On Thursday, February 26, 2026, explosive reports from Reuters and industry analysts revealed that the Chinese AI lab DeepSeek has broken standard industry protocol by withholding its upcoming flagship model, DeepSeek V4 (codenamed “MODEL1”), from U.S. chipmakers Nvidia and AMD.
Instead, DeepSeek has provided exclusive early access to Huawei Technologies, allowing the domestic giant to optimize its Ascend chips weeks before the model’s public release.
The “Optimization Blockade”
Typically, AI labs share pre-release models with Nvidia and AMD to ensure their software stacks (CUDA and ROCm) are tuned for peak performance on launch day. By excluding them, DeepSeek is intentionally creating a “performance gap”:
- Huawei’s Head Start: Huawei has had several weeks to tweak its drivers. On launch day, V4 is expected to run at maximum efficiency on Chinese hardware.
- Nvidia’s Disadvantage: Without lead time, V4 may run sluggishly or with bugs on Nvidia GPUs for several weeks or months until the U.S. firm can perform “reactive” optimizations.
- Strategic Intent: Analysts view this as a coordinated effort with the Chinese government to artificially disadvantage U.S. hardware within the Chinese market and bolster national “AI sovereignty.”
The “Blackwell” Controversy
The timing of this exclusion is linked to a massive geopolitical row. Earlier this week, a senior Trump administration official alleged that DeepSeek successfully trained V4 using a cluster of Nvidia Blackwell GPUs inside mainland China—a direct violation of U.S. export controls.
- The Allegation: DeepSeek reportedly used smuggled or redirected Blackwell chips located in Inner Mongolia.
- The “Hiding” Theory: Some U.S. officials believe DeepSeek is withholding the model from Nvidia to scrub technical indicators (like specific memory access patterns) that would prove the model was trained on banned American silicon.
- The Huawei Claim: DeepSeek reportedly plans to publicly claim that V4 was trained entirely on Huawei Ascend 910C clusters to avoid further U.S. sanctions.
What We Know About DeepSeek V4
Despite the “hiding” of the weights from chipmakers, technical leaks suggest V4 is a massive architectural leap:
| Feature | Specification (Leaked) |
| Architecture | 1-Trillion Parameter MoE (Mixture-of-Experts). |
| Context Window | Upgraded to 1 Million+ tokens (confirmed by a Feb 11 silent update). |
| Key Innovation | “Engram” System: Decouples factual recall from reasoning to save GPU memory. |
| Performance | Allegedly scores 90% on HumanEval, potentially beating GPT-5.2 and Claude 4.6 in coding. |
| Knowledge Cutoff | Updated to May 2025. |
Market & Geopolitical Fallout
- Nvidia’s Reaction: On Wednesday, Nvidia confirmed it has not generated any revenue from the limited, government-approved shipments of H200 chips to China, as the “DeepSeek controversy” has stalled final regulatory approvals.
- The “Distillation” Accusation: OpenAI and Anthropic have both recently accused DeepSeek of “distillation attacks,” claiming V4 was trained by “scraping” the reasoning outputs of their own flagship models to gain an unfair edge.
