China's LLMs quit the price war as DeepSeek API jumps 11x
China's AI labs are closing the book on the "cabbage price" era — the two-year race to give away model inference is reversing, and the bill is arriving fast. Three moves from the last 24 hours show the industry pivoting from land-grab to margin.
China's large model vendors are ending the price war, with DeepSeek leading the climb at API increases of up to 11x. DeepSeek's new peak/off-peak (峰谷) schedule puts output tokens at roughly $1.32 per million at peak hours and about $0.66 off-peak, while Zhipu (智谱) has lifted prices around 83% — and Baidu's index is now full of headlines declaring the "白菜价" (rock-bottom) age over. The shift is systemic, not a one-off: after two years of subsidized inference to win developer mindshare, the major Chinese labs are now pricing for sustainability ahead of expected public offerings. We covered DeepSeek's first move to peak-hour billing last week — DeepSeek hikes API prices up to 4.7x with new peak-hour billing — but this is the whole field turning, not a single vendor's tweak.
DeepRoute.ai (元戎启行) has stood up what it calls China's first Physical AI basic research laboratory, and it is staffed by ex-DeepSeek talent. The autonomous-driving firm brought in Ruan Chong (阮翀), a former core multimodal researcher from DeepSeek, to lead the new organization, betting that the next competitive frontier is "physical world fundamental capabilities" — world models that understand and act in real space rather than just language. The move fits a broader industry tilt: as pure text-model differentiation narrows, embodied and physical-AI capability is becoming the contested ground for the next wave of Chinese AI investment.
Ulanqab, in Inner Mongolia, is now home to a 172,000-P compute "super-monomer" — and the constraint is no longer chips, it's power. Reporting around Envision (远景) executive Zheng Zihao's comments frames large-model training as spawning ever-larger single-site clusters where green electricity is the binding bottleneck; the local tagline is literally "compute chases the green grid." The buildout underscores where China's AI buildout is headed: massive centralized clusters sited next to renewable generation, with energy procurement — not GPU supply — setting the pace of expansion.
What to watch: whether the price hikes stick once developers can route traffic to still-subsidized rivals — and whether Ulanqab-style green-powered super-clusters become the default template for national AI capacity.
Is the end of cheap Chinese inference a sign of a maturing market or a bet that developers have nowhere else to go? Tell us in the comments.
