China's most important AI lab just made its most important move yet. DeepSeek released V4 on Friday, an open-source model with 1.6 trillion parameters that matches or trails the best American closed-source models by a margin so thin it barely matters. The kicker: it was trained on Huawei's Ascend processors. Not Nvidia. Chinese chips.

Fifteen months after DeepSeek V3 wiped a trillion dollars off US tech stocks, the sequel is here. And it brings a message that Washington really does not want to hear: export controls did not work.

The Numbers That Matter

V4-Pro has 1.6 trillion total parameters with 49 billion active per token. It comes with a native one-million-token context window, which is table stakes for 2026 but notable because DeepSeek built the entire architecture around it rather than bolting it on. The model uses a hybrid attention mechanism that slashes inference costs by 73% at full context compared to V3.

On benchmarks, V4-Pro outperforms every other open-source model in agentic coding and reasoning. It scores competitively against Anthropic's Claude Opus 4.6, OpenAI's GPT-5.4, and Google's Gemini 3.1-Pro. DeepSeek's own tech report admits it "falls marginally short" of GPT-5.4 and Gemini 3.1-Pro, placing itself about 3 to 6 months behind the absolute frontier. On Codeforces, V4-Pro ranks 23rd among human competitors with a 3,206 rating.

Now look at the pricing. V4-Pro costs $3.48 per million output tokens. OpenAI charges $30 for GPT-5.5. Anthropic charges $25 for Opus 4.7. DeepSeek is offering roughly 90% of the performance at roughly 10% of the price. V4-Flash, the smaller variant with 284 billion parameters, costs $0.28 per million tokens. That is not a typo.

The Huawei Signal

This is the part that changes the geopolitics. Previous DeepSeek models relied on Nvidia GPUs, many of them allegedly acquired before export restrictions tightened. V4 was trained on Huawei's Ascend AI processors, manufactured by SMIC in China. Huawei confirmed on Friday that its latest AI computing cluster fully supports DeepSeek's V4 model.

SMIC shares jumped 10% in Hong Kong on the news. And they should have. What DeepSeek just demonstrated is that China can build frontier-competitive AI models on domestically produced chips. The entire thesis behind US export controls was that denying Nvidia GPUs would slow Chinese AI development by years. DeepSeek V4, trained on Huawei Ascend chips and performing within spitting distance of GPT-5.5, is the counter-argument in silicon.

The Open-Source Weapon

V4 is open-source. Weights on Hugging Face. Download, fine-tune, deploy anywhere. DeepSeek has been doing this since V3, and the strategy keeps paying off: open-source models gave DeepSeek a global developer army that no marketing budget could buy. It forced OpenAI to release its own open-source model last August. It inspired Alibaba, Moonshot AI, MiniMax, and Knowledge Atlas to do the same.

But V4 drops into a more crowded landscape than V3 did. Alibaba, Moonshot, and others have all released competitive open-source models this year. MiniMax and Knowledge Atlas, two DeepSeek competitors that went public in Hong Kong, both sank more than 9% on Friday. DeepSeek is eating its own ecosystem.

What Happens Next

DeepSeek is raising money for the first time. The Financial Times and The Information both report that Tencent and Alibaba are in talks to invest at a $20 billion valuation. DeepSeek does not need the money; its parent hedge fund High-Flyer is cash-rich. It needs the valuation to retain talent. AI researchers want equity in a company worth something, and DeepSeek's competitors are waving nine-figure retention packages.

DeepSeek expects to lower V4 prices further as Huawei scales production of its new Ascend 950 processors. Let that sink in. The model is already 90% cheaper than American alternatives, and the prices are going down.

OpenAI released GPT-5.5 yesterday at double the price of GPT-5.4. DeepSeek released V4 today at a fraction of the cost. One of these strategies works in a world where intelligence is abundant and getting cheaper. The other works in a world where intelligence is scarce and customers have no alternatives. That second world no longer exists.

First reported by Bloomberg. Technical report available on Hugging Face.