On April 24, a startup in Hangzhou quietly released the largest open source AI model ever built. Washington called it theft. Developers called it Christmas.
The price gap that should embarrass every American AI lab
I have been watching the AI pricing wars for two years, and nothing has landed quite like this. DeepSeek V4 Pro costs $3.48 per million output tokens. OpenAI charges $30 for the same work. Anthropic charges $25. That is not a discount. That is a demolition.
The thesis here is simple and I will not soften it: open source AI just won a round that Silicon Valley did not see coming. V4 Pro packs 1.6 trillion parameters, supports a 1 million token context window, and runs on Chinese chips. That last part is the real story.
DeepSeek partnered with Huawei, using its Ascend 950 chips and Supernode cluster technology to power V4. Counterpoint Research analyst Wei Sun put it plainly: V4 running on domestic chips could accelerate adoption and reduce global reliance on Nvidia. That is not a minor footnote. That is a geopolitical shift.
“It allows AI systems to be built and deployed without relying solely on Nvidia, which is why V4 could ultimately have an even bigger impact than R1.”
— Wei Sun, Principal Analyst, Counterpoint Research
What V4 actually does that the headlines miss
The architecture upgrade is genuinely impressive. V4 Pro requires only 27% of the compute FLOPs and 10% of the memory cache compared to DeepSeek V3.2 when running at 1 million token context. That is not incremental progress. That is a rethink of how large models should work.
On coding benchmarks, V4 performs comparably to GPT-5.4. On formal math, V4 Flash Max scored 81.0 on Putnam-200 Pass@8 versus 35.5 for the next best open model. These are not vanity numbers. These are the kinds of scores that make enterprise buyers switch vendors.
The smaller Flash model costs just $0.28 per million output tokens. That undercuts every comparable model from OpenAI, Google, and Anthropic. Every single one.
The distillation accusation is real but it misses the bigger point
Now for the counterpunch, because it deserves a fair hearing. Anthropic and OpenAI have accused DeepSeek of distillation: feeding their models thousands of questions, collecting answers, and training a new model to mimic the reasoning. The White House called it industrial-scale IP theft.
I do not dismiss that concern. If the distillation claims are proven, that is a serious problem. But here is what the accusation cannot explain: DeepSeek's architectural innovations are real and documented. The Hybrid Attention Architecture, the compressed sparse attention, the efficiency gains on Huawei chips. None of that is copied from a ChatGPT prompt.
Calling V4 stolen is convenient. It lets American labs avoid the harder question.
Open source is the strategy and it is working
V4 is available on Hugging Face under the MIT license. Developers can download the weights, run the model locally, and fine-tune it for their own use cases. This is the open strategy that American labs have largely abandoned in the race for revenue.
MIT Technology Review noted that V4 marks DeepSeek's most significant release since R1, the January 2025 model that stunned global markets by delivering near frontier performance for under $6 million in training costs. V4 is the follow-through. It shows that R1 was not a fluke.
The good edge here is undeniable: a developer in Lagos, Manila, or Nairobi can now access frontier-level AI reasoning for pennies per query. That is genuinely good for the world, regardless of where the model was built.
What Washington gets wrong about this moment
The bad edge is the policy response. Banning DeepSeek from government devices while doing nothing to close the price gap is unserious. Italy, South Korea, and the US have all restricted DeepSeek in government contexts. Germany pulled it from app stores entirely. None of that stops a developer from downloading V4 weights tonight.
The real policy question is not how to block DeepSeek. It is why American labs charge ten times more for comparable work and call that a competitive advantage. Export controls on Nvidia chips were supposed to slow China down. Instead, they pushed DeepSeek to build a model that runs on Huawei hardware. That is the opposite of the intended outcome.
“V4 could be an early sign that China is successfully building a parallel AI infrastructure.”
— MIT Technology Review
Would you trust a frontier AI model built on Huawei chips with your company's most sensitive data? That is a fair question. But it is a different question from whether V4 is technically impressive. It is both things at once.
The AI race just got a lot more interesting. And the team that was supposed to be losing just handed the world its best open source model for free. Tell me that is not worth paying attention to.