DeepSeek, the Chinese AI startup that upended the global AI race in early 2026, is now developing its own AI inference chip to reduce dependence on NVIDIA and Huawei hardware. Reported by Reuters on July 7–8, 2026, and confirmed by Bloomberg, the effort is still at an early stage — but DeepSeek is already in talks with manufacturing partners and quietly hiring chip engineers to staff the initiative.
What Reuters and Bloomberg Reported
According to Reuters, DeepSeek is developing its own chip specifically for inference — the process of running a trained AI model to generate responses — rather than for training new models from scratch. The distinction matters: inference chips can be more specialized, lower-power, and optimized for specific model architectures, making them a more tractable engineering target for a company at DeepSeek’s scale.
Multiple outlets confirmed the Reuters reporting. Per SiliconAngle’s coverage, DeepSeek has been exploring in-house accelerators for roughly a year but the initiative is described as early stage. No chip name has been announced, no manufacturing partner has been publicly identified, and no launch timeline has been given.
The chip’s stated goal, per sources cited by Reuters, is to reduce DeepSeek’s reliance on third-party chip providers. That means both NVIDIA — whose most advanced data center GPUs are blocked from China under US export controls — and Huawei, whose Ascend chips have become the de facto high-end AI hardware option for Chinese labs operating under those restrictions.
The Geopolitical Driver: Export Controls Are the Real Pressure
US export controls have steadily cut China’s access to the most advanced AI hardware. NVIDIA’s H100 and H200 chips — the industry standard for large-scale AI training and inference — are blocked from export to China. US curbs have also restricted access to high-bandwidth memory, a key component in AI accelerators. For a company running inference at DeepSeek’s scale, that’s a structural supply chain problem with no easy fix.
Building its own silicon is the long-term answer. If DeepSeek controls its inference chip, it sidesteps the sanctioned supply chain entirely. A domestically designed and manufactured chip — assuming Chinese foundry access can handle the required specs — is not subject to US export restrictions in the same way.
US chip designers working with Chinese customers also face constraints: per SeekingAlpha’s coverage of the Reuters story, US bans prevent Chinese chip designers from accessing the most advanced overseas foundries like TSMC’s leading-edge nodes. That means DeepSeek’s chip will likely be manufactured at a less advanced process node, which affects power efficiency and density. But for inference workloads — which are less computationally intensive per chip than training — a chip built on an older node can still be competitive for cost and volume.
DeepSeek Isn’t Alone — OpenAI and Anthropic Went First
DeepSeek is following a path already taken by its American counterparts. OpenAI recently unveiled its own custom inference chip, codenamed Jalapeno, built in partnership with Broadcom. Anthropic has also been reported to be developing custom silicon. The pattern is clear: every major AI lab eventually concludes that commodity GPU clusters are too expensive, too dependent on a single supplier (NVIDIA), and too generic for the specific inference workloads their models generate.
Designing a competitive AI chip typically takes years and significant capital — industry context puts the timeline at three to five years from tape-out to volume production for a new entrant. DeepSeek’s one-year-old exploration is genuinely early. But the direction of travel is set: the inference chip market, which is where the volume of AI compute lives day-to-day, is becoming a multi-competitor race.
What This Means for the AI Hardware Market
NVIDIA’s dominance in AI has rested on two pillars: the H-series chips for training, and an increasingly large share of inference infrastructure. The training market is harder to crack — it requires massive parallel bandwidth and memory that few companies can match. But inference is different. It’s the workload that runs 24/7 at every AI product company. Specialized inference silicon — from Groq, Cerebras, and now potentially DeepSeek — targets exactly this market.
DeepSeek’s models are already widely used. Earlier in 2026, Chinese AI models accounted for more than 30% of weekly token usage by US companies, peaking at 46%, according to CNBC and Bloomberg reporting. If DeepSeek’s inference chip ships and performs competitively, it could change the cost structure of that workload — and give DeepSeek a hardware moat that’s entirely outside the US-controlled supply chain.
Frequently Asked Questions
What chip is DeepSeek building?
DeepSeek is developing a proprietary AI inference chip — designed for running trained models rather than training new ones. As of Reuters’ reporting on July 7–8, 2026, no name, manufacturing partner, or launch timeline has been announced. The initiative has been in exploration for roughly a year and is described as early stage.
Why is DeepSeek building its own chip instead of using NVIDIA?
US export controls block China from accessing NVIDIA’s most advanced data center chips (H100, H200) and high-bandwidth memory. DeepSeek’s chip push is a strategic response to that supply chain constraint — a domestically designed inference chip would not be subject to US export restrictions in the same way. Reducing reliance on Huawei’s Ascend chips is also a stated goal.
What is an AI inference chip?
An AI inference chip is specialized hardware designed to run a trained AI model — generating responses, classifications, or outputs — rather than to train models from scratch. Training requires massive clusters of high-bandwidth GPUs. Inference chips can be more power-efficient and purpose-built for specific model architectures, making them a more tractable target for new entrants to the chip market.
Is DeepSeek’s chip a threat to NVIDIA?
Not immediately — competitive AI chip design typically takes three to five years from early development to volume production. But the trend matters: every major AI lab (OpenAI with Jalapeno, Anthropic, now DeepSeek) is developing custom silicon for the inference market, which is where the daily AI compute volume lives. NVIDIA’s dominance in inference is the segment most at risk from specialized competitors.
What manufacturing constraints does DeepSeek face?
US bans prevent Chinese chip designers from accessing the most advanced overseas foundries, including TSMC’s leading-edge nodes, per SeekingAlpha’s reporting. DeepSeek’s chip will likely need to be manufactured at a less advanced process node, which affects power efficiency and density. For inference workloads, which are less parallel-compute intensive than training, a chip on an older node can still be viable.
Sources: SiliconAngle · SeekingAlpha · TechNode (citing Reuters/Bloomberg, July 7–8, 2026)
