DeepSeek's $1 billion run rate, and what the number hides

China's cheapest frontier lab is growing fast and raising at a $75 billion valuation. But a run rate is not revenue, prices are going up, and compute and a data probe now set the limits.

DeepSeek has always been the lab that made AI cheap. This month it also became a real business: according to The Information, its annualised revenue run rate passed $1 billion, about double where it stood in July. The company is closing a $7.5 billion round at a $75 billion valuation by the end of October, with a listing on Shanghai’s STAR Market pencilled in for 2027.

Those are big numbers for a Chinese lab. They are still small next to the US leaders, and they need careful reading.

Growing fast, from a small base

Annualised revenue run rate, latest reported

Billions of US dollars. Anthropic's figure is from the end of July; investors expect it to pass $100 billion by year end.

Anthropic$65B
OpenAI$40B
Z.ai$1.8B
DeepSeek$1.0B
MiniMax$0.8B

US labs
Chinese labs

Source: TechCrunch, company statements via press reports, The Information, Silicon UK, China Biz Insider

Put together, the three Chinese labs with public figures add up to roughly $3.6 billion of run rate. OpenAI and Anthropic together are above $105 billion, about thirty times more. The gap is not closing in absolute terms, but the growth curves in China are steep.

Chinese labs: run rate over 2026

Millions of US dollars.

Z.ai
DeepSeek
MiniMax

Hollow points are approximate.

Source: China Biz Insider, KrASIA, Silicon UK, Global Semi Research, The Information

Price, not just demand

Part of DeepSeek’s jump is simply a higher price. Last month it raised the cost of its models by 2.3 to 4.5 times, with peak-hour prices for its V4 models roughly quadrupling. Founder Liang Wenfeng told investors the increase did not shrink the customer base. Even after it, DeepSeek remains far cheaper than US rivals.

The economics were already unusual. Reports put the gross margin of DeepSeek’s API business at 82.9% through July: even at rock-bottom prices it was charging several times what the compute cost. Other labs report very different figures, though the definitions are not identical.

Reported gross margins

Percent. Periods and definitions differ, so treat this as a rough comparison.

DeepSeek83%
Anthropic63%
OpenAI39%
MiniMax18%

Source: AI Weekly (DeepSeek API through July 2026; OpenAI Q1 2026; Anthropic 2026 projection), China Biz Insider (MiniMax H1 2026)

A run rate is not revenue

Run rate, or ARR, takes the latest month’s revenue and multiplies it by twelve. In a company whose sales double every few months, that single month runs far ahead of what has actually been booked. DeepSeek recognised about $71 million of revenue in the first seven months of 2026. Z.ai reported a $535 million run rate at the end of June against about $140 million booked in the first half. The pattern repeats at MiniMax.

Run rate against revenue actually booked

Millions of US dollars. Booked revenue for the period, scaled to a full year so the two are comparable.

Run rate at the end of the period
Revenue booked in the period, annualised

Z.ai$535M (June)$280M (H1)
DeepSeek~$450M (July)$121M (Jan to Jul)
MiniMax$400M (May)$233M (H1)

Source: PYMNTS and AI Weekly (DeepSeek, Jan to Jul 2026), Global Semi Research (Z.ai, H1 2026), China Biz Insider (MiniMax, H1 2026)

None of this makes the growth fake. It does mean a $75 billion valuation is 75 times a run rate and several hundred times the revenue DeepSeek has actually collected this year.

Compute is the ceiling

According to The Information, DeepSeek spends more than 70% of its computing power on training new models and less than 30% on serving customers. With US export controls limiting access to Nvidia’s most advanced chips, more revenue without more compute can only come from higher prices. The company is reportedly testing cheaper gaming GPUs from Nvidia to serve its models. Z.ai made the same point from the other side: it raised its year-end outlook only after a $5 billion raise covered its immediate compute needs.

The distillation probe

On 10 September Anthropic published a report accusing seven Chinese labs, including DeepSeek, Moonshot, Zhipu (Z.ai) and MiniMax, of routing customer requests through Claude to harvest its answers for training. China’s Cyberspace Administration has since focused on DeepSeek and Moonshot. Its question is not about Anthropic’s terms of service but whether sensitive data from Chinese users, including state-linked ones, reached US servers. The reported volumes are large: more than 23 million exchanges from Moonshot between May and July, and 12.1 million from DeepSeek over two weeks in July. No penalties have been announced.

What it means

For people building on AI: cheap Chinese models are not guaranteed to stay cheap. DeepSeek has shown it can raise prices sharply without losing customers, and its capacity is limited. Plan for price and availability changes, and keep a second provider ready.
For anyone reading AI headlines: treat run rates from fast-growing labs as a snapshot of their best recent month. Where booked revenue is available, it tells a more sober story.
For policy watchers: export controls are biting where they were meant to, at compute. The distillation fight has turned into a data-sovereignty question on both sides, and it lands just as the US and China prepare to talk about AI.
The bigger picture: Chinese labs are a real business now, with strong margins and fast growth, but they remain about thirty times smaller than the US leaders and more constrained by chips than by demand.