From Price Hike to $1B Run Rate: DeepSeek Doubles Revenue and Closes In on a $7.5B Round

DeepSeek — the Hangzhou lab whose open-weight models upended global assumptions about the cost of frontier AI — has crossed a line that few Chinese AI startups have reached. According to a report from The Information on September 24, 2026, the company’s annualized revenue run rate has hit $1 billion, more than doubling from under $500 million just a few months ago. The same report, citing two people with direct knowledge of the matter, says DeepSeek is simultaneously finalizing a $7.5 billion fundraising round as it prepares for a potential public listing.

The turnaround is stark. In July, the startup had told prospective investors it was suspending its second funding round after viral posts by an investor sparked controversy, as Fortune reported at the time. Less than three months later, the round is being closed — and at materially better economics than anyone inside or outside the company expected in the spring.

The numbers behind the run rate

Annualized run rate is a simple extrapolation: take the most recent month’s revenue and multiply by twelve. Crossing $1 billion on that basis means DeepSeek is now generating revenue at a pace that puts it in rare company — OpenAI, Anthropic, Google — and nearly all of it is earned in China, from Chinese developers and enterprises, priced in yuan.

The proximate driver is no mystery. On August 16, 2026, DeepSeek put into effect the most aggressive repricing of its API in the company’s history, raising rates on DeepSeek-V4-Flash and DeepSeek-V4-Pro by anywhere from roughly 57% to more than 1,100% depending on the model, token type, and time of day. Bloomberg reported that peak-hour pricing alone increased the company’s rates to more than four times previous levels. The move shocked developers who had built on the assumption that DeepSeek would remain the industry’s permanent discount rack.

It did not last in full. By early September — just 23 days after the hikes took effect — DeepSeek partially walked them back, cutting cache-hit input prices by up to 60% while leaving output prices modestly reduced and still roughly double their original launch levels. The rollback was read as a concession to developers threatening to migrate to open-source alternatives and to rival Chinese labs like Alibaba’s Qwen, which days later cut its own voice-API prices by as much as 95%.

But even the trimmed prices are dramatically higher than what DeepSeek charged through the first half of 2026, and demand has absorbed them. Per StoneX’s IPO analysis, DeepSeek generated roughly 475 million yuan (about $70.7 million) in revenue in the first seven months of 2026 — roughly ten times its full-year 2025 figure. Hitting a $1 billion annualized pace by September implies monthly revenue accelerated severalfold in a matter of weeks after the repricing landed.

The $7.5 billion round

The funding picture has moved just as quickly. DeepSeek’s first external round, closed around the end of May 2026, raised about $7.4 billion at a valuation of just over $50 billion, with founder Liang Wenfeng personally contributing roughly $3 billion and retaining control through an unusual structure that kept the company founder-led.

The second round now being finalized is slightly larger — $7.5 billion — and reportedly comes at a valuation near $71–75 billion, according to Financial Times and Dealroom reporting ahead of the close. The pause announced in July is over. Liang, whose quant fund High-Flyer incubated and continues to bankroll the lab, remains the controlling shareholder; Reuters noted earlier this year that his personal stake makes him one of China’s wealthiest AI founders.

What the money buys is compute and people. DeepSeek has been on a hiring surge — Dealroom reported roughly 150 new engineers as the round neared completion — and, like every frontier lab, it faces steep inference costs as usage of its models scales.

Why an IPO, and why now

The listing question frames everything. Chinese listings by AI companies have surged in 2026 as Beijing pushed labs toward domestic capital markets and away from US investors, and DeepSeek’s file is unusually clean from a Beijing perspective: it takes no foreign money, its models are open-weight, and its revenue is almost entirely domestic.

An IPO would also give DeepSeek something it has never had: public scrutiny of its finances. Skeptics have long noted that run-rate figures can be flattered by a single strong month, and that DeepSeek’s revenue is heavily API-driven in a market where rivals are cutting prices aggressively. Alibaba’s Qwen price cuts, Moonshot’s enterprise push, and the possibility that the CAC probe reported on September 23 extends into commercial practices all bear on how durable the $1 billion pace proves to be.

The bull case is equally concrete. DeepSeek’s models remain at or near the top of open-weight leaderboards, its inference costs per token are among the lowest in the industry thanks to architectural innovations like multi-head latent attention, and Chinese enterprise demand for domestic models keeps compounding as procurement rules favor them.

The bigger picture

Two structural trends collide in this story. The first is the end of free-or-nearly-free inference: every frontier lab — OpenAI, Anthropic, and now DeepSeek — has spent 2026 raising prices to bring model economics closer to actual cost, and DeepSeek’s $1 billion milestone is the clearest evidence yet that users will pay substantially more than the 2025 price book implied. The second is China’s AI sector maturing into revenue-generating businesses with public-market ambitions, no longer just state-subsidized research prestige projects.

For developers, the message is mixed. The era of using DeepSeek as a loss-leader arbitrage on compute is over, but the models are still, by most benchmarks, the best value per unit of capability available under an open license. For the industry, the message is simpler: open-weight AI can plausibly be a very large business — and it may soon be one you can buy on an exchange.