DeepSeek Locks Nvidia and AMD Out of V4 - Gives Huawei a Head Start

DeepSeek has broken from long-standing industry convention by denying Nvidia and AMD pre-release access to its upcoming V4 flagship model, instead granting Huawei and other Chinese chipmakers a multi-week head start to optimize their hardware for the new system.

The move, first reported by Reuters on February 25 and confirmed by multiple sources since, signals something larger than a vendor preference. It is the first time a major Chinese AI lab has deliberately locked American chipmakers out of its pre-release pipeline - a pipeline that, until now, worked the same way everywhere in the industry.

How the Pre-Release Pipeline Works

When an AI lab builds a new model, it typically hands pre-release versions to major chipmakers weeks before launch. Nvidia, AMD, and others use that window to tune their drivers, CUDA kernels, and inference stacks so the model runs well on their hardware from day one. It's routine, unglamorous plumbing work - and it matters enormously for adoption.

DeepSeek just cut Nvidia and AMD out of that process completely.

The battle over AI chips has shifted from hardware access to software optimization rights.

Instead, Huawei's Ascend chip division and other domestic Chinese suppliers received the V4 preview. That gives them weeks to optimize their silicon for what's expected to be a trillion-parameter model with a one-million-token context window - architectural features from DeepSeek's V3.2 lineage that demand tight hardware-software integration to run efficiently.

The Lockout by the Numbers

"The impact to Nvidia and AMD for general data accelerators is minimal - most enterprises are not running DeepSeek. It is a benchmarking model."

• Ben Bajarin, CEO of Creative Strategies

Bajarin added that the move reflects a broader Chinese government strategy "to try to keep US hardware and models disadvantaged" inside China.

Why This Is Different

AI labs in the US and China have always shared pre-release models with chipmakers regardless of nationality. Google gives Nvidia and AMD early access to Gemini optimizations. Meta does the same with Llama. Even DeepSeek's previous V3.2 release followed this convention.

The AI chip war has evolved from hardware export controls to software ecosystem strategy.

The lockout isn't about keeping V4 secret - the model will be open-sourced under Apache 2.0, and anyone will be able to download and run it. It's about who gets a head start on making it run well. If Huawei's Ascend chips are optimized for V4 at launch while Nvidia's CUDA stack isn't, enterprise buyers in China have a practical reason to choose domestic silicon - even if Nvidia's raw hardware is technically superior.

The Software Ecosystem Play

This is the detail that matters most. Hardware advantage means little if the software doesn't keep up. By giving Huawei early optimization access, DeepSeek is helping build something the Chinese AI industry has lacked: a competitive software stack around domestic chips. Every developer who tunes their workflow to run open-source models on Ascend chips becomes a node in a network that gets harder to displace.

The Irony of the H200

The timing makes the strategic calculus even sharper. Nvidia just secured a US government license to export a limited number of H200 chips to China - a less advanced processor than the Blackwell units DeepSeek reportedly used for training. But as of late February 2026, zero H200 units have actually shipped to Chinese customers. Nvidia did not include any China data center revenue in its Q1 sales outlook.

So the picture is this: the US approved chip sales that haven't happened, while DeepSeek locked American chipmakers out of the software access that would make those chips useful for V4 when they arrive.

China's data center infrastructure is increasingly designed to run on domestically optimized hardware and software stacks.

Counter-Argument

The lockout may be less significant than it appears. Most enterprise AI deployments outside China still run on Nvidia hardware, and open-source models tend to get community-driven optimizations for CUDA within days of release regardless of whether the lab cooperates. Nvidia's ecosystem of 4.7 million developers, extensive tooling, and battle-tested inference libraries won't collapse because one lab denied pre-release access.

Bajarin's framing - that DeepSeek is "a benchmarking model" rather than an enterprise workhorse - also matters. Chinese models dominate Hugging Face downloads, but downloads do not equal production deployments. Western enterprises building on frontier models from Anthropic, OpenAI, or Google aren't about to swap their stack for DeepSeek V4 regardless of chipmaker access.

The real question is whether this is a one-off or a pattern. If Alibaba's Qwen team or Baidu follows DeepSeek's lead and starts locking US chipmakers out of pre-release access too, the collective effect on Nvidia's China strategy becomes harder to dismiss.

What the Market Is Missing

The chip war isn't about chips anymore. Washington spent three years trying to deny China access to advanced semiconductors, and the result is a Chinese AI lab that trained its flagship model on smuggled Nvidia hardware and then locked Nvidia out of the software that runs on top of it. The export controls created exactly the incentive structure you would expect: if you cannot buy the hardware legally, build the software ecosystem that makes the hardware replaceable.

DeepSeek is not just building a model. It's building a reason for Chinese enterprises to choose Huawei over Nvidia - not because Huawei's chips are better, but because Huawei's chips will run the model better on launch day. That's a moat that doesn't show up on Nvidia's earnings call, but it is the kind of moat that matters over five years.

The 75 million downloads on Hugging Face are the leading indicator. The software ecosystem around Chinese silicon is the lagging one. DeepSeek just connected the two.

Sources:

• Exclusive: DeepSeek Withholds Latest AI Model From US Chipmakers Including Nvidia - Reuters via Yahoo Finance
• DeepSeek Withholds V4 Model from US Chipmakers, Grants Huawei Exclusive Early Access - The China Academy
• DeepSeek Gives Huawei a Head Start on V4 Model, Blocks US Chip Giants Out - News9live
• DeepSeek Shuts Out Nvidia, AMD in V4 Move Tied to US Chip Tensions - Digitimes
• Nvidia Prepares Shipment of 82,000 AI GPUs to China as Chip War Lines Blur - Tom's Hardware
• DeepSeek V4 Targets Coding Dominance - Introl Blog