DeepSeek Raised Prices and Doubled Its Revenue Run Rate

When DeepSeek raised its API prices in August, the obvious question was how many developers would leave. Six weeks later the company has an answer, and it is not the one the price-war narrative predicted. DeepSeek's annualized revenue run rate has passed $1 billion, more than double the rate of a few months ago, The Information reported on September 23. Founder and CEO Liang Wenfeng disclosed the figure at an investor meeting and told investors the increase had not reduced the customer base.

The price increase was not a trim. DeepSeek announced on August 13 that V4-Flash and V4-Pro prices would rise on August 16, by 2.3 to 4.5 times depending on the model, per The Information and Reuters. V4-Flash output tokens went from $0.28 per million to $1.32 at peak. Reuters, which carried the report, said it could not independently verify the figures.

The run rate outruns the books

The scale of the jump is clearest next to DeepSeek's booked numbers. The Information reported in August that DeepSeek generated about 475 million yuan, roughly $70.7 million, in the first seven months of 2026. That was about ten times its revenue for all of 2025, according to Digital Today's summary of the report, and it put the run rate at the time at under $500 million.

A $1 billion run rate works out to roughly $83 million a month. That means DeepSeek is now annualizing a monthly figure larger than everything it booked from January through July combined. Some of that is the steepness of the growth curve. Most of it, in the timing, is the price hike landing on volume that did not shrink.

The margins were already unusual before the hike. Per The Information's August figures, DeepSeek's overall gross margin was 44.6% for the first seven months, while its API business ran at an 82.9% gross margin. For comparison, the same reporting put OpenAI's first-quarter gross margin at 39% on $5.7 billion of revenue. DeepSeek still lost money, a net loss of about 715 million yuan over the period, because the company spends heavily on the thing that does not show up in API gross margin: training.

Scarce compute, not loyal customers

The detail that explains why the hike worked sits in the compute split. DeepSeek devotes more than 70% of its computing capacity to training new models and less than 30% to inference for users, according to The Information's reporting as summarized by Tech Startups.

That changes how the price increase should be read. DeepSeek did not test how much developers would pay in a market with spare capacity. It repriced a product it could not supply enough of at the old price. When the company introduced peak and off-peak rates in August, it said it wanted to "allocate resources more reasonably," a polite way of saying demand at working hours exceeded what its inference fleet could serve. A price hike under those conditions does not so much shed customers as ration them. The ones who stay pay more, and the ones pushed to off-peak windows fill capacity that was sitting idle overnight.

That is also why the demand claim is believable, even unverified. Even at the new peak rates, DeepSeek remains far cheaper than Western frontier APIs. In July, OpenRouter routing data analyzed by CNBC showed DeepSeek as the single largest vendor on that platform, at 17.6% of routed tokens. Developers who built pipelines on DeepSeek were not choosing between DeepSeek and free. They were choosing between a 4x increase on a very small number and a migration to a vendor still many times more expensive.

What pricing power means for Chinese labs

For two years the working assumption about Chinese model labs was that they competed on price because they had nothing else to compete on, and that any lab that raised prices would be undercut by the next open-weight release. DeepSeek's run rate is the first large piece of evidence against that assumption. A Chinese lab raised prices sharply, in a market full of cheaper alternatives, and revenue went up.

The honest caveats are real. The $1 billion number comes from the CEO, delivered to investors in the middle of a fundraise, and no one outside the company has audited it. A run rate is a snapshot of recent weeks multiplied by twelve, not a year of revenue. And DeepSeek's open weights mean the same models are available from other hosts, so its pricing power applies to its own first-party API, not to the models themselves.

Still, the direction matters. Pricing power in inference comes from three things: model quality, reliability of supply, and switching costs built up in production code. DeepSeek appears to have enough of all three to charge more. That should worry the U.S. labs less than it worries the hosts and resellers who built businesses on the gap between DeepSeek's old price and everyone else's.

The money behind the timing

The run-rate disclosure did not arrive by accident. DeepSeek is finalizing a second funding round of 50 billion yuan, about $7.5 billion, at a 500 billion yuan valuation, and expects to close it by the end of October, per Reuters and PYMNTS. It is also preparing a listing on the Shanghai Stock Exchange's STAR Market. A company in that position wants its most recent numbers to show pricing discipline and high-margin revenue, and a price hike that did not dent demand is the cleanest version of that story it could tell.

The consumer DeepSeek chatbot remains free, which means essentially all of this revenue comes from developers paying for API access. That is a narrower base than OpenAI or Anthropic can claim, and it is concentrated in exactly the customers most sensitive to price. So far, they have paid.

What to watch

The next test is capacity. If the new round buys more inference hardware, DeepSeek will have to decide whether to keep prices where they are and bank the margin, or cut them again to take share from Alibaba's Qwen and Moonshot's Kimi. The first choice is what a company preparing for public markets usually does. The second is what DeepSeek did for most of its history. The August hike, and the revenue that followed it, suggest the company has already decided which kind of business it wants to be.