DeepSeek is designing its own AI accelerator, according to a Reuters exclusive published July 7 citing three people familiar with the matter, and Nvidia shares fell roughly 1.6% in premarket trading on the report. The Hangzhou lab has spent about a year in quiet talks with chip-design, foundry, and memory partners, and, per two of the sources, has been hiring silicon engineers off public job boards for months.

It's the logical next move for a lab whose entire brand is cost-per-token. Inference, as Reuters notes, runs on specialised chips that can be cheaper and less power-hungry than general-purpose GPUs, which is exactly where the economics of Chinese-market H800s and Huawei Ascend parts stop being kind to gross margins.

The context tightens the story. Liang Wenfeng, DeepSeek's founder, spent a rare 2024 interview describing a strategy of refusing external capital and naming US chip export controls as the binding constraint on Chinese AI. Two years later the lab is reportedly slated to raise $7 billion at a $52 billion to $59 billion valuation, per Reuters reporting from June. The posture has shifted from ascetic to industrial.

The broader field has arrived at the same conclusion in parallel. OpenAI unveiled its custom inference chip, Jalapeno, with Broadcom last month; Anthropic was reported in April to be deliberating on its own part. DeepSeek is the first Chinese lab to publicly cross that line, and it's doing so while V4, released in April, was adapted for Huawei's Ascend line, and Huawei confirmed its processors were used in part of the V4-Flash training run. Ascend 950 orders from Chinese cloud buyers surged after V4 shipped. The lab isn't running from Huawei so much as hedging against being defined by it.

The demand backdrop is real. Chinese-built models have accounted for more than 30% of weekly tokens routed through OpenRouter since February 8, up from an average of 4.5% across the first half of 2025. US developers are, quietly, routing around OpenAI and Anthropic on price.

Then there's the ceiling. "Nvidia is at zero in China and staying there. DeepSeek has almost no chance of selling silicon outside of China unless it gets access to leading edge manufacturing," Richard Windsor of Radio Free Mobile told Reuters. That's the whole geometry in two sentences. A domestic inference part solves DeepSeek's cost problem inside China's walls; it doesn't solve the foundry and HBM problem that keeps the walls where they're. DeepSeek didn't respond to requests for comment.

The 2024 interview posture, read against the 2026 chip effort and $7 billion round, is the tell. Refusing outside capital was a story about principles when the compute was there. It stopped being about principles the moment it wasn't.

Sources

https://www.usnews.com/news/top-news/articles/2026-07-07/exclusive-chinas-deepseek-developing-its-own-ai-chip-sources-say
https://www.bloomberg.com/news/articles/2026-07-07/chinese-ai-startup-deepseek-developing-own-ai-chip-reuters-says
https://www.cnbc.com/2026/07/07/chinese-ai-models-costs-us-openai-anthropic.html
https://www.engadget.com/2209378/deepseek-reportedly-developing-ai-chips/
https://finance.yahoo.com/technology/ai/articles/exclusive-chinas-deepseek-developing-own-103437335.html