DeepSeek is designing its own inference chip to lean less on Nvidia and Huawei, Reuters reports

Reuters reported on July 7 that DeepSeek, the Chinese lab whose R1 model rattled AI stocks last year, is designing its own AI chip. The report rests on three people familiar with the matter, and DeepSeek did not comment. Read it as a credible scoop about intent, not a product anyone can benchmark yet.

What the chip is for

The chip is designed for inference. It is not a training chip, and that split is the whole point. Training is the expensive, one-time job of building a model. Inference is the cheaper per-run job of answering each query, repeated billions of times once the model is live. Inference silicon can be simpler and more specialized, and inference is where the recurring compute bill actually sits. Going after it first is the pragmatic call: it is the more tractable design problem and the bigger long-run cost.

Why DeepSeek would bother

DeepSeek trains and serves its models on Nvidia and Huawei hardware. US export controls bar Chinese firms from buying Nvidia's top parts, and Beijing has been pushing its national champions onto domestic silicon. Huawei's Ascend line is the main homegrown option, and Reuters notes Huawei supplies roughly half of China's $50 billion domestic AI-chip market. A DeepSeek part would cut its dependence on both a restricted foreign supplier and a domestic rival it would rather not rely on.

It also follows the pattern of every large model shop. Google has TPUs. Amazon has Trainium and Inferentia. OpenAI has its own inference chip in progress with Broadcom. If you run models at scale, owning the inference hardware is how you control cost and supply. DeepSeek wanting the same is not surprising.

What is not established

No tapeout, no confirmed foundry, no performance numbers, no ship date. "In talks with design, foundry, and memory partners for about a year" and "hired chip engineers quietly" describes a program that is early. Chip design is a multi-year slog for teams that have done it before, and manufacturing an advanced part inside China's export-constrained toolchain is a second hard problem stacked on the first. A design on paper is a long way from a chip in a rack.

The market reaction was small and about right. Nvidia slipped roughly 1.6% in premarket trading, then the story moved on. That is the market pricing in a possibility, not reacting to a threat that has landed.

For scale, Reuters reported in June that DeepSeek was set to raise about $7 billion at a valuation between $52 billion and $59 billion, its first outside capital after years of refusing any. A funded lab designing its own inference silicon is a real signal about where it thinks its constraint is. Whether the chip works is a question for 2027 at the earliest.