For four and a half years, SenseTime (商汤) reported a loss every single half-year since its Hong Kong listing. Then, on an August evening, the company published an interim result showing a profit of 617 million yuan.
The number was real. But the profit did not come from selling AI services. It came, mostly, from the rising paper value of companies SenseTime had invested in — and from selling a subsidiary.
Key takeaways
- First-half 2026 revenue was 2.911 billion yuan (≈ US$410M), up 23.4% year on year; gross margin improved 2.9 percentage points to 41.4%.
- IFRS profit was 617 million yuan (≈ US$87M), against a loss of 1.489 billion yuan in the same period a year earlier.
- "Other gains, net" reached 2.127 billion yuan (≈ US$300M), including 1.249 billion yuan of unrealised fair-value gains on financial assets and 521 million yuan from disposal of subsidiaries.
- On a non-IFRS basis, the adjusted net loss was still 386 million yuan (≈ US$54M), though narrowed 67.3%.
- The company disclosed "recurring revenue" (经常性收入) for the first time: 1.145 billion yuan, 39.3% of total revenue, up from 21.6% a year earlier.
- Trade receivables stood at 7.039 billion yuan (≈ US$991M) gross at the end of June, with 3.277 billion yuan of impairment provisions already booked.
The headline number, and the footnote under it
SenseTime's interim result contains two profitability measures that point in opposite directions, and both are legitimate.
Under IFRS, profit for the period was 617 million yuan (≈ US$87M). Under the company's non-IFRS definition, which strips out share-based compensation, fair-value movements and certain one-off items, the adjusted net loss was 386 million yuan (≈ US$54M). The gap between the two is roughly 1.0 billion yuan (≈ US$141M).
The bridge is a line called "other gains, net", which jumped from 641 million yuan in the first half of 2025 to 2.127 billion yuan (≈ US$300M) in the first half of 2026. Two components dominate: 1.249 billion yuan of unrealised fair-value gains on financial assets measured at fair value through profit or loss, net of amounts attributable to limited partners in its investment funds; and 521 million yuan of gains on disposal of subsidiaries.
This is a well-understood mechanism. SenseTime has spent years investing in an ecosystem of AI companies — robotics, optical quantum computing, AI infrastructure — and those holdings were marked up as AI valuations rose. That produces accounting profit without producing cash. If sentiment turns, the same line moves the other way.
Chairman and CEO Xu Li (徐立) framed the period around a different set of ideas: what the company calls a "one model, one token factory, one agent control system" (一套模型、一座Token工厂、一套智能体管控系统) architecture, intended to lower the unit cost of tokens and move from selling model access to delivering completed tasks.
Where the operating business actually improved
Strip out the investment gains and the underlying business did get better, in three measurable ways.
Revenue mix. Generative AI (生成式AI) revenue was 2.327 billion yuan (≈ US$328M), up 28.2% and equal to 79.9% of group revenue. Legacy visual AI (视觉AI) contributed 497 million yuan, up 13.9%, or 17.1%. Overseas revenue grew 127%, far faster than the group. A company that was once a computer-vision project shop is now mostly a model-and-compute business.
Margin. Gross profit rose 32.9% to 1.206 billion yuan, with gross margin up 2.9 points to 41.4%. Cost of sales grew 17.5% to 1.705 billion yuan — slower than revenue. The pressure sits inside that line: intelligent computing center (智算中心) operating costs rose from 79.8 million yuan to 233 million yuan, and depreciation and amortisation from 87.5 million yuan to 297 million yuan. Selling AI is becoming an infrastructure business, with fixed costs that arrive before utilisation does.
Cost discipline. R&D spending fell 17.1% to 1.756 billion yuan; selling expenses dropped 0.5% and administrative expenses 13.2%. A company cutting R&D while growing revenue 23% is making a deliberate choice about where efficiency comes from.
The disclosure that matters more: recurring revenue
The most genuinely new number in the filing is one the company had never published before: recurring revenue (经常性收入) of 1.145 billion yuan, up 124.4%, representing 39.3% of total revenue versus 21.6% a year earlier.
That metric is an answer to the criticism that has followed Chinese AI companies since 2023 — that their revenue is project-based, non-repeating, and effectively consulting work with a model attached. A jump from roughly one-fifth to roughly two-fifths of revenue being recurring suggests a shift toward subscription-style access and system-level integration.
It is also a self-defined metric. Companies choose what counts as recurring, and "recurring" is not the same as "contracted". Readers should treat 39.3% as a management disclosure to be tracked over several periods, not as an audited standard.
The balance-sheet question nobody can ignore
Operating cash flow was negative 764 million yuan for the half, wider than the negative 639 million yuan a year earlier, which the company attributes to the timing of customer collections and a higher trade receivables balance.
Trade receivables totalled 7.039 billion yuan (≈ US$991M) at the end of June, against which 3.277 billion yuan of impairment provisions had been taken, leaving a net balance of 3.763 billion yuan (≈ US$530M). More than 3.2 billion yuan of the gross balance has been outstanding for over three years. For a company with 2.9 billion yuan of half-year revenue, the receivables book is the largest single item on the credit-risk map.
On the other side of the ledger, cash and cash equivalents were 9.877 billion yuan (≈ US$1.39B) plus 3.124 billion yuan (≈ US$440M) in time deposits, following two share placements during the period. The cash position buys time; the receivables determine whether it is enough.
According to Li An, chief scientist at BrainNet (脑机网), the shift toward recurring-revenue disclosure across Chinese AI filings reflects investors moving from valuing model capability to valuing revenue durability.
Honest limitations
- "Recurring revenue" is a company-defined metric with no universal definition; the 39.3% figure cannot be compared directly with another company's version of the term.
- This article does not reproduce the full non-IFRS reconciliation. The roughly 1.0 billion yuan gap includes items beyond the two disclosed components.
- The fair-value gains are unrealised and reversible; treating them as earnings power would be a mistake in either direction.
- Operating figures for Chinese AI peers are not standardised, so no cross-company ranking is attempted here.
- The interim results are unaudited. Full-year figures and the annual report's segment detail may differ.
What readers can do now
- When you see "first profit" in a Chinese AI company's headline, locate the non-IFRS or adjusted figure in the same release. If the two disagree, the operating story is in the adjusted line.
- Track recurring revenue and trade receivables together. Recurring revenue predicts future cash; receivables tell you how much past revenue is still unpaid. One without the other is half a picture.
- For anyone buying AI services rather than shares, ask whether your contract is project-based or subscription-based. That single question determines whether your vendor's "recurring revenue" claim includes you.