DeepSeek says its revenue run rate hit $1 billion -- almost entirely off a price hike, not new demand

DeepSeek chief executive Liang Wenfeng told investors on Sept. 23 that the company's annualized revenue run rate has hit $1 billion, more than double the under-$500-million pace of a few months earlier. The figure lands as DeepSeek finalizes a second funding round targeting roughly 50 billion yuan (about $7.5 billion) at a valuation near 500 billion yuan (about $75 billion), with a close targeted for end of October and a Shanghai STAR Market listing -- lead-underwritten by CITIC Securities -- expected in 2027.

A run rate is one period's revenue multiplied out to a full year, and the period doing the multiplying here is a loud one. Last month DeepSeek raised API prices 2.3x to 4.5x across its model lineup -- the kind of increase that would normally cost a vendor customers. Liang told investors it didn't: demand held. That single claim is carrying most of the headline number, and it comes from the company itself, one month after the fact, with no independent usage data behind it yet.

There's a reason the price hike is doing so much work. DeepSeek built its reputation, and much of its user base, on undercutting Western labs on price -- its models have run at a fraction of GPT and Claude API rates since the original DeepSeek-R1 launch triggered a market-wide repricing in January 2025. An August increase that still leaves DeepSeek, in Liang's own framing, "among the lowest" prices for a major model is less a reversal than a recalibration: the company found out how much of its discount it could claw back before users left, and the answer -- so far, for one month -- was most of it.

The margin split matters because it's the API business -- the 82.9% line -- that the price hike touched directly, while the 44.6% company-wide figure is what actually funds everything else DeepSeek does, including the model training that produced the 82.9% margin in the first place. Set against rivals, DeepSeek's API economics look unusually efficient: OpenAI posted a 39% gross margin in the first quarter of 2026 on $5.7 billion in revenue, while Anthropic posted $11.5 billion in second-quarter revenue with a margin it expects to climb from 40% in 2025 toward 63% by the end of 2026. DeepSeek's API slice already clears both -- though DeepSeek is a fraction of either company's absolute revenue, and its company-wide margin trails both.

Run the two headline numbers against each other and the valuation is asking for 75 times the $1 billion run rate -- a multiple that only looks ordinary in a market where AI labs are routinely priced on where their revenue is headed, not where it is. That's the same logic investors have applied to funding rounds across the sector this year, but it means DeepSeek's valuation is a bet on the price hike sticking through a second, third and twelfth month, compounding into the roughly $1 billion of actual annual revenue the run rate implies -- not a bet on the $70.7 million the company had actually booked by the time the round started coming together.

The skeptical read isn't that DeepSeek is lying -- it's that a founder's investor-meeting figure, repeated to reporters rather than filed with a regulator, is exactly the kind of number that looks strongest right when a company most needs it to. DeepSeek is mid-fundraise and pre-IPO; a $1 billion run rate is a materially better story to tell a prospective backer than $70.7 million in booked first-half revenue, even if both numbers are technically accurate and describe the same company a few weeks apart. Nothing here suggests the figure is fabricated. It does mean the run rate is a claim about the future built on one month of the past, at a moment when the company has every incentive to make that one month look representative.

None of this is DeepSeek's first valuation headline this year. This is the same round that was paused in July after a leaked investor transcript went viral in China, then resumed Aug. 6 with existing backers Monolith, Shixiang Capital and CATL confirmed -- and reported nearing a close at roughly $74 billion by late August. The September figures don't represent a new round; they're the same fundraise, closer to done, with a fresher revenue number attached and the valuation essentially unchanged from a month earlier.

A Shanghai listing would also be DeepSeek's first real test of investor appetite beyond the founders and venture backers who've priced it so far. CXMT's 466% debut and Unitree's six-times pop both happened on a STAR Market that has rewarded AI-adjacent listings enthusiastically all year, which is exactly the kind of receptive market a company wants to go public into -- and exactly the kind of market where an early pop can say more about scarcity of AI-themed shares on a constrained exchange than about the specific company's fundamentals. Whether DeepSeek's own debut, whenever it lands, reads more like durable demand or like the same pattern repeating is a question the prospectus and the first trading day will answer separately.

That prospectus, whenever it's filed, would be the first document in this entire story carrying an independent auditor's signature rather than an investor-meeting readout -- the gap this whole account has been circling. Until then, the $1 billion figure, the $75 billion valuation, and the 82.9% margin are all one company's own account of itself, at the exact moment that account is most valuable to have believed.