DeepSeek, the Chinese AI lab that caught global attention in early 2025, is developing its own chip to run AI models, according to a Reuters exclusive reporting on multiple sources familiar with the project. The Hangzhou-based company is already in talks with manufacturing partners and has begun hiring engineers Reuters.

The chip is designed for inference — the work of running an already-trained model to generate outputs, as opposed to training a model from scratch. Think of training as creating a blueprint; inference is using that blueprint to build something. Inference is generally less computationally demanding than training, and it's the part that happens every time a user sends a prompt and waits for a response. The goal is to reduce DeepSeek's reliance on chips from Huawei and Nvidia, the two suppliers that have powered most of the company's computing to date.

A Pattern of Hedging

DeepSeek's pursuit of custom silicon follows an uneven eighteen months of shifting strategy around chip supply. In June 2025, a US official alleged that DeepSeek had acquired Nvidia H100 chips after the United States had already banned their sale to China, and that the company had aided Chinese military efforts while evading export controls Reuters. Two months later, DeepSeek released an upgraded model explicitly compatible with Huawei's Ascend chips — China's homegrown alternative to Nvidia for AI work.

By February 2026, the picture shifted again: a Chinese official said DeepSeek's newest model had actually been trained on Nvidia's most advanced chips despite the US export ban Reuters, and days later Reuters reported that DeepSeek was withholding that same model from US chipmakers, including Nvidia itself, saying it was for performance tuning — a departure from the open-access practice common in the industry Reuters. Then in April, DeepSeek unveiled a new model, V4, tailored specifically to Huawei's Ascend chips. The release drew a muted response from the market; competition from rival Chinese models like Kimi and Qwen has intensified Reuters.

Taken together, these moves suggest DeepSeek is diversifying its compute strategy rather than betting everything on one supplier. The company is keeping Nvidia access when it can, supporting Huawei's chips when it's strategically useful, and now building its own silicon. This playbook — using available hardware now, adapting to domestic alternatives when needed, and moving toward vertical control later — reduces exposure to any single bottleneck.

Why Inference Chips Are a Logical Start

Building a custom inference chip is materially easier than building a training-class chip. Inference is less demanding: the system can use lower-precision arithmetic (fewer decimal places in calculations), simpler memory structures, and doesn't need as much data transfer bandwidth. This is why Google's TPU line, Amazon's Inferentia, and a host of other custom accelerators have historically started with inference before — if ever — attempting to tackle training. DeepSeek's move lines up with that same logic: optimize for cost-per-query at scale rather than raw computational horsepower.

The Manufacturing Question

Reuters' reporting indicates talks are underway with manufacturing partners, but does not name the foundry or the chip manufacturing process that will be used. This matters because of US export controls on advanced semiconductor equipment. Any Chinese-based fabrication of a competitive inference chip would almost certainly be limited to older, less advanced manufacturing nodes than what Nvidia and TSMC currently ship. That constraint has direct consequences for power efficiency and chip yield — the percentage of usable chips produced in each manufacturing batch — regardless of how clever DeepSeek's chip design is.

Geopolitical Layer

The broader context here is that DeepSeek operates under heightened scrutiny from Washington following the June 2025 assertion that its chip acquisitions had supported Chinese military activities. A company designing its own AI accelerator while already under investigation for chip procurement is a different policy calculus for US regulators than a software-only model release would be. Export-control enforcement will almost certainly apply pressure at multiple points in the development cycle.

For the broader AI industry, the pattern is clear. Vertical integration — building and controlling both the model software and the silicon it runs on — has been the strategy of Google, Amazon, and increasingly OpenAI through hardware partnerships. DeepSeek following the same path suggests the economics of large-scale inference are compelling every major AI lab toward custom silicon, regardless of geography or political constraints. Whether DeepSeek can execute chip design with the same speed and cost efficiency it brought to model training is an open question that will take years to answer, as semiconductor manufacturing timelines are measured in years, not product cycles.