The AI cold war just got a lot colder.

DeepSeek — the Chinese lab that shook the industry last year with models matching Western competitors at a fraction of the cost — has cut Nvidia and AMD out of its upcoming V4 launch. No early access. No optimization window. Instead, Huawei got a weeks-long head start to tune V4 for its own Ascend processors.

This isn’t a technical decision. It’s a geopolitical signal flare.

How Model Launches Usually Work

Before a major AI model drops, developers share pre-release versions with chipmakers like Nvidia. Hardware teams optimize their software stacks — drivers, libraries, compilers — so the model runs efficiently on their GPUs. It’s been standard practice for years. DeepSeek played along before, working closely with Nvidia’s technical staff on previous releases.

For V4, they broke the pattern entirely.

Chinese chipmakers got early access. American ones got nothing. When V4 launches, it’ll run better on Huawei’s hardware out of the box. That’s never happened before with a model this anticipated.

The Blackwell Bombshell Makes It Worse

The withholding story alone would be significant. But Reuters dropped a companion revelation that makes it explosive: a senior Trump administration official claims DeepSeek trained V4 using Nvidia’s Blackwell chips — the company’s most advanced AI processor — at a data center in Inner Mongolia.

Problem: US export controls bar Blackwell shipments to China. Full stop.

Even more provocatively, US officials allege DeepSeek may be planning to scrub technical indicators revealing its use of American chips and publicly claim V4 was trained entirely on Huawei hardware.

“Given China’s leading AI companies are brazenly violating U.S. export controls, we obviously cannot expect that they will comply with U.S. conditions,” one official told Reuters.

This is the nightmare scenario export control hawks have warned about: restricted chips finding their way into Chinese hands through grey markets — and the resulting models being optimized away from American hardware after the fact.

What Makes V4 So Dangerous

DeepSeek isn’t just any Chinese AI company. Its models have been downloaded more than 75 million times on Hugging Face. Among models released in the past year, Chinese downloads have surpassed those from any other country on the platform.

V4 looks like their most ambitious release yet:

  • Engram architecture — a “cyborg” approach separating static memory (CPU RAM) from dynamic reasoning (GPU compute), reportedly cutting VRAM requirements by 30%
  • Massive context windows — built with DeepSeek Sparse Attention for repo-level context, slashing long-context costs by roughly 50%
  • Aggressive pricing — leaks suggest around $0.27 per million tokens, approximately 40x cheaper than premium US models
  • Benchmark rumors — unverified claims suggest V4 outperforms Claude 4.5 and GPT-5.2 on real-world coding tasks

A “V4 Lite” variant has already leaked through unofficial channels. Markets are nervous enough that analysts predict the Nasdaq could take a hit when V4 officially drops.

The Ecosystem Lock-In Play

Ben Bajarin, CEO of Creative Strategies, offered a measured take: the immediate impact to Nvidia and AMD is minimal since most enterprises aren’t running DeepSeek. But he identified the larger pattern — this fits a broader Chinese government strategy “to try to keep U.S. hardware and models disadvantaged” within China.

That’s the real game. If China’s most influential AI lab optimizes for Huawei first, other Chinese developers follow. Enterprise customers buy Huawei. The flywheel spins. American chips become second-class citizens in the world’s second-largest AI market.

Meanwhile, AMD’s MI308 generated $390 million in China sales last quarter. That revenue is now at risk if China’s AI ecosystem deliberately routes around American hardware.

What This Means for Everyone

For Nvidia: Limited financial damage — most revenue comes from Western hyperscalers. But if the world’s most-discussed AI lab runs better on Huawei than Nvidia, that’s a narrative problem money can’t fix.

For Huawei: A massive win. Huawei’s Ascend chips have struggled with stability and slower inter-chip connectivity. Having DeepSeek optimize V4 for Ascend first gives them a showcase customer that could accelerate their entire AI hardware roadmap.

For US export controls: Deeply embarrassing. If China’s top AI lab can train on America’s most restricted chips while simultaneously cutting American companies out, the entire enforcement regime looks like it’s failing at both ends.

For developers: V4 will eventually run well on Nvidia regardless — optimization tools are improving rapidly. But the performance gap at launch could matter for benchmarks and enterprise decisions.

The Uncomfortable Truth

The export control allegations will intensify congressional scrutiny. Expect hearings and calls for stricter enforcement. The AI chip trade is about to become an even hotter flashpoint in US-China relations.

But here’s what neither side wants to acknowledge: the genie is out of the bottle. DeepSeek has already proven world-class AI can be built at a fraction of Western costs. Whether V4 was trained on Blackwell or Ascend chips matters legally and politically, but the model will exist regardless. And once it’s open-source, it runs everywhere.

The question isn’t whether the AI world is splitting into two camps. It’s whether anyone can stop it.

Sources: Reuters, Reuters, Kilo AI, Dataconomy