Invest1 distinct publisher2 min readPublished
The seven-month numbers behind DeepSeek's pre-IPO round pair the best gross margin anyone reports with one of the smallest revenue lines. A Shanghai filing would put both on the record.
The Investor · Invest desk
Compiled by The InvestorSomething wrong?How this is made
The only line in these numbers that beats anybody is the API margin. DeepSeek reports 82.9% on API access against 44.6% blended [6][5], which means whatever sits outside the API drags the company down by more than 38 points [8], and nothing disclosed says what that is. The Information attributes the margin to models that need fewer chips to stay competitive [14], and the price list is consistent with that: V4 Pro charges $0.87 per million output tokens, roughly a seventh of what Grok 4.6 charges, according to Cryptopolitan's earlier reporting [13].
Cheap inference cuts both ways. Apply the blended 44.6% to 475 million yuan of revenue and gross profit for the seven months is about 212 million yuan; to reach a 715 million yuan net loss, roughly 927 million yuan of cost has to sit below that line [6]. There is no revenue line here big enough to absorb it. OpenAI's first quarter alone, $5.7 billion at a 39% gross margin [8], is about 81 times DeepSeek's seven-month total, and Anthropic's $11.5 billion second quarter about 163 times it [7]. DeepSeek is selling the better unit economics and the much smaller business.
That business is being asked to absorb a 50 billion yuan round for the second time this year. June's round put in 50 billion yuan, about $7 billion, at a valuation above $50 billion [11]. The same nominal ticket at roughly $74 billion buys about 9.5% of the company rather than about 14% [10], which is the part of the story a revenue multiple does not capture: the price moved much faster than the operating history. The expected buyers, per the reporting, include returning backers Monolith, Shixiang Capital and CATL, first-time investors CPE, Legend Capital and Stony Creek Capital, and funds tied to GigaDevice and Hefei state investment vehicles [12]. State-linked money on the register is one reason a valuation can reprice that quickly without a public market involved.
One intervening detail matters for a company said to be picking banks for a Star Market listing: the second round was paused by founder Liang Wenfeng over leaked meeting comments [11]. A Nasdaq-style board in Shanghai [10] replaces the problem of controlling what leaks with the obligation to publish on a schedule, in an audited form, with segment detail. That is the real test of the disruptor-to-listed-company move, and it has not been taken yet: DeepSeek has not publicly confirmed the IPO talks [10], and the $74 billion currently rests on $70.7 million of reported revenue [1] and an implied 2025 base of around 47.5 million yuan [1].
Ranked by verification strength, evidence, and original report placement.
DeepSeek reported revenue of 475 million yuan ($70.7 million) for the first seven months of 2026, figures first reported by The Information.
The seven-month 2026 revenue figure is about ten times DeepSeek's full-year 2025 total.
Net loss for the first seven months of 2026 was 715 million yuan, against 935 million yuan for all of 2025.
DeepSeek's overall gross margin was 44.6%.
DeepSeek's API access business gross margin was 82.9%, a level neither OpenAI nor Anthropic reports.
OpenAI's first-quarter revenue was $5.7 billion at a 39% gross margin.
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Thin: one aggregator relaying leaked figures, with an internal contradiction
Every number in the cluster comes from a single crypto-and-tech outlet relaying The Information, IndexBox and its own earlier reporting about a private company that has confirmed nothing. There are no filings, audited statements, company statements or independent second sources, and the article contains an unreconciled inconsistency: $70.7 million over seven months cannot outperform a $400m-$500m annualized estimate. Peer comparisons for OpenAI and Anthropic are quoted without sourcing or period alignment. The arithmetic that can be checked internally is consistent, which keeps this above the floor.
Real paid usage, very small absolute base
Revenue is the only adoption proxy available and it points to genuine, fast-growing paid API consumption: 475 million yuan over seven months, roughly ten times the 2025 base, earned at a quoted $0.87 per million output tokens with an 82.9% API gross margin. But the absolute base is tiny - about $121 million annualised, roughly 1/81st of a single OpenAI quarter and 1/163rd of an Anthropic quarter - and the cluster supplies no user counts, token volumes, enterprise deployments or named customers. Adoption is therefore evidenced in direction but small and indirectly measured.
Overstated: growth and valuation framing outrun the figures
The framing consistently reads better than the arithmetic it is built on. A 'tenfold' jump annualises to about $121 million against a $74 billion estimated valuation, roughly 610 times revenue; the loss described as 'reduced' is running about 31% above the 2025 monthly pace; and the article asserts an estimate was beaten by a factor its own numbers reverse. The 82.9% API gross margin is a genuine, non-hyped datum and the piece does concede DeepSeek 'cannot compare' on scale, which keeps this short of the extreme.
High: leaked financials mid-raise, relayed by a self-citing aggregator
The figures surface while DeepSeek is reportedly trying to close a 50 billion yuan raise at a 500 billion yuan valuation before end of August and to line up a Star Market filing - precisely the moment at which flattering, unaudited numbers benefit the company and its returning backers, and after a prior round was paused over leaked meeting comments. The publisher adds its own incentive layer: it cites its earlier reporting as support for the price comparison and closes with a newsletter solicitation. Nothing here is company-attributed or on the record, so no accountability attaches to the numbers.
Low-moderate: arithmetic is solid, inputs are not
Confidence is limited by the inputs rather than the reasoning. The derived figures - annualisation, revenue multiple, loss run rate, below-the-line cost, implied stake - follow deterministically from numbers the article states, so the internal analysis is robust. But those numbers are single-sourced, unaudited and partly self-contradictory, the company confirms nothing, and there is no second publisher in the cluster to cross-check. A Star Market filing or independent corroboration would move this materially.
build
Developer habit, priced at $965B: what Anthropic's run actually proves1 distinct publisher
product
A 27B laptop model scores like a rented one, and thinks three times as hard to do it1 distinct publisher
build
Three frontier launches in a day, all pitched on price. Open weights set the ceiling.4 distinct publishers
build
Your Multi-Key Failover Is The Most Expensive Line On Your Coding Agent Bill1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 26, 2026