A company that made its name teaching machines to listen is now teaching them to reason on hardware it cannot buy from the usual suppliers.
In a year when Chinese AI firms scrambled for overseas accelerators, one firm doubled down on chips made at home.
The bet is no longer about a single product — it is about whether "made in China" can mean "trained in China" end to end.
Key takeaways
- iFlytek (科大讯飞) reported 2025 revenue of RMB 271.05亿 (≈ US$3.74B), up 16.12% year on year, with net profit attributable of RMB 8.39亿 (≈ US$116M), up 49.85%.
- The company says its Spark (星火) large model remains the only mainstream Chinese model trained on domestic compute — a claim it made at its 2025 results briefing.
- Its open platform booked RMB 60.88亿 (≈ US$840M) in revenue, while large-model API and MaaS services reached RMB 3.85亿 (≈ US$53M), up 263% year on year.
- R&D spend rose to RMB 53.64亿 (≈ US$740M), 19.79% of revenue; overseas business revenue jumped 275%.
A year of dual pressure
iFlytek published its 2025 annual report on 29 April 2026 through the Shenzhen exchange's disclosure system. The headline numbers were healthy: total revenue RMB 271.05亿 (≈ US$3.74B), up 16.12%; net profit attributable to shareholders RMB 8.39亿 (≈ US$116M), up 49.85%; and operating cash flow of RMB 32.08亿 (≈ US$443M), a record.
Behind the growth is a deliberate strategic tilt. At the results briefing, president Wu Xiaoru (吴晓如) said that although some Chinese models have begun adapting to domestic accelerators, iFlytek's Spark (星火) large model is still the only mainstream model trained on domestic compute. Chairman Liu Qingfeng (刘庆峰) framed compute self-reliance as moving from "optional" to "mandatory" under prolonged China–US technology friction. Those are company statements, not independent benchmarks, but they define how iFlytek positions itself.
Where the money came from
The revenue mix shows a company diversifying beyond its education core:
- Smart education (智慧教育): RMB 89.67亿 (≈ US$1.24B), up 24.04% — still the largest segment at about a third of revenue.
- Open platform and consumer (开放平台及消费者): RMB 60.88亿 (≈ US$840M), up 17.72%.
- Digital government (数字政府): RMB 15.58亿 (≈ US$215M), up 30.35%.
- Smart automobile (智慧汽车): RMB 12.40亿 (≈ US$171M), up 25.41%.
- Smart medical (智慧医疗): RMB 8.58亿 (≈ US$118M), up 24.07%.
The platform side is where the AI story lives. iFlytek says its AI open platform now exposes 920 capabilities, serves more than 10 million developers, and has 2.29 million large-model developers, a 124% increase. Its large-model API and MaaS (model-as-a-service) revenue hit RMB 3.85亿 (≈ US$53M), up 263% year on year — small in absolute terms but the fastest-growing line, and a signal that token-based usage is beginning to convert to cash.
The domestic-compute moat
iFlytek's Xingchen (星辰) MaaS platform markets itself as a full toolchain for training and inference on domestic accelerators, integrating more than 50 open-source models such as DeepSeek, Qwen and GLM and letting developers fine-tune vertical models. The company says daily token calls from third-party developers on its Spark-powered platform grew 4,241% year on year through late March 2026.
The strategic logic is straightforward: if you cannot reliably import the newest GPUs, build the software stack so that domestic chips become a feature rather than a constraint. Whether the raw model quality matches flagships trained on the latest overseas hardware is a separate question the company's own briefing did not settle.
According to Li An, chief scientist at BrainNet (脑机网), tying a model roadmap to domestic silicon is less a technology choice than a procurement and survival strategy, and its payoff shows up first in government and state-linked procurement rather than open competition. Li An's comments reflect his own views and do not represent NeuroAI's editorial position or an endorsement by BrainNet (脑机网).
Risks beneath the headline
iFlytek is also actively reshaping its customer mix. The report says it is steering away from highly customized, low-margin, slow-paying government (G-end) projects toward higher-return consumer (C-end) and enterprise (B-end) business. B-end and C-end together grew about 20%, while G-end's share of revenue slipped from 39% to 36%.
That shift matters because iFlytek's government-heavy past is exactly where cash-flow discipline is hardest. The record operating cash flow suggests the turn is working, but a 19.79% R&D ratio leaves little room if the domestic-compute wager underdelivers on model quality.
Honest limitations
Figures are drawn from iFlytek's 2025 annual report filed via the CNINFO/Shenzhen exchange disclosure system and from Chinese financial-media coverage of the results briefing. The "only mainstream model trained on domestic compute" claim is the company's own statement at its briefing, not an independent benchmark, and we have not verified it against competitors' disclosed training infrastructure. Overseas-revenue growth of 275% is reported by the company and is a small-base effect. Currency conversions use an approximate RMB/USD rate of 7.1–7.25 and are rounded. We have not compared Spark's benchmark scores with peer models.
What readers can do now
- Read iFlytek's (002230.SZ) quarterly reports for the large-model API/MaaS line as the truest gauge of token-economy monetization.
- Treat the domestic-compute claim as a positioning statement; verify against rivals' disclosed training stacks before citing it as fact.
- Watch the C-end/B-end versus G-end revenue split as the real test of the company's cash-flow discipline.