DeepSeek released on Friday the preview of V4-Pro, a 1.6 trillion parameter model the company says rivals what closed American labs ship. The catch is not just the size: the model was optimized to run on Huawei Ascend chips, not Nvidia GPUs.
Why it matters
A year ago, DeepSeek V3 shook the market when it proved you could train a frontier model spending a fraction of what OpenAI and Anthropic spend. Now V4 crosses the line that still seemed safe: dependency on American hardware. If a model of this size really runs well on Chinese silicon, the thesis that Nvidia holds the world by the throat starts to crack.
- Preview ships in two versions: V4-Pro (1.6T parameters, 1M context) and the cheaper V4-Flash
- Pricing lands at $3.48 per million output tokens — vs $25-30 for Western competitors
- Huawei published a note confirming "full support" with the Ascend 950 line and Supernode architecture
- The US government raised the volume the same day, accusing DeepSeek of intellectual property theft
- Anthropic leads benchmarks in March, but the margin got narrow
What this means for creators
If you use an LLM API in your pipeline — script generation, assisted editing, moderation automation — the ceiling on cost just dropped again. A frontier model at 10% of the price of the leading GPT changes what is worth building: that idea that was too expensive to bother with now pays for itself. For anyone outsourcing inference to a Brazilian or European provider, watch closely — many will start offering DeepSeek V4 hosted in weeks, not months.
The other angle is geopolitical and touches you indirectly: if Washington tightens the screws on Chinese models, you may have to choose between running cheap (DeepSeek) or running "clean" for US enterprise clients.
The real read
It is not V4 that matters. It is the signal: Nvidia's hardware moat is leaking, and the capital moat of Western labs is no longer decisive. Anyone who was waiting for inference prices to drop another order of magnitude — congratulations, it started.
Sources
- CNN Business (April 24, 2026): China's AI upstart DeepSeek drops new model. Will it make waves like last year? — https://www.cnn.com/2026/04/24/tech/chinas-ai-deepseek-v4-intl-hnk