HONG KONG — An analyst keeps two windows open: a term sheet for a company that has not listed, and a results announcement for one that has.
The documents describe businesses of similar age, similar revenue and similar ambition. They do not describe similar prices.
That gap is now the most informative thing in Chinese AI finance.
Key takeaways
- Zhipu reported RMB 954 million (≈US$134m) of revenue for H1 2026, up 399.7%, and an adjusted net loss of RMB 1.964 billion (interim results, 31 August 2026).
- API and open-platform revenue reached RMB 825 million, 86.5% of Zhipu's total; on-premises deployment fell from 84.8% to 13.5% of revenue.
- Zhipu's blended gross margin fell from 50% to 26.4% as the mix shifted toward metered interface sales.
- Moonshot closed a private round at about US$50 billion and targets a Hong Kong listing in Q1 2027, raising up to US$5 billion (Bloomberg, 6 October 2026).
- Zhipu's market value was about HK$279 billion on 14 October 2026, against more than HK$1 trillion in June (Shanghai Securities News).
Two markets, two methods
Start with the public side. Zhipu's shares were worth about HK$279 billion on 14 October 2026, down from a peak above HK$1 trillion in June (Shanghai Securities News, 14 October 2026). MiniMax, listed a day later in January, was worth about HK$71.3 billion. Both numbers are set every trading day by people who can read the accounts.
Now the private side. Bloomberg reported on 6 October 2026 that Moonshot closed its final pre-IPO round at about US$50 billion, up from US$31.5 billion in a round earlier in the summer, and is preparing a Hong Kong listing in the first quarter of 2027 that could raise up to US$5 billion. Reuters reported on 3 September that the company had filed confidentially. DeepSeek, which has never published accounts, is raising more than RMB 80 billion (≈US$11.3bn) in a round that opened at a RMB 500 billion (≈US$70bn) valuation, and has hired CITIC Securities for a possible Shanghai STAR Market listing (Reuters and Bloomberg, October 2026).
Neither private figure is a price: each is the post-money mark of the last round, set by investors who cannot sell and cannot see a daily quote.
What the filings actually show
Zhipu's interim report, published on 31 August 2026, is the more instructive of the two because the company changed shape inside six months.
- Revenue of RMB 954 million (≈US$134m), up 399.7% and above the whole of 2025.
- Open platform and API revenue of RMB 825 million, up more than 27 times, or 86.5% of the total.
- On-premises deployment (私有化部署) falling from 84.8% to 13.5% of revenue.
- Gross margin of 26.4%, down from 50%; the API line alone moved from −0.4% to 24.6%.
- R&D of RMB 2.13 billion (≈US$300m), up 33.6%, and an adjusted net loss of RMB 1.964 billion.
MiniMax, the other listed lab, reported US$117 million of revenue for the same half, up 283.1%, a gross margin of 17.9% against 12.1% a year earlier, and an adjusted net loss of US$293 million (company results, August 2026).
The mechanism behind Zhipu's margin is worth stating plainly. A private deployment is a delivery: hardware, integration and licence booked once, at a margin that reflects the work. An API call is metered — priced per token (词元), costed against compute bought in advance. Moving from one to the other trades a lumpy, higher-margin line for a recurring, lower-margin one. Zhipu's blended margin, now 26.4%, sits close to the 24.6% it earns on API sales, which is what a business looks like when four-fifths of it is metered.
Why the private price is different
Moonshot's US$50 billion against roughly US$1 billion of annualised revenue, the run rate Bloomberg reported, is about 50 times sales. DeepSeek's RMB 500 billion opening valuation against a comparable run rate is higher still. Both are far above what the two listed labs command relative to their disclosed revenue.
Three things explain the distance, and none of them is fraud:
- Scarcity. There are very few allocatable positions in Chinese frontier labs, and pre-IPO demand is not rationed by price discovery.
- Optionality. A private mark prices the listing, not the current P&L. If the listing slips, the mark does not adjust that day.
- Disclosure. A public price incorporates a quarterly loss; a private one incorporates a pitch.
What to check before quoting a valuation
- Revenue mix. Metered revenue compounds; project revenue does not. Ask for the split before comparing two labs.
- Gross margin by line. A blended margin hides which business is being sold.
- R&D against revenue. Zhipu spent more than twice its revenue on R&D in the half.
- The definition of ARR. Annualised run-rate is not booked revenue, and a price increase lifts it without new customers.
- The label on the number. Last-round post-money and market capitalisation are different instruments and belong in different sentences, with dates.
Honest limitations
Interim results are company-prepared and should be read with their review notes; they are not the same artefact as an annual audit. The Moonshot and DeepSeek figures come from unnamed sources in Bloomberg and Reuters reports, not from filings, and neither company has confirmed them publicly. Annualised revenue is a run rate, not a reported figure, and it moves with price changes as well as usage. Market values cited here are from 14 October 2026 and are already stale. Nothing in this article is investment advice, and this column does not cover individual shares as recommendations.
What readers can do now
- Analysts: rebuild the revenue mix from the two Hong Kong interim reports before comparing them with US labs, and note that one sells projects and the other sells interfaces at different margins.
- Buyers: ask any Chinese model vendor whether the contract is metered API or a private deployment. The first is priced against compute costs that are rising; the second is priced against a delivery.
- Newsrooms: label every valuation as last-round post-money or market capitalisation, add the date, and never put the two in the same sentence. The gap is the story, not a contradiction.