Key Takeaways
China's DeepSeek Quietly Building Its Own AI Chip to Ditch Nvidia and Huawei Control
- DeepSeek is designing a custom chip focused on inference, not training
- The company is raising $7 billion at a $52-59 billion valuation
- US export controls make this effort harder but also more urgent
DeepSeek, the Chinese AI lab that shocked the industry with its low-cost training methods, is now designing its own AI chip. Reuters reports the chip targets inference workloads, the compute-intensive phase where a trained model actually responds to users. Three people familiar with the project confirmed the effort is underway.
The move signals a strategic shift. DeepSeek built its reputation on squeezing remarkable performance from limited hardware. Now it wants to control that hardware directly.
Why inference, not training?
Training a model is a one-time cost. Inference happens every time a user sends a prompt. For a company scaling its API and consumer products, inference costs compound relentlessly. Every query, forever.
DeepSeek reportedly trained its V3 model for around $5.6 million, a fraction of what competitors spend. But serving that model to millions of users? That bill keeps growing. A custom inference chip, optimized for DeepSeek's specific architecture, could slash those ongoing costs.
The project remains early-stage. DeepSeek is in discussions with chip design firms, manufacturers, and memory companies. The company has been quietly recruiting chip engineers for months, avoiding public job postings.
Export controls make this harder and more urgent
US export restrictions block Chinese companies from buying the most advanced chips and chipmaking equipment. That constraint forced DeepSeek to get creative with older Nvidia hardware. It also makes the case for domestic chip development stronger.
Right now, DeepSeek relies on some combination of Nvidia and Huawei chips. Nvidia dominates the global AI accelerator market with roughly 80-90% share. But each new round of US restrictions tightens what Chinese firms can access. Building in-house reduces that dependency.
The path is not simple. Chip design requires deep expertise. Manufacturing at advanced nodes requires access to equipment from ASML and others, which export controls limit. Memory bandwidth matters enormously for inference workloads, and the best high-bandwidth memory comes from Samsung and SK Hynix in South Korea.
DeepSeek is also raising serious money
Reuters reports DeepSeek is seeking $7 billion in outside capital at a valuation between $52 billion and $59 billion. This would be the company's first external fundraise.
That valuation puts DeepSeek in the same tier as Anthropic and not far behind OpenAI. The funding would finance chip development, model training, and global expansion. Chip projects burn cash for years before yielding silicon.
DeepSeek joins a crowded field
OpenAI and Anthropic are both working on custom chips. Google has TPUs. Amazon has Trainium and Inferentia. Microsoft has Maia. Meta is developing its own accelerators. The logic is consistent: at scale, owning your silicon means lower costs, better optimization, and less dependence on Nvidia's pricing power and allocation decisions.
For DeepSeek, the calculus includes geopolitical risk. Even if Nvidia wanted to sell its best chips to Chinese AI labs, US regulations prevent it. Self-sufficiency is not just economically attractive; it is strategically necessary.
Logicity's Take
DeepSeek's chip ambitions make sense given its constraint-driven innovation culture. The company already proved it can match frontier models on a shoestring training budget. If it can build inference silicon tuned specifically to its architecture, it could widen the cost gap with Western competitors. The harder question is execution. Chip programs take 3-5 years to reach production. Export controls complicate access to advanced manufacturing and memory. DeepSeek is betting it can solve hardware problems the same way it solved software ones: through unconventional optimization. For AI teams watching from the outside, the broader lesson is that vertical integration is becoming table stakes for frontier labs.
What this means for the chip market
Nvidia's dominance remains intact, but the trend lines point toward fragmentation. Every major AI lab is now either building chips or seriously evaluating it. That does not mean Nvidia loses, at least not soon. Custom chips take years. Nvidia's CUDA ecosystem is deeply entrenched. But the long-term trajectory is clear: the biggest AI players want to own their stack, top to bottom.
For smaller AI companies and startups, the immediate impact is minimal. They will keep buying Nvidia GPUs or renting them from cloud providers. But the pricing leverage Nvidia enjoys today may erode as alternatives mature.
Frequently Asked Questions
What kind of chip is DeepSeek building?
An inference chip, designed to run trained models efficiently rather than to train new ones. Inference workloads dominate operational costs at scale.
How much is DeepSeek raising?
The company is seeking $7 billion at a valuation between $52 billion and $59 billion, according to Reuters.
Why can't DeepSeek just buy Nvidia chips?
US export controls restrict Chinese companies from accessing the most advanced Nvidia and AMD chips, forcing DeepSeek to rely on older hardware or domestic alternatives.
Are other AI companies building their own chips?
Yes. OpenAI, Anthropic, Google, Amazon, Microsoft, and Meta all have custom chip programs at various stages.
How the leading Western AI labs are competing for developer mindshare with compute credits.
Need Help Implementing This?
Building AI products and want to understand infrastructure tradeoffs? Reach out to our team at Logicity for strategic guidance on chip economics, cloud compute, and model deployment decisions.
Source: The Decoder / Maximilian Schreiner
Manaal Khan
Tech & Innovation Writer
Produced with AI assistance and reviewed by the Logicity editorial team. Learn more in our Editorial Policy.
