DeepSeek Develops In-House AI Inference Chips to Reduce Reliance on Nvidia and Huawei
Alina Collins
DeepSeek is building its own AI inference chip, Reuters reports, aiming to cut its dependence on Nvidia and Huawei; the move marks a high-stakes pivot from pure software into silicon for the company known for doing more with less.
Why is DeepSeek making its own chip?
DeepSeek currently relies on Nvidia H800 GPUs for training and Huawei Ascend chips for deployment — both supply chains carry external risk.
The H800 was banned from export to China in late 2023; Huawei Ascend is available but still constrained in capacity and performance.
This means → whichever supplier DeepSeek picks, the risk of a cut-off or a bottleneck remains. Building in-house is an attempt to take back control.
Why an inference chip, not a training chip?
Inference — the stage where a trained model generates responses for users — requires chips that are typically cheaper and less power-hungry than training chips.
Inference is also the fastest-growing segment of AI compute demand: more users means more inference load.
In plain terms = training chips are the power plant; inference chips are the local transformer. DeepSeek is entering at the easier, bigger end of the market first.
How far along is the project?
The chip effort started roughly one year ago and remains in an early stage.
DeepSeek has begun engaging chip-design, foundry, and memory partners, while quietly expanding its chip-design engineering team through private hiring — no public job postings.
This reflects a project that has not yet reached tape-out, but the company is making real moves in staffing and supply-chain outreach — not a paper exercise.
How much pressure does this put on Huawei?
Huawei holds roughly half of China's approximately $50 billion AI chip market; DeepSeek is a significant customer.
In April, DeepSeek released its V4 model optimized for Huawei Ascend; Huawei said its processors helped train the lighter V4-Flash variant.
Yet Alibaba, Baidu, and other giants are already building their own chips. If DeepSeek joins them, Huawei's addressable market shrinks further.
This means → Huawei Ascend's moat is not the technology itself but the fact that customers have few alternatives. Each new in-house effort makes that moat shallower.
What are the biggest obstacles to building this chip?
Design: a competitive AI chip typically takes years and heavy capital to develop.
Manufacturing: U.S. export controls bar Chinese designers from the most advanced overseas foundries; high-bandwidth memory (HBM — a key component in AI chips) is also restricted.
In plain terms = even if DeepSeek finishes the design, finding a fab to build it and sourcing the critical components are each a chokepoint.
Are other AI companies on the same path?
OpenAI last month unveiled "Jalapeno," its first in-house inference chip co-developed with Broadcom; Anthropic is also reportedly evaluating a custom-chip effort.
This reflects a growing consensus among top AI companies: single-supplier dependence on Nvidia is a strategic risk, and in-house silicon is the hedge.
DeepSeek founder Liang Wenfeng said in a rare 2024 interview that chip export controls are a major challenge for the company.
Content is for reference only, not financial advice.