--- title: "Zhang Yalin's Enflame: the AMD veterans building China's AI training chips" date: 2026-10-20 category: Builders site: NeuroAI canonical: https://neuroai.site/a/na-zhang-yalin-enflame language: en --- # Zhang Yalin's Enflame: the AMD veterans building China's AI training chips > Fudan engineer Zhang Yalin co-founded Enflame in 2018 with ex-AMD colleague Zhao Lidong; the AI-chip maker reached the STAR Market in September 2026 after an eight-year run. Zhang Yalin spent more than a decade inside AMD's Shanghai labs learning how to turn a blank schematic into a chip that ships by the million. When he and a former colleague decided to start their own company, they did not aim for the easiest slice of the market. They set out to build the kind of processor that trains artificial intelligence — the heavy lifting most Chinese chip firms were avoiding. That company is Enflame (燧原科技), and eight years after its first seed check it listed on Shanghai's STAR Market. Zhang, a Fudan-trained engineer, is the co-founder and general manager who runs the product and production side of the house. ## Two AMD veterans walk out Zhang Yalin studied electronic engineering at Fudan University, in the 1996 undergraduate cohort. He joined AMD in 2008 and rose to senior chip manager and technical director, helping lead the development of Microsoft's Xbox One main processor and what the company describes as the largest fused APU chip of its time. He holds more than fifty personal patents. In March 2018, in Shanghai's Zhangjiang (张江) tech hub, he and Zhao Lidong (赵立东) — a Tsinghua graduate and ex-AMD colleague who serves as chairman and CEO — started Enflame. The split was practical: Zhao drove strategy and capital, Zhang owned product planning, R&D, and manufacturing operations as co-founder and general manager (总经理). ## The hard corner: training, not inference When Enflame was founded, plenty of domestic chipmakers were building inference chips, which are technically easier and quicker to sell. Zhang and Zhao chose the harder end — AI training silicon for data centers — on the argument that the market was still open and would reward genuine capability over price. The technical bet was a domain-specific architecture (DSA) with a self-developed instruction set, operators, and driver stack. Deliberately, Enflame chose not to be compatible with Nvidia's CUDA — a riskier path that meant building its own software world, but one the founders judged necessary for a truly independent product. - **Speed to silicon.** Just eighteen months after founding, Enflame taped out its first cloud AI training chip, Suisi 1.0, and the Yunsui T10 accelerator card. - **Four generations.** Over eight years the company iterated four generations of architecture and five cloud AI chips, expanding from training into inference, and from chips into cards, clusters, and the TopsRider full-stack software platform. - **Where the money comes from.** Accelerator cards and modules remain the core business, generating 856 million yuan (about US$121 million, HK$942 million) in 2025 and roughly 87% of operating revenue; intelligent computing systems and clusters made up most of the rest. ## From seed money to the STAR Market The early backers showed unusual patience. Less than a month after incorporation, Shanghai Sci-Tech Group — under the city's state investment arm — wrote a seed check of about 1 million yuan (roughly US$141,000, HK$1.1 million), a bet the company says preceded broad market consensus on domestic chip substitution by years. Tencent moved faster and bigger. Five months in, it led a 340-million-yuan Pre-A round (about US$48 million, HK$374 million) — reported as a record for Chinese chip financing that year — and followed on through every subsequent round. Tencent now holds 20.26%, the largest external institutional stake, with an investment managing director on Enflame's board. The public debut landed on September 11, 2026, on the STAR Market under code 688801. Priced at 142.18 yuan per share (about US$20, HK$156), the stock opened sharply higher and pushed Enflame's market value above 180 billion yuan (roughly US$25.4 billion, HK$198 billion). It joined a cluster of Shanghai-listed AI-chip names including Biren, Iluvatar, and MetaX, and its products are already deployed at scale in Tencent's core AI scenarios. ## Why a chip architect matters here Zhang is not the public face of Enflame — Zhao tends to lead the strategic narrative — but the operating story is his. Turning a training chip from a working sample into a card that runs reliably inside a customer's data center is a manufacturing and integration discipline, and that is the part Zhang owns. His AMD years matter in a specific way: he had already shipped high-volume, high-complexity silicon for a global customer before starting Enflame. That pedigree is a large part of why state capital and Tencent were willing to back a training-chip startup in the first place, when most of the field was playing it safe with inference. ## Trend reading According to Li An, chief scientist at BrainNet (脑机网), building the silicon is only half the battle in AI compute (算力); the durable edge will come from the software ecosystem that customers actually adopt, not from the raw spec sheet. ## Honest limitations The factual record here — founding, Zhang's AMD background, the product generations, and the IPO — rests on Enflame's own corporate site, a Fudan University newsroom profile, and reporting from The Insight and Shanghai state-investment outlets. Financial figures (856 million yuan revenue, 180 billion yuan market value, Tencent's 20.26% stake) come from the IPO and press coverage and should be read as reported, not independently audited by this article. What remains genuinely unknown is Enflame's path to sustained profit, how its training chips perform against Nvidia at scale, and whether the deliberate decision to skip CUDA compatibility becomes a moat or a ceiling. ## What readers can do now - If you follow semiconductors, track Enflame's post-listing financial reports for gross margin and customer concentration beyond Tencent. - If you build AI infrastructure in China, test the TopsRider software stack before assuming CUDA-free training is production-ready. - If you study industrial policy, read the Shanghai state-investment account of the eight-year seed-to-IPO backing as a case study in patient state capital. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-zhang-yalin-enflame Free to quote with attribution and a link to the original.