Chinese AI startup DeepSeek has made a permanent 75 percent reduction to the price of its V4‑Pro model. Under the new pricing, token costs range from $0.0035 to $0.83 per million tokens—substantially lower than any comparable US frontier model.
What changed technically and on price
V4‑Pro had previously been priced at twelve times the cost of a lighter sibling model, a premium justified by DeepSeek on the basis of higher compute capacity requirements. In the second half (H2), Huawei deployed Ascend 950 supernodes, which provided the compute capacity that allowed DeepSeek to run V4‑Pro more cheaply and pass a large share of savings to customers.
Regulatory and geopolitical context
US export controls aimed to limit China’s access to high‑end AI hardware in order to slow Chinese advances in foundation and large models. The recent developments undercut key assumptions behind those controls:
- Huawei shipped Ascend 950 supernodes;
- DeepSeek deployed V4‑Pro on that hardware;
- prices for frontier‑level service fell sharply.
As a result, one of the cheapest frontier‑tier AI offerings on the market today runs on a chip the US had sought to restrict.
Market strategy and implications
Companies such as Anthropic and OpenAI have positioned themselves partly on the idea that developers will pay a premium for top performance. DeepSeek’s strategy, by contrast, targets broad uptake of lower‑priced “good‑enough” models. If this approach continues, the commercial meaning of the AI frontier could shift away from where developers buy the most expensive systems toward where the most affordable, sufficiently capable systems are available.
Conclusion
In short: Washington banned a specific chip to impede Chinese AI progress; China produced and deployed a comparable chip, and a very low‑cost frontier AI product now operates on that hardware. The chain of events raises questions about the effectiveness of export controls and may reshape price competition in the global AI market.
