Enflame
Last reviewed
Jun 3, 2026
Sources
15 citations
Review status
Source-backed
Revision
v1 · 1,732 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
15 citations
Review status
Source-backed
Revision
v1 · 1,732 words
Add missing citations, update stale details, or suggest a clearer explanation.
Enflame Technology (Chinese: 燧原科技, romanized Suiyuan) is a Chinese fabless chip company that designs AI accelerators for data-center training and inference. Founded in March 2018 and based in Shanghai, it is best known for being backed and bought from by Tencent, which is both its largest shareholder and by far its largest customer. Its accelerators are built around an in-house architecture the company calls the GCU (General-purpose Computing Unit), and it sells training cards, inference cards, cluster systems, and a software stack named TopsRider. Enflame is usually grouped with Biren, Moore Threads, and MetaX as one of the Chinese GPU and AI-accelerator startups positioned as challengers to NVIDIA in China, a cohort that Chinese press calls the "four little dragons" of domestic GPUs. [1][2]
Enflame was founded in March 2018 by two former AMD engineers: Zhao Lidong (also rendered Lidong Zhao), who became chief executive and who had led AMD's global GPU product development, and Zhang Yalin (Arthur Zhang), who became chief operating officer. The company is registered in Shanghai. [3][1]
Tencent has backed Enflame since its earliest rounds, alongside the venture firm ZhenFund. Over time the shareholder base widened to include the China Integrated Circuit Industry Investment Fund (the "Big Fund"), the photo-app company Meitu, Primavera Capital, and CITIC-affiliated investors. By the time Enflame filed to go public, Tencent was its single largest shareholder with a stake of roughly 20 percent (the prospectus reports 20.26 percent). [1][3][4]
The Tencent relationship runs deeper than equity. Tencent deploys Enflame's accelerators in its own data centers, and in Enflame's IPO prospectus, sales to Tencent and related parties accounted for 71.84 percent of revenue in the first nine months of 2025. That makes Enflame in effect a captive supplier to its largest shareholder, a concentration the prospectus itself flags as a risk and expects to persist for some time. [4][5]
Enflame's chips are general-purpose accelerators rather than graphics GPUs, and the company markets the underlying design as the GCU, or General-purpose Computing Unit. The compute silicon itself carries the internal name DTU, for Deep Thinking Unit (邃思 in Chinese). A GCU core is organized hierarchically: the first-generation DTU grouped 32 scalable processors into four clusters and combined tensor, vector, and scalar units, with an on-chip interconnect for multi-card scaling. The architecture targets the matrix multiply (GEMM) and convolutional workloads common in deep learning, and it supports mixed-precision math including FP32, FP16, BF16, and INT8. [6][7]
Enflame's first product, announced in December 2019, was the CloudBlazer T10 (Chinese: Yunsui T10) built on the first-generation DTU 1.0 chip, fabricated on GlobalFoundries' 12LP FinFET process with 2.5D packaging and HBM2 memory. The DTU 1.0 used a PCIe Gen4 x16 interface and proprietary chip-to-chip links for clustering. [8][6]
The second generation, branded Suisi 2.0 (DTU 2.0), arrived in 2021 and powered the CloudBlazer T20 card and the T21 OAM module. Enflame's published figures put the second-generation training chip at about 20 teraflops of FP32 and 80 teraflops of FP16 or BF16, with two HBM2 stacks delivering 512 GB/s of bandwidth; the T21 followed the Open Compute Project's OAM form factor. Later in 2021 Enflame split out a dedicated inference line with the CloudBlazer i20, based on the Suixi 2.5 chip with an updated GCU-CARE core, GlobalFoundries 12nm fabrication, and 16 GB of HBM2e rated at up to 819 GB/s. [7][9]
Enflame's third generation centered on the S60 inference accelerator, launched in 2024. By company and press accounts it became the most widely deployed domestic alternative of its class, with more than 70,000 cards shipped, and it anchored what was described as China's first 10,000-card domestic inference cluster, built in Qingyang, Gansu province, under the country's "east data, west compute" program. In July 2025, at the World AI Conference in Shanghai, Enflame unveiled a fourth-generation part, the L600 (邃思L600), positioned as a single chip for both training and inference. The company says the L600 carries 144 GB of on-package memory and 3.6 TB/s of bandwidth and supports FP8. As with most Chinese accelerators, these are vendor figures and have not been independently benchmarked. [10][11]
| Product | Chip | Year | Role | Process | Memory | Notes |
|---|---|---|---|---|---|---|
| CloudBlazer T10 | DTU 1.0 | 2019 | Training | GlobalFoundries 12LP | HBM2 | First product; 32 cores, 2.5D packaging [8] |
