A Billion-Dollar Run Rate, Built on a Price Hike: DeepSeek Doubles Revenue and Readies a ¥50 Billion Shanghai Raise

The company that taught the AI industry to fear the phrase “price war” has just demonstrated the other end of the trick. According to a report by The Information on September 24, DeepSeek’s annualized revenue run rate has crossed $1 billion — more than double the figure of under $500 million reported just a few months ago. Reuters picked up the story the same day, citing the same reporting: the Chinese AI lab founded by quantitative-finance billionaire Liang Wenfeng is now finalizing a fundraising round of roughly ¥50 billion (about $7.45 billion) at a valuation of ¥500 billion, with a target of closing by the end of October ahead of a planned Shanghai listing.

The numbers behind the doubling

The most striking detail in the report is not the headline figure but what produced it. Liang told investors that the surge followed an increase in API prices of between 2.3x and 4.5x implemented in August — and that customers, by and large, stayed. Demand held even after the hikes, with reports of the price increase spreading across Chinese tech media in late August under headlines bidding farewell to DeepSeek’s old identity as the industry’s “price slasher.”

That pricing power shows up directly in the unit economics. The Information reports that DeepSeek’s API business ran an 82.9% gross margin through July — a figure that would put it in the same neighborhood as the software industry’s most profitable franchises, and a stunning reversal for a company that spent 2024 and early 2025 deliberately selling inference at prices competitors called unsustainable.

There is a caveat worth stating plainly, because critics already are: a run rate that doubles on the back of a 2.3x–4.5x price increase is not the same thing as 2x usage growth. Skeptical commentary within hours of the report argued the milestone reflects repricing more than organic expansion. That reading is fair as far as it goes — but it undersells what the price hike itself proves. In a market where Alibaba’s Qwen team cut voice-API prices by up to 95% this very week, DeepSeek raised prices as much as 4.5x and kept its customers. In commodity markets, that does not happen. It happens when a supplier has something buyers cannot easily get elsewhere: in DeepSeek’s case, frontier-class open-weight models whose cost-performance profile still anchors a large share of the global inference and fine-tuning ecosystem.

Where the money goes

Liang also gave investors a breakdown of compute allocation that is unusual in its bluntness: more than 70% of DeepSeek’s computing power is devoted to training new models, leaving less than 30% for inference serving. That is the allocation of a company that still thinks of itself as a research lab with a revenue line, not a business optimizing an existing product. It is consistent with what Liang said publicly in July, when Reuters reported him telling an audience that DeepSeek prioritizes the pursuit of AGI over maximizing profit and is likely to keep its strongest models open-weight.

The distinction matters for how to read the fundraise. The ¥50 billion round — reported in earlier iterations since mid-July as a $71–74 billion valuation deal, with The Wall Street Journal confirming in late August that the number was near $74 billion — is not primarily a war chest for go-to-market. It is a training budget. Under the reported structure, the round would be tied to a listing on Shanghai’s STAR Market, with CITIC Securities tapped to lead, according to reports earlier in September. At ¥500 billion, the valuation would make DeepSeek one of the most valuable AI companies in China and among the most valuable private AI companies anywhere.

From price slasher to price setter

The irony of the moment is hard to miss. In January 2025, DeepSeek’s R1 release upended global AI economics by offering near-frontier reasoning at a fraction of prevailing API prices, wiping hundreds of billions off the market caps of Western AI infrastructure stocks in a single day and forcing price cuts across the industry. Eighteen months later, the same company is charging 4.5x more at peak hours — and the market absorbs it.

The lesson the industry should take from this sequence is uncomfortable for anyone who declared inference a commodity in 2025. Commodities cannot raise prices 4.5x during a glut. DeepSeek could, because its open-weight models created an ecosystem dependency: enterprises, startups and national AI programs built around models they could download, fine-tune and serve themselves, and a meaningful share of them still buy the official API when it matters. Open weights, it turns out, can function as a moat with a revenue model attached — the downloads build the dependency, and the API monetizes it.

The gross margin figure sharpens the point. An 82.9% API gross margin through July is not the margin profile of a company subsidizing usage to buy market share. It is the margin profile of a company that has stopped competing on price because it no longer has to.

The Shanghai window

The listing plan is as significant as the revenue number. Chinese AI companies have spent 2026 splitting their exit paths: Moonshot AI has been reported preparing a Hong Kong IPO at a potential $50 billion valuation, while DeepSeek — smaller by headcount, more storied by reputation — is going onshore, tapping domestic capital markets at a moment when Beijing has made technological self-sufficiency in AI a national priority. A ¥500 billion STAR Market debut would be a milestone for that market and a test of how much domestic liquidity is willing to pay for frontier-lab exposure.

It also lands at a peculiar geopolitical moment. Xi Jinping’s state visit to Washington this week deliberately excluded DeepSeek’s Liang and Moonshot’s Yang Zhilin from the delegation — commentators read the omission as risk management around US exposure. Meanwhile the company itself lines up a domestic listing and reports to investors in yuan. The bifurcation of the global AI capital market — US labs raising record rounds at trillion-adjacent valuations, Chinese labs monetizing domestically — is no longer a forecast. It is a filing away from being a fact on the Shanghai Stock Exchange.

What to watch

Three things will tell us whether the $1 billion run rate is a milestone or a plateau. First, whether post-hike usage holds over a full quarter — a run rate is a trailing snapshot, and churn arrives slowly. Second, what the next open-weight release does to the dependency: every model DeepSeek ships at frontier level renews the ecosystem lock-in that makes its pricing power possible, while a weak release would erode it just as fast. Third, the actual IPO: a ¥500 billion valuation on ¥7+ billion of annualized revenue is roughly 70x ARR — a multiple that assumes the price-hike playbook has several more turns left in it.

For now, the company that once priced AI like a loss leader has become the industry’s cleanest demonstration that frontier capability, given away at the weights and charged for at the API, can be both the cheapest game in town and one of its most profitable businesses. The billion-dollar run rate is the receipt.