Inside Enflame's Eight-Year, $5 Billion Bid to Build a Chinese AI Stack from Scratch

Enflame has spent eight years and roughly 36.76 billion yuan (about $5 billion) in cumulative R&D building a custom AI compute stack, betting the next decade's binding constraint is cluster software and developer gravity rather than single-chip speed. The bet is now public because the company cleared the STAR Market listing committee in April 2026, with a 19.95% stake held by its largest customer, Tencent.

The architectural choice is the story. Most Chinese AI accelerator startups, including Moore Threads, MetaX, Biren, Hygon, and Cambricon, have chosen some form of compatibility with Nvidia's GPU model and, where they can, with CUDA, the proprietary software platform that has become the de facto operating system of AI training. CUDA is the developer gravity. Tools, models, debugging, profiling, and the habits of every ML engineer trained since 2014 are baked into it. A chip that does not speak CUDA does not inherit that gravity; it has to build its own. Enflame's founding bet, articulated in its 2018 roadmap and re-stated in the IPO prospectus, is that this is worth doing because the binding constraint is no longer how fast one chip runs a model but how a 10,000-card cluster behaves under years of real production load.

The company calls its approach a custom instruction set and a domain-specific architecture. Concretely, that means the GCU-CARE compute unit, now in its fourth generation with native FP8 support, paired with a high-speed chip-to-chip interconnect called GCU-LARE, and a software stack named TopsRider that covers drivers, compilers, operator libraries, and deep learning framework integration. According to the Leiphone feature that drew on the prospectus, nearly 1,000 AI models have been adapted to the platform, across more than 300 application scenarios spanning both internet and non-internet workloads. The company also claims that thousand-card and ten-thousand-card intelligent-computing projects have moved from validation to commercialization. These are company-sourced figures pending independent confirmation; they describe what Enflame says it can do, not what has been independently benchmarked.

Tencent holds 19.95% of Enflame, the largest single stake, and is simultaneously the company's largest customer, a fact noted in both the prospectus and the New Beijing News IPO observation. The relationship began in 2019 with a single-scenario small-scale validation, expanded to multi-scenario large-scale validation, and is now, by Enflame's account, a routine batch deployment. The chips underpin what the company describes as "national-level applications" inside Tencent's AI business, a phrase the Leiphone article does not tie to a specific product. Tiger Brokers' English coverage of the IPO treats the same arrangement as anchor-investor-plus-anchor-customer concentration. Either reading carries the same structural implication: the bet is being underwritten, validated, and risk-priced by a single counterparty.

The headcount tells the second half. Of Enflame's 838 employees at the end of 2025, 643, or about 77%, are in research and development. That ratio is consistent with a company whose product is not silicon alone but the full layer above silicon, and it is also the reason the cumulative R&D burn over 2023 through 2025 reached the equivalent of roughly $5 billion. The prospectus does not yet disclose net profitability, and the New Beijing News report treats the IPO as a question of whether the burn can be turned into recurring cluster revenue before capital runs thin.

For non-beat readers, the relevant question is not whether Enflame can out-ship Nvidia, an Nvidia-shaped contest the company is not contesting. The question is whether a custom-ISA, non-CUDA stack can accumulate enough developer mindshare and cluster operating experience to be the second-best option inside China when domestic AI compute procurement moves from single-chip substitution to multi-year cluster deployment. The bet says yes, and the bet's price is now on the public record. What the prospectus still does not show is whether the validator customer, on its own, can supply the developer gravity a CUDA-compatible stack would inherit by default. That is the figure worth watching in the next filing.