NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

Phancy Swings to Adjusted Profit as Token Billing Reshapes an Enterprise AI Platform

The former Fourth Paradigm reports ¥7.1 billion of 2025 revenue, an adjusted profit and a three-line business built around models, API calls and agents.

2026-11-22 · 881 words · NeuroAI
Phancy Swings to Adjusted Profit as Token Billing Reshapes an Enterprise AI Platform

An enterprise software company used to sell a licence and invoice once a year. Phancy is moving to bills measured in calls and outcomes. Its 2025 accounts show what that transition looks like in numbers.

Key takeaways

  • Phancy Group Co., Ltd. (范式智能技术集团股份有限公司, 06682.HK), formerly Beijing Fourth Paradigm Technology Co., Ltd. (北京第四范式智能技术股份有限公司), reported 2025 revenue of ¥7.135 billion (≈ US$999M / HK$7.78B), up 35.6% year on year, in a results announcement filed with HKEX on 30 March 2026.
  • Gross profit rose 10.6% to ¥2.483 billion (≈ US$348M / HK$2.71B).
  • On a non-IFRS basis, adjusted net profit was ¥6.33 million (≈ US$0.89M / HK$6.9M) and adjusted profit attributable to owners was ¥17.84 million (≈ US$2.5M / HK$19.4M), against losses in both measures in 2024.
  • Revenue divides into AI Platform at ¥6.552 billion (≈ US$917M / HK$7.15B), Agentic AI at ¥503 million (≈ US$70.4M / HK$548M) and API at ¥80 million (≈ US$11.2M / HK$87M).

What the filing shows

The headline figures are large for a Chinese enterprise AI company, but the shape of the announcement matters more than the total.

Revenue grew 35.6% from ¥5.261 billion (≈ US$736M / HK$5.73B) in 2024. Gross profit grew only 10.6%, from ¥2.245 billion (≈ US$314M / HK$2.45B). The gap means blended gross margin fell: roughly 42.7% in 2024 against 34.8% in 2025.

That is the first thing to notice. A company can add revenue quickly by taking on lower-margin work, and the filing shows exactly that pattern. The announcement ties growth to domestic substitution demand and to the "AI+" action, both demand-side factors rather than pricing gains.

The profit line is a non-IFRS measure. Adjusted net profit of ¥6.33 million (≈ US$0.89M / HK$6.9M) and adjusted profit attributable to owners of ¥17.84 million (≈ US$2.5M / HK$19.4M) exclude items such as share-based compensation and fair-value changes. Against ¥7.135 billion (≈ US$999M / HK$7.78B) of revenue, the adjusted margin is thin — a fraction of a percentage point.

The three business lines

The company now reports three segments:

Segment2025 revenueGrowthShare
AI Platform¥6.552B (≈ US$917M / HK$7.15B)+32%about 92%
Agentic AI¥503M (≈ US$70.4M / HK$548M)+93.2%about 7%
API¥80M (≈ US$11.2M / HK$87M)+129.2%about 1%

AI Platform remains the financial base. It bundles compute-related business with platform software and is where the domestic-substitution theme lands. The company names three products in this line: ModelHub XC, described as a model community for domestic-chip adaptation with more than 30,000 models adapted and a stated target of one million by 2027; HAMi vGPU, a GPU sharing and scheduling tool; and PhanthyCloud, an AI platform-as-a-service layer.

The API line is small but growing fastest. The company describes a pay-as-you-use model in which charges track token consumption rather than seats. Agentic AI sits between the two, where the company is testing a different billing basis.

What "pay for outcome" means

The most distinctive element of the report is the Agentic AI pricing model, described as pay-for-outcome (按结果付费).

The stated application is electricity trading. The company describes a full-chain service — prediction, decision, risk control and review — for spot and medium-to-long-term power trading, with pilots in several electricity provinces and coverage of wind, photovoltaic and storage assets. It says the approach is being extended to manufacturing and finance.

There is a simple economic logic here. A power trader loses money on a wrong forecast, so an AI system that improves the forecast can be paid against the result. That is an easier sale than a generic assistant, but it also means the vendor carries measurement risk: the outcome must be attributable and the baseline agreed.

The company also says it is pursuing internationalisation and a consumer market in 2026. Those are stated directions, not reported revenue lines.

The margin question

Two numbers pull in different directions. Adjusted profit is positive, which is a genuine change from 2024. Gross margin is lower, which means the additional revenue arrived at a lower quality.

For a platform business, the interesting question is whether the mix reverses. API and Agentic AI together are under 9% of revenue but grew far faster than the core. If those lines scale while keeping higher margins, the blended figure could improve. If growth continues to come mainly from hardware-linked platform work, margin will stay compressed.

The filing does not break out margin by segment, so this cannot be settled from the announcement alone. It is the right question to ask at the next reporting date.

Honest limitations

All financial figures come from the HKEX annual results announcement and the company's own news release. I have not examined the full notes to the accounts, so the exact composition of the non-IFRS adjustments is described at the level the announcement uses. Segment shares are calculated from the company's reported figures; the company's target of one million adapted models by 2027 is a stated goal, not a result.

What readers can do now

  • Read non-IFRS profit alongside gross margin. A positive adjusted result can coexist with a lower-quality revenue mix.
  • Track the API and Agentic AI share over the next two reporting periods; that is where the pricing experiment is measurable.
  • If you buy enterprise AI, ask whether the contract bills per seat, per call or per outcome — the three models create very different incentives.

Related coverage

Cite this article

NeuroAI (脑机网), "Phancy Swings to Adjusted Profit as Token Billing Reshapes an Enterprise AI Platform," 22 November 2026, https://neuroai.site/a/na-corp-phancy-2025-results

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Companies & Stack” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories