Anthropic’s revenue for the six months to 30 June 2026 was $16,286 million, according to its amended IPO filing of 2 November. Zhipu (Z.ai) and MiniMax, the two Chinese model makers listed in Hong Kong, reported about $142 million and $116.6 million for the same half in their interim results. By our arithmetic, Anthropic brought in about 63 times the two combined.
IndexBox, citing the South China Morning Post, wrote that Anthropic “reported a gross profitmargin of 85.4 per cent” in the half, against 26.4% for Zhipu and 17.9% for MiniMax. The two Hong Kong figures match the companies’ filings. Anthropic’s filing does not report a gross profit line; the 85.4% is a calculation from a footnote, and it is not built the same way as the Hong Kong numbers.
Business Pill · Gross Margin
An evergreen Business Engineer explainer on gross margin: what an AI company keeps from each dollar of revenue. It explains the general idea and does not discuss the companies in this story.
The key insight: Anthropic’s first-half revenue was about 63 times Zhipu and MiniMax combined, by our arithmetic; the 85.4% margin set against their 26.4% and 17.9% is calculated from Anthropic’s expense note, not reported, and rests on a different cost basis.
Anthropic’s S-1/A, filed with the SEC on 2 November, gives quarterly revenue of $4,731 million for the three months to 31 March 2026 and $11,555 million for the three months to 30 June. Together that is $16,286 million, the same six-month figure its first S-1 gave on 26 October. Anthropic reports in US dollars.
Zhipu, which trades in Hong Kong as Z.AI (stock code 2513), reported revenue of RMB953.9 million for the six months to 30 June 2026, up 399.7% year on year. The company itself puts that at “approximately USD142 million”. Its open platform and API business made up 86.5% of revenue.
MiniMax (stock code 0100) reports in US dollars. Its revenue for the half was $116.6 million, up 283.1%, “exceeding our total revenue of US$79.0 million for the full year of 2025.”
Social Capital’s account on X posted the same comparison on 3 November: “Anthropic generated $16.3B in revenue” and the two Chinese labs “reported $259M combined.” That matches the filings, using Zhipu’s own dollar figure. Our coverage of the two Hong Kong listings is in MiniMax, Zhipu Shed $150B From Peaks Before January Unlock.
Revenue for the six months to 30 June 2026. Anthropic from its S-1/A; Zhipu reports RMB953.9 million and gives “approximately USD142 million” itself; MiniMax reports in US dollars. The 63 times is our arithmetic; no other exchange rate is used.
Our Arithmetic: About 63 Times
Anthropic’s $16,286 million divided by the two labs’ combined $258.6 million ($142 million plus $116.6 million) is about 63, by our arithmetic. Zhipu reports in renminbi; we use the dollar figure Zhipu gives itself and apply no other exchange rate.
The second quarter alone carries most of it: Anthropic’s $11,555 million for April to June is about 45 times the two labs’ six-month total, by our arithmetic. We reported the speed of that climb in Anthropic Hit a $17B Quarter in 5 Years; Google Took 16.
What the 85.4% Actually Measures
The 85.4% is revenue minus “cost of revenue”, divided by revenue. Anthropic shows cost of revenue only in a supplemental note that regroups its expenses by function. For the first half of 2026 that note gives $2,380 million against $16,286 million of revenue, which is 85.4% by our arithmetic. We first worked it out on 27 October in Anthropic: Models Under 6 Months Old Made 80%+ of Q3 Revenue.
The S-1/A says that during 2025 Anthropic stopped presenting “cost of revenue, gross profit (loss), research and development, sales and marketing, and general and administrative” as separate line items and moved to expenses grouped by nature. So the filing shows no gross profit line, though IndexBox’s wording implies one.
The note says “Substantially all of cost of revenue consists of compute and infrastructure costs associated with customer-serving workloads.” Compute for model development sits in research and development. Channel partner costs, and the compute used to give free access, sit in sales and marketing.
The S-1/A updates the note to nine months: $4,812 million of cost of revenue on $33,242 million of revenue, or 85.5% by our arithmetic.
Why the Margins Are Not Like for Like
Zhipu’s 26.4% and MiniMax’s 17.9% are the gross profit margins the two companies report, from the gross profit line in their interim results. Zhipu says its cost of sales rose “primarily attributable to the increase in computing service fees”. MiniMax says its cost of sales rose “primarily in line with the increase in revenue.”
As we read the three filings, they agree on one point: training sits below the gross line. Zhipu ties its research spending to “investment in model training”. MiniMax ties its own to “cloud services expenses related to training activities”. Anthropic puts model-development compute in research and development.
