Nvidia shares slipped nearly two percent in early trading on Tuesday after a report that Chinese artificial intelligence laboratory DeepSeek is developing its own inference chip, the latest signal that the gap between US chip designers and their Chinese counterparts is closing faster than Western semiconductor executives had expected. The report, carried by International Business Times on July 7, sent ripples through the broader semiconductor index and reignited a debate about how durable Nvidia’s near-monopoly on the hardware that trains and runs frontier AI models really is, given that the company’s largest customers are now actively trying to replace it.
DeepSeek’s move is the most visible piece of a broader pattern inside China’s AI industry. The Hangzhou-based lab stunned Western competitors in early 2025 when it released a frontier-grade open-weight model trained on a comparatively small cluster of Nvidia chips, and it has been telegraphing its chip ambitions ever since. The new inference processor, which people familiar with the project describe as a purpose-built design optimised for the specific matrix math that large language models use at serving time, would be DeepSeek’s first commercial silicon and the first credible Chinese inference accelerator designed by a model lab rather than a legacy chipmaker.
Why an inference chip, not a training chip
The decision to target inference rather than training is deliberate and revealing. Training a frontier model is dominated by Nvidia’s H100 and B200 accelerators, hardware that is so tightly optimised around the company’s CUDA software stack that no Chinese alternative has yet managed to match the throughput that frontier labs demand. Inference, by contrast, is a more heterogeneous workload. Modern deployments mix batch processing, low-latency serving for chat applications, and high-throughput pipelines for code generation, each of which stresses different parts of a chip’s design. The performance bar is also lower: an inference chip that delivers 70 to 80 percent of Nvidia’s throughput at half the price is commercially attractive, particularly for a Chinese market whose customers are increasingly cost-sensitive.
DeepSeek’s design reportedly focuses on high-bandwidth memory and aggressive on-chip caching, two areas where Chinese fabrication has closed the gap with Taiwan Semiconductor Manufacturing Company. It is expected to be manufactured at a domestic foundry using a mature process node, sidestepping the export controls that have blocked the most advanced lithography equipment from reaching China. That constraint also explains why DeepSeek has chosen inference over training. Building a training chip that competes with the B200 requires bleeding-edge fabrication that no Chinese foundry can currently provide, while building a competitive inference chip is achievable with the process nodes the country already has.
What the market is reading into the report
- 2 percent drop in Nvidia shares on the news, the first negative session of the week for the stock.
- Spillover weakness across the broader semiconductor index, with peer names also giving back gains.
- Renewed interest in Chinese chip design houses, which would benefit from a credible domestic inference customer.
- A reminder that the “one customer, one chip” model that has defined the AI hardware market is breaking down, as model labs begin to design for themselves.
For Nvidia, the practical impact is muted in the short term. DeepSeek’s chip, if it ships on the timeline people are suggesting, will serve a small fraction of the global inference market, and only inside China. The larger signal is strategic. If the most demanding AI labs in the world are designing their own silicon because they have grown frustrated with Nvidia’s pricing and supply constraints, the company’s customers are quietly telling it that the current arrangement is no longer stable. That is the kind of message that moves stock prices more than any single quarter’s earnings can.
The wider chip race
DeepSeek is not alone. Huawei’s Ascend product line has matured into a credible alternative for many domestic training workloads, and several smaller Chinese design houses are working on inference accelerators aimed at the same market. Anthropic is reported to be in early discussions with Samsung about a custom chip, and Amazon’s Trainium and Google’s TPU are already eating into the inference share of the largest hyperscalers. The story the market is slowly waking up to is that “Nvidia everywhere” was always a transitional state, and that the transition is now visibly underway.
Whether the move ends up material for Nvidia’s earnings is a separate question. The company has spent the last two years building out its own product lines beyond the data-centre accelerator, including a growing networking business, a software platform, and an industrial pipeline for autonomous systems and robotics. Those businesses are not directly threatened by a Chinese inference chip, and the bulk of the company’s revenue still comes from a small number of American hyperscalers whose purchasing decisions are shaped by a different set of incentives. But the era in which Nvidia could price its chips as if no substitute existed is ending, and the market is right to mark that down, however slightly, whenever the next credible substitute appears in the press.