--- title: "iFlytek bets on a \"second choice for the world\" built on domestic compute" date: 2026-10-08 category: Companies & Stack site: NeuroAI canonical: https://neuroai.site/a/na-corp-iflytek-spark-global language: en --- # iFlytek bets on a "second choice for the world" built on domestic compute > How iFlytek (科大讯飞) turned a GPU shortage into a global go-to-market for its Spark (星火) large model (大模型). Most Chinese AI labs complain they cannot get enough GPUs. iFlytek (科大讯飞) decided the shortage *was* the strategy. Its answer is a large model (大模型) trained entirely on domestic chips — and a pitch to countries that want an alternative to US stacks. ## The constraint became the product For two years, the loudest story in China's model race was who could buy the most Nvidia silicon. iFlytek took the opposite bet. Chairman Liu Qingfeng put the gap bluntly at the company's February 2026 annual meeting: the firm holds roughly 20,000 domestic compute cards against the hundreds of thousands held by the best-funded rivals — a gap of more than ten times. Instead of treating that as a handicap, iFlytek built its 2025 story around it. Spark (星火) is, by the company's repeated claim, the only mainstream Chinese large model (大模型) whose key versions were trained end-to-end on fully domestic compute (全国产算力). That claim is a company position, echoed across Chinese financial and state media, not an independently benchmarked fact — but it shapes how iFlytek sells. The engineering payoff is concrete. On domestic hardware, the team lifted training efficiency for deep-reasoning models from 30% to over 84%, and for mixture-of-experts (混合专家) models from 30% to 93%, according to figures shared at the same meeting. Those numbers are the moat: a model that runs on local chips is easier to deploy inside banks, hospitals, and government offices that cannot touch export-controlled hardware. ## What "fully domestic compute" actually bought The 2025 full-year results, published on 28 April 2026, show the bet converting into revenue. iFlytek reported: - Revenue of **27.11 billion yuan** (~$3.76 billion / ~HK$29.5 billion), up 16.12% year on year. - Net profit attributable to shareholders of **839 million yuan** (~$116 million / ~HK$912 million), up 49.85%. - Operating cash flow of **3.21 billion yuan** (~$446 million / ~HK$3.49 billion), a record. - R&D spend of **5.36 billion yuan** (~$745 million / ~HK$5.83 billion), 19.79% of revenue. None of this is a moonshot multiple. It is steady, profitable scaling — exactly what a company selling into regulated industries needs to show. In March 2026, iFlytek also won Shenzhen Stock Exchange approval for a **4 billion yuan** (~$556 million / ~HK$4.35 billion) private placement to expand its large-model and compute platform, a signal that the domestic-compute roadmap is funded for years. ## The global wedge: languages, not labs Where iFlytek diverges from most Chinese model makers is overseas. The Spark multilingual large model (大模型) now covers **more than 130 languages**, and the company reports 2025 overseas revenue grew **275%** year on year, crossing the **1 billion yuan** (~$139 million / ~HK$1.09 billion) line for the first time. The go-to-market is not "ship a chatbot." It is vertical and硬件-led: - An ASEAN (东盟) multilingual model base launched for the ten Southeast Asian nations. - The AI office book (办公本) line passed **1 billion yuan** in 2025 revenue, with **68,000 units** sold overseas; the AINOTE 2 won a Guinness record as the world's thinnest e-ink tablet and topped Japan's Makuake crowdfunding charts. - iFlytek became the designated model supplier for the China Pavilion at Expo 2025 Osaka. - A multilingual meeting system signed its first overseas smart-office deal with the Legislative Council of Hong Kong. Liu's framing is deliberate: China should give the world a *"second choice"* beyond US models. In parts of Latin America and ASEAN, he says, Spark already beats US counterparts on specific languages. That is a positioning claim, not a blanket performance claim — and it is the right kind of claim to make when selling sovereignty, not benchmarks. ## Three engines, one moat iFlytek runs three commercialization tracks that reinforce each other: - **Industry applications.** In 2025 the company's large-model project wins were valued at **2.32 billion yuan** (~$322 million / ~HK$2.52 billion) — the top slot in China by both count and value, per third-party bid tracking cited in its earnings materials. - **AI open platform.** Developers on the platform surpassed **10 million**, including **564,000 overseas**; large-model developers reached **2.29 million**. Daily token calls on the platform rose 45-fold year on year. - **Consumer hardware.** Learning devices held the top sales slot across JD.com and Tmall festivals for a third straight year. The moat is the combination: a domestic-compute model that satisfies regulators, vertical data from two decades in education and healthcare, and a hardware channel that already reaches consumers abroad. ## Honest limitations - The "only mainstream model on domestic compute" line is iFlytek's own claim, repeated by Chinese media; treat it as a positioning statement, not an independent audit. - Overseas revenue is growing fast off a small base — **1 billion yuan** is still under 4% of the **27.11 billion yuan** total, so "global" is aspirational, not yet material. - Raw model capability trails the front-runners by what Liu himself calls half a generation (3–6 months); iFlytek is betting on deployment and compliance, not topping leaderboards. - All figures above come from iFlytek's own reports and Chinese financial/state outlets; we found no Western independent benchmark confirming the domestic-compute training-efficiency numbers. ## What readers can do now - If you evaluate AI vendors for regulated or government work in Asia, shortlist iFlytek as the domestic-compute option and pressure-test the "fully domestic" claim with a written architecture disclosure. - Watch the **4 billion yuan** placement and the 2026 Spark X2 iteration as the real test of whether domestic-compute training can keep pace. - Track overseas developer count (now **564,000**) quarter by quarter — it is the cleanest leading indicator of whether the "second choice" pitch is landing outside China. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-corp-iflytek-spark-global Free to quote with attribution and a link to the original.