By blocking Nvidia and AMD from early access to its V4 model, DeepSeek isn't being petulant — it's systematically helping Chinese chipmakers catch up. This is industrial policy dressed as corporate decision-making.
DeepSeek, the Chinese AI lab that sent shockwaves through the global semiconductor market with its efficient R1 model last year, is doing it again — not with a model release, but with a strategic denial. According to multiple reports, DeepSeek has withheld early access to its upcoming V4 model from Nvidia and AMD, giving Huawei and other Chinese chipmakers a head start of several weeks to optimise their software for the new architecture. The model is expected as early as March 2026. The chip access decision, reported by Reuters and confirmed by industry sources, deserves more attention than the model specs do.
What Happened
DeepSeek's upcoming V4 model has been anticipated since late 2025, initially expected around the Lunar New Year holiday in late January 2026. The release has slipped, but preparations are clearly under way. What's notable about those preparations is who's been excluded.
Major AI chip companies — including Nvidia, AMD, and their software teams — typically receive early access to new frontier models so they can optimise their CUDA libraries, ROCm stacks, and hardware drivers for new model architectures before public release. This benefits everyone: the model runs faster on more hardware from day one, and chipmakers can market their hardware as compatible. DeepSeek has broken this convention.
Instead, DeepSeek gave Chinese chipmakers — including Huawei with its Ascend line — an early access window. According to sources cited by Reuters, this lead time has allowed Huawei's software teams to prepare optimised drivers and inference kernels specifically for V4, meaning that Huawei chips may have better out-of-the-box performance with V4 than Nvidia's at launch. The FT additionally reported that DeepSeek has been working directly with Huawei to reduce reliance on Nvidia infrastructure. AMD's MI308, a chip that had reportedly achieved $390 million in sales partly from Chinese AI company demand, is also locked out.
Why It Matters
The conventional analysis of China's AI position focuses on model capability benchmarks and export controls on Nvidia's H100 and H20 chips. This story reveals a different and arguably more durable dynamic: deliberate ecosystem bootstrapping.
DeepSeek's decision to exclude US chipmakers from early access is not a product decision — it's industrial policy. By guaranteeing Chinese chipmakers an optimisation head start, DeepSeek is accelerating the feedback loop that China needs to develop viable alternatives to Nvidia. When a major frontier model launches performing well on Huawei Ascend chips, it validates those chips to other Chinese AI companies. They evaluate, adopt, build for, and request improvements to the hardware. The ecosystem thickens. Nvidia's position in the Chinese market — already hobbled by export controls — weakens further.
This matters beyond China's domestic market. Export controls on advanced Nvidia chips were designed to slow China's AI development. DeepSeek's V3 model, released late last year, already demonstrated that Chinese labs could achieve competitive performance at dramatically lower compute costs than US counterparts — a finding that destabilised AI chip valuations globally. V4 may extend that lead. But the more durable implication is that even if Nvidia chips remain technically superior, China may be building a parallel AI hardware ecosystem that doesn't need them.
For Nvidia specifically, losing optimisation partnerships with leading Chinese AI labs is a meaningful long-term threat. The company reported $193.7 billion in annual data centre revenue, up 75% year on year — but China was already a constrained market. If Chinese labs build their workflows around Huawei chips optimised for Chinese models, re-entry becomes increasingly difficult even if export controls are eventually relaxed.
Wider Context
China's approach to AI hardware independence has been under development for years, but the pace has accelerated dramatically since the US tightened export controls in 2022 and again in 2023. Huawei's Ascend series — particularly the 910B and 910C — has been adopted by Chinese AI labs including Baidu, Alibaba, and ByteDance to varying degrees, though performance gaps versus Nvidia's H100 have remained significant for training large models.
DeepSeek's efficiency innovations — the mixture-of-experts architectures and aggressive quantisation approaches that made R1 competitive at low cost — may have the secondary effect of making the gap between Huawei and Nvidia hardware less consequential. If you can achieve competitive performance with fewer FLOPS, the absolute performance of any given chip matters less. DeepSeek may be inadvertently (or deliberately) reducing the relevance of hardware benchmarks that currently favour Nvidia.
Meanwhile, the geopolitical context is not static. US export controls have already proven difficult to enforce across the full supply chain, with reports of smuggling, third-country routing, and grey-market access to restricted chips. China's strategy of reducing dependency on Nvidia is therefore a hedge against current controls and any future tightening.
The Singularity Soup Take
The chip access story is easy to miss in the noise about DeepSeek V4's expected benchmarks, but it's arguably the more important development. DeepSeek is not just building competitive AI models — it's actively shaping the hardware ecosystem those models run on. That's a sophisticated industrial strategy, and it's working.
Western analysis of China's AI ambitions has too often focused on the wrong metrics: model parameter counts, benchmark scores, and chip export numbers. The ecosystem story is harder to quantify but more consequential. If Chinese AI labs normalise building and optimising for Chinese chips, Nvidia's long-term position in that market erodes regardless of absolute performance. The flywheel spins in one direction.
The US response has been reactive — restricting chip exports without building a positive alternative. That approach can slow the timeline but not the direction. DeepSeek's V4 release, when it comes, should be watched not just for what it can do, but for what it can do on Huawei hardware.
What to Watch
The V4 launch — currently expected around early March 2026 — will be the test case. Benchmark comparisons specifically on Huawei Ascend hardware versus Nvidia will be telling. Watch also for whether other Chinese AI labs (Baidu, Zhipu, Moonshot) follow DeepSeek's early-access convention when launching their own models. If the pattern becomes industry-standard rather than a DeepSeek-specific decision, it represents a systemic shift in how Chinese AI development relates to global chip markets. US Commerce Department responses — in the form of tightened or re-scoped export controls — are also likely.
Sources
PC Gamer — "DeepSeek has reportedly denied Nvidia and AMD early access to its new V4 AI model"
Futurism — "American AI Industry Trembles as DeepSeek Prepares to Release New Model"
Manila Times — "DeepSeek withholds latest AI model from US chipmakers"
Wikipedia — "DeepSeek" (for timeline and Nvidia chip training context)