that has captured global attention for the efficiency of its LLMs, is reportedly designing its own AI chip. This is not a technical footnote, but a turning point for the industry.

The stated motive is to sidestep US export restrictions on advanced semiconductors. But the stakes are higher: if until now DeepSeek has written software optimized to run on Nvidia GPUs (or local alternatives like Huawei's Ascend), it now wants to define hardware tailored to its workloads. It means abandoning the software-first model and embracing extreme vertical integration, where silicon architecture and algorithms are co-designed.

The geopolitical push and the domestic chip race

DeepSeek's move does not happen in a vacuum. Since the United States tightened export controls on GPUs and AI accelerators to China, Chinese companies have accelerated development of local alternatives. Huawei with its Ascend series, Biren Technology and other startups have tried to fill the gap left by Nvidia, with mixed results. DeepSeek fits into this pattern with a crucial difference: it is not a semiconductor company, but a lab that trains and serves models. It knows first-hand the bottlenecks of inference and training, so it can sculpt silicon around the real needs of its workloads, bypassing intermediaries.

This shift is driven by structural necessity. Western chips are no longer available, or arrive in downgraded versions. For a lab aiming at the frontier of LLMs, controlling the entire chain—from the compute node to the data center—becomes vital. It is no longer just about performance, but about technological sovereignty and operational continuity.

The hidden costs: EDA, foundries and dependencies

Designing an AI chip involves more than just architecture. It requires electronic design automation (EDA) tools, dominated by US companies like Synopsys and Cadence, which are also subject to restrictions. And it requires a foundry capable of manufacturing the chip at advanced nodes: today, SMIC, China's leading foundry, struggles to match TSMC's levels. DeepSeek will therefore have to navigate a path full of technical and geopolitical hurdles that go well beyond writing RTL code.

The real signal, however, is that frontier AI in China is no longer content to adapt to existing offerings. It wants to set the specifications. This is the same logic that drove Google to develop TPUs or Amazon to create Trainium chips: serving internal workloads with efficiency and costs unattainable with general-purpose products. The difference is that DeepSeek operates in a context of geopolitical bottleneck, not pure economic optimization.

Winners and losers (and what is cracking)

In the short term, Nvidia is not trembling: shipments to China are already restricted, and DeepSeek represents a tiny fraction of its market. But in the medium term, the story changes. Every major Chinese lab that adopts domestic silicon erodes the CUDA ecosystem, which for years has acted as a competitive moat. It is not a matter of direct sales, but of standards: if Chinese AI software develops on non-Nvidia platforms, the entire ecosystem of tools and libraries could bifurcate, making it harder for the US company to maintain its centrality.

For Chinese foundries, the demand for custom AI chips represents an enormous opportunity, but also a technological proving ground. And for the open-source community, there is another layer: DeepSeek has already released models under open licenses. If one day it were to share parts of its hardware design as well, it could open a season of distributed innovation that is hard to predict.

In any case, the direction is clear. AI labs are no longer mere consumers of compute, but producers of infrastructure. The silicon supply chain is splitting along political lines, and AI is becoming a battleground for digital sovereignty.