# DeepSeek Building Its Own AI Chip

> Source: [https://icharles.com/articles/deepseek-proprietary-ai-chip-development](https://icharles.com/articles/deepseek-proprietary-ai-chip-development) (canonical)
> Author: iCharles News — iCharles, https://icharles.com
> Published: 2026-07-07

## TL;DR

DeepSeek, the Chinese AI startup behind a widely used open-source large language model, is developing its own AI chip, according to three people familiar with the matter. The company has launched a major recruitment drive for semiconductor design talent. This comes as a Chinese official confirmed DeepSeek trained its latest model on Nvidia's H800 GPUs — chips that were already under US export restrictions at the time.

## What is DeepSeek building?

**DeepSeek** is a Chinese AI startup known for developing open-source large language models. According to [Reuters reporting via StreetInsider](https://www.streetinsider.com/Reuters/Exclusive-Chinas+DeepSeek+developing+its+own+AI+chip,+sources+say/26738057.html), three people familiar with the matter say the company is now developing its own proprietary AI chip.

DeepSeek has launched a major recruitment drive for semiconductor design talent. That hiring push is the clearest signal yet of its in-house chip ambitions, per a DigiTimes Asia report cited by [TechPowerUp](https://www.techpowerup.com/332565/deepseek-reportedly-pursuing-development-of-proprietary-ai-chip).

## What chips has DeepSeek used so far?

DeepSeek has relied heavily on Nvidia hardware. Industry watchers believe the company has access to 10,000 sanction-approved Nvidia "Hopper" H800 AI chips, as well as 10,000 H100 AI GPUs — the latter of which are now banned from export to China.

A Chinese official confirmed to Reuters that DeepSeek trained its AI model on Nvidia's best available chip despite the US ban. That confirmation adds weight to earlier speculation about DeepSeek's hardware stack.

Scale AI CEO Alexandr Wang separately claimed, around late January, that DeepSeek could use up to 50,000 H100 chips for model training. That claim was unsubstantiated and raised questions given current US-China trade tensions.

## What do we know about the chip design itself?

Here's what we know so far: not much. The DigiTimes Asia report did not provide detailed insight into the rumored in-house chip design. No specifications, timeline, or manufacturing partner were disclosed.

TechPowerUp notes that DeepSeek faces a real structural challenge. The Chinese semiconductor industry trails leading regions in process node technology. Whether local foundries can supply an advanced enough manufacturing process for AI chip production remains an open question the sources raise but do not answer.

## Why is DeepSeek recruiting semiconductor talent?

The recruitment drive signals that DeepSeek wants to reduce its dependence on foreign chips. Several large organizations have already moved toward proprietary AI hardware — DeepSeek appears to be following that path.

Reports have also noted that DeepSeek has been running inference workloads on Huawei Ascend 910C AI GPUs, suggesting the company is already exploring alternatives to Nvidia. The in-house chip effort would go further, toward full design ownership.

This context matters for builders and founders watching the [US export controls AI chips](https://www.reuters.com/world/china/chinas-deepseek-trained-ai-model-nvidias-best-chip-despite-us-ban-official-says-2026-02-24/) story. The pressure of those restrictions appears to be accelerating DeepSeek's push toward hardware independence — a pattern also visible in how other AI labs are thinking about [OpenAI inference cost](/articles/openai-inference-cost-halved-optimization) and compute efficiency.

## Who confirmed DeepSeek's chip training details?

A Chinese official confirmed to Reuters on February 24, 2026 that DeepSeek trained its AI model on Nvidia's best chip despite the US export ban. That statement is the most direct official acknowledgment of DeepSeek's hardware situation to date.

The [Business Post](https://www.businesspost.ie/article/deepseek-is-reportedly-developing-its-own-ai-chip/) reported the chip development story citing three people familiar with the matter — consistent with the Reuters sourcing.

## Key facts at a glance

| Detail | What the sources report |
|---|---|
| Chip development status | In progress, per three sources |
| Recruitment | Major semiconductor design talent drive launched |
| H800 chips (approved) | ~10,000 reportedly accessible |
| H100 chips (now banned) | ~10,000 reportedly accessible |
| H100 claim (Wang) | Up to 50,000 for training (unsubstantiated) |
| Training chip confirmed | Nvidia H800, per Chinese official |
| Chip design details disclosed | None |

## What does this mean for the broader AI hardware race?

DeepSeek's move mirrors what other major AI organizations have done — building custom silicon to control cost and performance. The difference is the export control backdrop. US restrictions on advanced chips to China are a direct pressure on DeepSeek's supply chain.

For developers and founders tracking AI infrastructure, this story connects to wider debates about compute access and model efficiency. Labs like Anthropic are also navigating hardware and policy pressures — see [Anthropic's clash with Alibaba](/articles/anthropic-alibaba-claude-distillation-attack) over AI model use as another example of the competitive tensions shaping this space. Meanwhile, questions about how AI labs manage training costs are central to stories like [Claude's user growth](/articles/claude-paid-users-growth-2026) and the economics behind it.

The most confirmed fact on record: a Chinese official stated on February 24, 2026 that DeepSeek trained its model on Nvidia's H800 GPU despite the US ban — and the company is now recruiting to build its own chip so it doesn't have to rely on that supply chain going forward.

## Frequently asked questions

**Is DeepSeek developing its own AI chip?**

Yes. According to three people familiar with the matter, as reported by Reuters and Business Post, DeepSeek is developing a proprietary AI chip. The company has launched a major recruitment drive for semiconductor design talent. No specifications or timeline for the chip have been disclosed publicly, and no manufacturing partner has been named in the reports.

**What Nvidia chips does DeepSeek currently use?**

Industry watchers believe DeepSeek has access to approximately 10,000 sanction-approved Nvidia H800 chips and 10,000 H100 GPUs, which are now banned from export to China. A Chinese official confirmed to Reuters that DeepSeek trained its AI model on Nvidia's best available chip despite the US export restrictions in place at the time.

**Why is DeepSeek building its own chip?**

The sources do not state DeepSeek's official reasoning. However, the company's recruitment drive for semiconductor talent signals a push toward hardware independence. US export controls have restricted DeepSeek's access to advanced Nvidia GPUs, and the in-house chip effort appears to follow a broader pattern of Chinese AI firms seeking to reduce reliance on foreign hardware.

**How many H100 chips does DeepSeek have access to?**

Industry watchers believe DeepSeek has access to around 10,000 H100 AI GPUs, which are now banned from export to China. Scale AI CEO Alexandr Wang separately claimed, around late January, that DeepSeek could use up to 50,000 H100 chips for model training — a claim described by TechPowerUp as unsubstantiated and one that raised eyebrows given current geopolitical tensions.

**What challenge does DeepSeek face in building its own AI chip?**

TechPowerUp notes that the Chinese semiconductor industry trails leading regions in advanced process node technology. Whether local foundries can manufacture chips at the node required for competitive AI workloads is an open question. The DigiTimes Asia report that first surfaced the recruitment news provided no detail on which foundry DeepSeek might use or what process node it is targeting.