DeepSeek’s inference chip project is the clearest sign yet that the global AI industry is moving away from a world where Nvidia supplies everyone. DeepSeek is designing a custom AI inference chip to cut reliance on Nvidia and Huawei, according to sources familiar with the matter, in what is described as an early-stage strategic shift. For Pakistani developers who already use DeepSeek’s low-cost API, and for companies watching Pakistan’s fast-growing data centre sector, this hardware move matters more than it might first appear.
What is an inference chip and why does it matter?
To understand this story, it helps to know the difference between training and inference. Training is the expensive, power-hungry process of teaching an AI model. Inference is what happens every time you ask an AI a question and it gives you an answer. It is the step that runs millions of times a day once a model is live.
The chip DeepSeek is developing is intended for inference rather than training. Inference is the stage where trained AI models generate responses for users, and it has become one of the fastest-growing areas of AI computing, with chatbots, coding assistants, enterprise tools, and autonomous agents handling millions of requests every day.
Inference workloads now account for the majority of AI chip demand as companies deploy models at scale. Custom inference chips are typically cheaper and more power-efficient than the general-purpose GPUs sold by Nvidia, making them attractive for companies seeking to compress costs. In simple terms, a chip built just for inference can do the same job at a lower cost and with less electricity than a general-purpose GPU.
DeepSeek’s inference chip plan, step by step
The effort began about a year ago and stays early-stage, with the Hangzhou company quietly hiring chip-design engineers through private channels and holding talks with design, foundry, and memory firms. No public job listings have been posted, and no timeline for tape-out or production has been disclosed.
The move comes weeks after DeepSeek closed a record $7.5 billion funding round, its first outside capital, at a valuation exceeding $50 billion, with backing from Tencent, CATL, and founder Liang Wenfeng himself. That war chest gives the company room to sustain a multi-year hardware effort even if the chip takes many years to reach production.
Since facing tighter Nvidia restrictions, DeepSeek has leaned more heavily on Huawei’s Ascend GPUs. In April it released its most advanced V4 model, which was designed to run on Huawei’s chips, leading to a surge of sales of the Ascend 950 to Chinese technology firms. Building its own chip would reduce that dependence on Huawei too, not just on Nvidia.
The bigger picture: every major AI firm is doing this
DeepSeek is not doing something unusual. It is following a clear pattern set by the world’s biggest tech companies.
- OpenAI and Anthropic have both disclosed custom chip programs, while Google has long run its TPU chips for both training and inference. Amazon designs its Trainium and Inferentia chips for AWS customers.
- OpenAI unveiled its first custom inference chip, Jalapeno, developed with Broadcom, last month, while Anthropic has reportedly been weighing development of its own AI silicon.
- Amid US restrictions on AI chip exports to China, Huawei has expanded its domestic market share and now controls around half of China’s AI chip market, which is valued at approximately $50 billion. However, that dominance is increasingly being challenged as several local technology companies, including Alibaba and Baidu, develop their own AI chips.
Industry surveys indicate Chinese firms plan to shift 46% of their AI-accelerator budgets to domestic suppliers within 12 months, up from roughly 30% today. DeepSeek’s chip project fits right inside that broader move toward domestic silicon.
What about Nvidia?
The news sent Nvidia shares down about 1.6% in premarket trading, reflecting investor attention to a broader trend reshaping the AI infrastructure market. However, analysts are not panicking about Nvidia just yet.
According to analyst Richard Windsor of Radio Free Mobile, the development of DeepSeek’s in-house inference chip will pose no serious threat to Nvidia, as the company will be unable to attract any major customers outside of China. US export controls that block Nvidia’s best chips from reaching China also limit DeepSeek’s ability to use the most advanced overseas factories to make or sell its own chip beyond Chinese borders.
While the market’s attention is on the technical pivot away from Nvidia and Huawei, the development points to a larger geopolitical reality: Western chip restrictions appear to be accelerating Chinese firms’ push toward technological independence rather than containing it. That is the real long-term story here.
The challenges DeepSeek still faces
Building a chip is very different from building a model. Developing a competitive AI processor demands years of engineering work, billions of dollars in investment, and access to manufacturing technology that remains heavily restricted for Chinese companies.
For a Chinese company, the challenge is compounded by US export controls that bar access to the most advanced overseas foundries and restrict shipments of high-bandwidth memory, a component critical to inference chip performance. So even with $7.5 billion in fresh funding, a finished, production-ready chip is likely still several years away.
Why Pakistani AI developers and cloud buyers should watch this
Pakistan’s tech sector has embraced DeepSeek’s API heavily since early 2025 because it offers powerful AI at a fraction of what US labs charge. The data centre industry is also growing fast, and cloud infrastructure decisions made today will shape costs for years. Here is why this hardware story matters locally.
If DeepSeek eventually runs its models on its own cheaper silicon, the company already cut prices on its V4-Pro API by 75% in May 2026, intensifying a price war among Chinese AI providers. Controlling its own silicon would give DeepSeek further leverage to undercut rivals on inference pricing. That means Pakistani startups and enterprises using the API could benefit from even lower costs in the future.
At the same time, the broader trend of every major AI company building custom chips means Nvidia’s dominance is slowly softening. For Pakistani cloud buyers and IT companies evaluating infrastructure, the AI hardware landscape in two to three years may look very different from today. Locking into any single vendor right now carries risk. Watching how this custom-chip trend unfolds, including DeepSeek’s effort, is a smart move for anyone planning AI workloads. Pakistan’s own IT export growth also depends on the competitiveness of the AI tools Pakistani firms use globally, so cheaper inference costs abroad translate directly into better margins at home.
Frequently Asked Questions
What exactly is DeepSeek building?
DeepSeek is building a custom chip designed specifically for AI inference, meaning the step where a trained AI model answers a user’s question. It is not building a training chip. The project is still in early stages, with no public timeline given.
Why does DeepSeek want to stop using Nvidia and Huawei chips?
Two reasons. First, US export controls block China from buying Nvidia’s most advanced GPUs, making supply uncertain. Second, building a custom chip lets a company cut running costs and control its own technology without depending on outside suppliers.
How long will it take for DeepSeek to release its own chip?
Nobody knows for certain, and DeepSeek has not said. But designing a competitive AI chip typically takes years of engineering work plus access to advanced chip factories. Most analysts expect a finished product is still several years away at minimum.
How does this affect Pakistani developers using DeepSeek’s API?
In the short term, nothing changes. But if DeepSeek successfully runs its models on cheaper in-house silicon, it could push API prices even lower than the 75% price cuts already seen in 2026. Pakistani startups using AI APIs would benefit from falling inference costs over the medium to long term.
