DeepSeek is expected to release V4 this week - a trillion-parameter multimodal model timed to coincide with China’s annual “Two Sessions” parliamentary meetings beginning March 4. The release signals something more significant than a new AI model: it’s China’s most deliberate attempt yet to build an AI ecosystem independent of American hardware.
V4 isn’t just optimized for domestic Chinese chips. DeepSeek has blocked Nvidia and AMD from pre-release access entirely, giving Huawei’s Ascend division exclusive early optimization. In a field where American companies have dominated, DeepSeek is building walls.
What V4 Actually Is
According to leaked specifications, V4 is a mixture-of-experts model with approximately 1 trillion total parameters, activating around 32 billion per token. That architecture allows frontier-level capability without crippling hardware requirements - similar to DeepSeek’s V3, which shocked the industry by matching Western models at a fraction of the cost.
The key specifications:
| Spec | V4 |
|---|---|
| Total parameters | ~1 trillion |
| Active parameters | ~32B per token |
| Context window | 1 million tokens |
| Multimodal | Native (text, image, video, audio) |
| Training approach | MoE with 16 active experts |
V4 is natively multimodal - trained on text, image, video, and audio data simultaneously rather than bolting visual capabilities onto a text-only base. This matches the direction of GPT-5 and Claude’s rumored next generation.
Benchmark claims circulating online suggest 80%+ performance on SWE-bench and ~90% on HumanEval, with inference costs projected at $0.10-0.30 per million input tokens - up to 50x cheaper than GPT-5. These claims remain unverified pending official release.
The Nvidia Lockout
DeepSeek broke from industry convention by denying Nvidia and AMD pre-release access to V4 for performance optimization. Instead, Huawei’s Ascend division received a multi-week head start.
The standard practice is for AI labs to share early models with all major chipmakers regardless of nationality. This allows hardware vendors to optimize drivers and kernels before public release. DeepSeek’s previous model, V3.2, followed this convention.
V4 breaks the pattern. One analyst quoted by Awesome Agents noted that “the impact to Nvidia and AMD for general data accelerators is minimal - most enterprises are not running DeepSeek.” But the strategic signal matters: DeepSeek is actively building what the report calls a “parallel AI software ecosystem on Chinese silicon.”
The rationale is straightforward. By giving Huawei’s Ascend division exclusive early optimization access, DeepSeek creates practical performance advantages for domestic hardware at launch. “It’s about who gets a head start on making it run well,” one source explained.
Huawei’s Ascend Chips
V4 is specifically optimized for Huawei’s Ascend 910C chips, which combine two 910B dies with high-bandwidth memory. According to CSIS analysis, DeepSeek has evaluated Ascend chips and found them “unattractive for training AI models” but capable of delivering roughly 60% of Nvidia H100 performance for inference.
That 60% figure isn’t necessarily a dealbreaker. Inference - running trained models to generate outputs - is becoming more important than training as AI capabilities mature. Barclays estimates that by 2026, 70% of AI compute demand will focus on inference rather than training.
For DeepSeek’s open-source strategy, inference matters more than training. Once V4 is trained, it needs to run cheaply at scale. If Huawei chips can do that at 60% of Nvidia’s performance but at substantially lower cost (and without US export control constraints), the economics work.
The challenge is software. Nvidia’s CUDA ecosystem provides substantial competitive advantages through established developer tools and libraries. Huawei’s CANN (Compute Architecture for Neural Networks) is comparatively immature. CSIS estimates 2-3 years minimum for ecosystem maturation even with substantial investment.
The Distillation Accusations
V4’s release comes amid accusations from Anthropic and OpenAI that DeepSeek and other Chinese AI companies systematically extracted capabilities from their models.
According to Fortune, Anthropic alleged that three Chinese firms - DeepSeek, Moonshot AI, and MiniMax - generated “over 16 million exchanges with Claude through approximately 24,000 fraudulent accounts” to train their own models on Claude’s outputs.
The technique is called distillation: training a smaller model using the outputs of a more advanced one. Anthropic acknowledges distillation is standard practice but argues the scale and deception involved here crossed lines.
Specifically, Anthropic claims DeepSeek alone generated over 150,000 exchanges targeting reasoning tasks, rubric-based grading suitable for reinforcement learning reward models, and - notably - “censorship-safe rewrites of politically sensitive queries.”
OpenAI has made parallel allegations. The company stated that “DeepSeek’s next model should be understood in the context of its ongoing efforts to free-ride on the capabilities developed by OpenAI and other US frontier labs.”
DeepSeek has not publicly responded to these accusations. But online commentators have noted the irony: Anthropic recently settled a $1.5 billion copyright lawsuit with authors over bulk data collection practices. The AI industry’s relationship with intellectual property is not straightforward.
Export Controls: What Worked, What Failed
V4’s existence reflects both the successes and failures of US export controls on advanced AI chips.
CSIS’s analysis traces the problem to a 2022-2023 window when Nvidia exploited redundancy mechanisms in chip design to create degraded versions (A800, H800) that technically complied with export restrictions while maintaining near-state-of-the-art performance. During that year, Nvidia generated over $9 billion in revenue from China - largely from these controlled chips.
The Biden administration closed this loophole in October 2023, but substantial quantities had already reached China. DeepSeek CEO Liang Wenfeng has been candid about the impact, stating that Chinese firms need “two to four times” the computing power to achieve equivalent results using H800 chips instead of H100s.
A senior Trump administration official told Reuters that DeepSeek’s V4 was trained using Nvidia’s most advanced Blackwell chip on a cluster located in mainland China. If true, this would appear to violate current US export controls - though verification is difficult.
Meanwhile, Huawei acquired over 2 million Ascend 910B logic dies through shell company arrangements with TSMC before enforcement of the Foundry Rule. This stockpile, combined with over a year’s worth of high-bandwidth memory, provides substantial near-term capacity for domestic chip deployment.
What V4 Means
DeepSeek’s V4 represents the most serious test yet of whether export controls can slow China’s AI development.
On one hand, the evidence suggests controls are working: DeepSeek needs inefficient workarounds and domestic chips that deliver 60% of Nvidia’s performance. The software ecosystem gap adds years to any timeline for true independence.
On the other hand, V4 exists. If it performs anywhere near the claimed benchmarks, it will demonstrate that frontier AI development is possible without continuous access to American hardware - just harder and more expensive.
The deliberate timing with China’s parliamentary meetings, the Huawei optimization, and the Nvidia lockout all signal strategic intent. This isn’t just an AI lab releasing a model. It’s a demonstration that China is building an alternative AI infrastructure stack.
For Western observers, the uncomfortable question is whether “harder and more expensive” is enough. Dario Amodei, Anthropic’s CEO, has observed that efficiency gains typically redirect investment toward greater model capability rather than reduced compute spending. If DeepSeek achieves frontier performance at lower costs, those savings don’t disappear - they fund the next generation.
The Bottom Line
V4’s release this week won’t settle the US-China AI competition. But it will clarify how that competition is evolving: not as a sprint where one side wins decisively, but as parallel ecosystem development where both sides build increasingly independent infrastructure.
DeepSeek is betting that open-source models optimized for domestic chips can compete with closed American systems running on Nvidia hardware. The distillation accusations suggest desperation from American firms watching their capabilities extracted. The Nvidia lockout suggests confidence from DeepSeek that they no longer need American cooperation.
Watch for V4’s actual benchmarks when they arrive. The claims are extraordinary. If they’re real, the AI landscape just shifted.