| CloudBlazer T20 / T21 | Suisi 2.0 (DTU 2.0) | 2021 | Training | 12nm FinFET | HBM2, 512 GB/s | ~20 TFLOPS FP32, ~80 TFLOPS FP16/BF16; T21 is OAM/OCP [7] |
| CloudBlazer i20 | Suixi 2.5 | 2021 | Inference | GlobalFoundries 12nm | 16 GB HBM2e, 819 GB/s | GCU-CARE core; PyTorch/TensorFlow/ONNX support [9] |
| CloudBlazer S60 | third generation | 2024 | Inference | TSMC N6 | N/A | ~70,000 units shipped; Gansu 10,000-card cluster [10][11] |
| L600 | 邃思L600 | 2025 | Training and inference | N/A | 144 GB, 3.6 TB/s | FP8; vendor specs, announced at WAIC [11] |
Enflame ships a software platform called TopsRider that sits between its hardware and mainstream AI frameworks. It provides drivers, libraries, and a compiler toolchain, and supports TensorFlow, PyTorch, and the ONNX model format, with C++ and Python interfaces. The company has repeatedly credited TopsRider optimization for performance gains across chip generations, and the platform handles virtualization, multi-tenant isolation, and the cluster management used to scale many cards together. A capable software layer is widely regarded as the harder half of competing with NVIDIA, whose CUDA ecosystem is the default for AI developers. [7][9]
Enflame's standing under US export controls is frequently misstated, so the distinction matters. Unlike Biren and Moore Threads, which were added to the US Commerce Department's Entity List in October 2023, and unlike Cambricon, Enflame has not been placed on the Entity List. That status is significant because it left Enflame able, in principle, to keep using foreign foundries. To stay within the rules, Enflame submitted downgraded chip designs to TSMC in late 2023 so that the parts would fall below the thresholds of the advanced-computing export rules introduced that year. [12][1]
The arrangement drew scrutiny in 2025. The analysis firm TechInsights reported that Enflame's S60 accelerator contained silicon manufactured by TSMC on a 6nm-class process (die marking SCORPIO-AO, process node identified as N6NTO-HPC), and questioned whether the chip's capabilities crossed the controlled threshold. TSMC contested that classification, saying the part did not meet the criteria for a controlled AI chip, and TechInsights described its analysis as ongoing. Reporting at the time, citing a senior Commerce Department official, noted that Enflame had not received an export license in 2023, 2024, or 2025. The episode illustrates the gray zone many Chinese designers operate in: not formally blacklisted, but dependent on foundry access that the controls are meant to limit. [13][12]
In August 2024 Enflame disclosed that it intended to list on the Shanghai Stock Exchange's STAR Market and had engaged China International Capital Corporation (CICC) to handle the required pre-IPO tutoring. A pre-IPO funding round in June 2025 valued the company at about RMB 20.5 billion (roughly US$2.8 billion), up sharply from the roughly US$1.65 billion at which Pitchbook had pegged it in September 2023. The Shanghai exchange accepted Enflame's listing application on 22 January 2026, with the company seeking to raise about RMB 6 billion to fund development of its fifth- and sixth-generation chips. The application moved into the inquiry stage on 12 February 2026. [14][5][2]
The prospectus shows a company growing fast and losing money. Revenue rose from RMB 90.1 million in 2022 to RMB 301.2 million in 2023 and RMB 722.4 million in 2024, reaching RMB 540.2 million in the first nine months of 2025. Net losses remained heavy, including about RMB 1.51 billion in 2024 and RMB 887.8 million for the first nine months of 2025, with cumulative losses over three years near RMB 4.29 billion. The filing also discloses steep customer concentration beyond Tencent: the top five customers made up 96.41 percent of revenue in the first nine months of 2025. Enflame's accelerator card and module shipments reached about 38,800 units in 2024, which it estimates at roughly a 1.4 percent share of China's AI accelerator card market. [4][5]
Enflame sits in a crowded field of Chinese accelerator makers that expanded quickly as US controls cut into NVIDIA's China sales. Its closest peers in the "four little dragons" framing are Biren, Moore Threads, and MetaX (Muxi), most of which moved to list on Chinese exchanges in 2025 and early 2026: Moore Threads debuted on the STAR Market in December 2025, and MetaX's application was accepted in mid-2025. The larger competitive picture in China also includes Huawei's Ascend line, which holds a far bigger share of domestic AI compute than any of the startups, and Cambricon, an earlier-listed designer. Enflame's particular profile within this group is its tight coupling to Tencent: that relationship gives it a guaranteed buyer and a path to scale that its rivals lack, while making it unusually exposed to a single customer. [2][15][5]