Where they differ is the rest. Anthropic’s cost of revenue leaves out channel partner fees and free-tier compute, which it books as sales and marketing. The Hong Kong filings do not split their cost of sales the same way. That is why we compare the margins in direction only, not to the decimal.
A stricter cut uses Anthropic’s own income statement. Its “compute and infrastructure” line, which includes training, was $8,945 million in the half. Revenue minus all of it leaves 45.1% of revenue, by our arithmetic. There is no matching line for Zhipu or MiniMax.
Training Spend Against Revenue
Zhipu’s research and development expenses were RMB2,131.2 million in the half, about 2.2 times its revenue, by our arithmetic. MiniMax’s were $296.9 million, about 2.5 times revenue. Anthropic’s research and development in its functional note was $9,741 million, about 0.6 times revenue.
Zhipu’s loss for the period was RMB2,072.0 million. MiniMax’s was $358.0 million. Anthropic’s operating loss for the half was $693 million; its net loss of $80,748 million was mostly a $72,388 million change in the fair value of convertible notes, derivative liabilities and SAFEs.
IndexBox’s two Chinese margins, 26.4% and 17.9%, are the figures in the companies’ interim results. Its headline says Anthropic’s IPO filing “Shows 85.4% Margin”. The filing shows the inputs, not the margin: 85.4% is a derived figure from Anthropic’s functional expense note, and Anthropic does not call it a gross profitmargin.
IndexBox attributes the comparison to the South China Morning Post. We did not read the SCMP article; the figures here are checked against the filings.
The Business Lever
The business lever is cost to serve. As we read the filings, all three companies keep training below the gross line, so the gross figure tracks what it costs to serve a paying customer. On that reading, Zhipu kept about 26 fen of each renminbi of revenue after cost of sales and MiniMax about 18 cents of each dollar, while Anthropic kept about 85 cents of each dollar before channel fees and free-tier compute, by our arithmetic.
The Structural Read
As we read it, the revenue gap is the cleaner comparison: each figure is the company’s own reported revenue for the same six months, and only Zhipu’s needs a currency step, which Zhipu supplies itself.
As we read the cost notes, the margin gap holds in direction but not in size: Anthropic’s 85.4% leaves out channel partner costs and free-tier compute, which the Hong Kong filings do not separate.
As we read the research lines, Zhipu and MiniMax spent about 2.2 and 2.5 times their revenue on research and development in the half, against about 0.6 times at Anthropic, by our arithmetic.
Anthropic, Form S-1/A, supplemental operating expense note, 2 November 2026
“Substantially all of cost of revenue consists of compute and infrastructure costs associated with customer-serving workloads.”
Three Implications
FOR IPO INVESTORS The S-1/A shows cost of revenue only in a functional note, so any gross margin quoted from it is a calculation; the nine-month version is 85.5%, by our arithmetic.
FOR ZHIPU AND MINIMAX SHAREHOLDERS Zhipu says its open platform and API gross margin rose to 24.6% in the half; MiniMax says its 17.9% was “primarily driven by improving infrastructure efficiency.”
FOR READERS OF CROSS-BORDER COMPARISONS Zhipu reports in renminbi and gives its own dollar figure, while MiniMax and Anthropic report in dollars, so any multiple across the three depends on that conversion.
The framework says: “Inference is a revenue engine: continuous, distributed across millions of applications and agents, incurring marginal cost with every query.”
As we read it, the gross line in all three filings is that marginal cost: what it takes to serve each paying query, with training booked further down. The framework was published in March 2026 and does not discuss Zhipu, MiniMax or Anthropic’s filing.
Anthropic’s 85.4% and the Hong Kong gross margins are not measured on the same basis, so the gap between them is directional. Zhipu does not state the exchange rate behind its “approximately USD142 million”; a different rate would move the 63 times only slightly.
Anthropic’s figures for the half are unaudited, as are both Hong Kong interims. We did not contact Anthropic, Zhipu, MiniMax or IndexBox.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growth story. This week it started being priced as a credit story.
On the companies’ own filings, Anthropic’s first half of 2026 brought in about 63 times the revenue of Zhipu and MiniMax combined, by our arithmetic. The 85.4% margin in circulation is a calculation from Anthropic’s expense note, not a reported gross margin; the Hong Kong labs’ 26.4% and 17.9% are reported ones, and the cost lines behind them are drawn differently.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. Anthropic’s figures are from its Form S-1/A filed on 2 November and its Form S-1 of 26 October on SEC EDGAR. Zhipu’s and MiniMax’s are from their interim results announcements on HKEXnews. IndexBox’s wording is from its article; the Social Capital figures are from its X post. The 63 times, the 85.4% and the other ratios are our arithmetic. We use Zhipu’s own US-dollar figure and no other exchange rate. We did not contact the companies.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
