DeepSeek raised its prices and kept its customers. Now it's raising $7.5 billion.

The underdog is now a billion-dollar business

DeepSeek, the Chinese lab that touched off an AI price war in early 2025 by releasing capable models at a fraction of Western prices, has reportedly hit a $1 billion annualized revenue run-rate and is finalizing a raise of about 50 billion yuan, roughly $7.5 billion, at a valuation near 500 billion yuan, or about $75 billion. It is also preparing to list on the Shanghai Stock Exchange. The reporting, from The Information on 23 September, is based on figures CEO Liang Wenfeng reportedly shared with investors, not a public filing, and DeepSeek has not confirmed them, so treat every number here as reported rather than audited.

Even reported, the trajectory is steep. DeepSeek generated only about $70.7 million, or 475 million yuan, across the first seven months of 2026. A $1 billion annualized run-rate is roughly a tenfold jump on last year, and it is the kind of curve that turns a research darling into an IPO candidate.

It got there by charging more

Here is the part that should stop you. The lab famous for making AI cheap raised its prices last month, by a reported 2.3 to 4.5 times depending on the model, and, according to what Liang told investors, demand did not fall. The company DeepSeek's whole legend was built on undercutting everyone chose to test its pricing power, and the test reportedly passed.

That is the opposite of the story the industry tells itself. The assumed law of AI pricing is a race to zero, where models are interchangeable and the only direction for prices is down. DeepSeek just ran the other way and, if the reporting holds, kept its customers.

Still the cheapest game in town

The reason the price hike stuck is visible on DeepSeek's own pricing page, and it is the one hard, checkable fact in this story. Even after raising prices, DeepSeek is dramatically cheaper than the US frontier labs. Its pro-tier model lists around $0.66 per million input tokens and $1.98 per million output off-peak, and its lighter model runs $0.15 and $0.60. Set that against the GPT-6 and Claude prices that fell in September: $2 and $10 per million for GPT-6 Sol, $4 and $20 for Claude Opus 5.5. DeepSeek can raise prices 4x and still undercut a US lab by an order of magnitude.

So the customers who stayed were not being generous. They were doing arithmetic. When your baseline is a tenth of the alternative, a price increase that leaves you still the cheapest option is a margin expansion, not a gamble. DeepSeek did not defy the laws of pricing. It exploited how much room it had underneath everyone else.

What this does to the "cheap AI" story

The lazy version of DeepSeek is the scrappy underdog, the efficiency hackers who embarrassed Silicon Valley on a shoestring. That framing is now out of date. A company at a reported $1 billion run-rate, raising $7.5 billion at a $75 billion valuation, with enough pricing power to hike rates without losing customers, is not an underdog. It is a large, high-margin business with a moat made of cost structure, heading for a public listing.

The efficiency narrative quietly became a profit narrative, and that is the more important story. DeepSeek's cheapness was never charity; it was a wedge. The wedge worked, the wedge is now being monetized, and the same low-cost engineering that won attention in 2025 is what lets it charge more in 2026 and still win on price.

The caveats that matter

Read all of this through the fog that surrounds DeepSeek. The revenue and raise figures come from a private investor briefing relayed by one outlet, not from audited accounts, and the company has said nothing publicly, so there is no on-the-record quote to stand behind, only paraphrased reporting. A Shanghai listing also carries its own weather: Chinese regulatory approval, US export controls on the chips that train these models, and geopolitics that can reprice a company overnight. The $75 billion number is a target, not a close.

Our take

Strip the mythology and DeepSeek looks less like a disruptor and more like a winner consolidating. The reported numbers, a $1 billion run-rate and a $7.5 billion raise, are unconfirmed but credible, and the one fact anyone can check, its published prices, explains the rest: DeepSeek raised prices and kept customers because it is still cheaper than everyone who matters. The interesting lesson is not that AI is cheap. It is that being the cheapest is a business model with pricing power, and DeepSeek has quietly figured out how to charge for it.