Enflame Technology: What Makes It Different?

It closed at 397 yuan, up 179%, with a market capitalization of 170.8 billion yuan. Enflame Technology, which landed on the STAR Market on September 11, brought the capitalization of the "four little dragons" of domestic GPUs to a close — Moore Threads, MetaX, Biren Technology and Enflame Technology (hereinafter referred to as "Enflame") have all made it across within a single year.

After the excitement comes the cold question. All four go by the name of AI chip company, so what makes Enflame Technology different?

Rejecting CUDA, Claiming Architectural Sovereignty

Domestic AI chips face an unavoidable judgment call: follow Nvidia, or not?

Most choose to follow. The GPGPU path, CUDA compatibility, software-stack alignment — the logic is also hard to argue with: millions of developers around the world earn their living on CUDA, and the lower the migration cost, the easier a deal is to sign.

Enflame Technology, however, chose not to: it adopts a DSA dedicated architecture and has developed its own "Yusuan TopsRider" software stack in-house. It has so far rolled out five cloud AI chips across four generations of architecture, and has built a rich product matrix centered on its self-developed chips, as shown in the figure below.

At present, China's homegrown cloud AI chip vendors include DSA architecture vendors represented by Huawei HiSilicon, Kunlunxin, Cambricon and Enflame, and GPGPU architecture vendors represented by Moore Threads, MetaX, Iluvatar CoreX and Biren Technology.

It is worth pointing out that on deterministic workloads, a dedicated architecture delivers better performance-per-watt than a general-purpose yet redundant GPGPU. As AI inference shifts from competing on peak computing power to competing on unit cost, this advantage will become increasingly valuable. The more practical layer is this: not depending on CUDA means not being bound by Nvidia's ecosystem rules.

Based on model forecasts from Goldman Sachs Global Investment Research, the shipment share of non-GPGPU architecture (i.e., DSA architecture) chips among the AI chips in AI servers will show a clear upward trend, and is expected to grow gradually from 36% in 2024 to 45% in 2027.

Tencent, the One You Cannot Get Around

Tencent has a dual identity with respect to Enflame: largest shareholder and largest customer.

After Enflame's listing, Tencent Technology (Shanghai) Co., Ltd. holds 17.95% of its equity, making it the company's largest shareholder.

In August 2018, Tencent led Enflame's 340 million yuan Pre-A round, and has kept adding to its position in almost every subsequent financing round, accompanying the company for eight years; in the strategic placement for this IPO, Tencent-affiliated Shanghai Qishan was allocated 1.7471 million shares, worth about 248 million yuan — the largest allocation and the longest lock-up period among external strategic investors. More critically, the two sides have upgraded from buying and selling to joint R&D: from Hy3 to Hy4, the Hunyuan large model has engaged in deep "chip-model synergy" with Enflame's chips at the design stage, aligning precision layer by layer from the software stack down to the operator library, with the model architecture even designed in reverse to fit the chip's characteristics. This kind of Co-Design has welded the technology roadmaps of the two companies together.

The numbers on the customer dimension are even more striking. From 2023 to 2025, Tencent-related sales as a share of revenue went from 33.34% to 37.77%, and then to 83.79%. In 2025, more than 80% of revenue came from a single customer, a concentration level that is the highest among the four little dragons.

While bringing Enflame direct revenue, Tencent also brings a large-scale validation ground that peers cannot obtain. Having a chip proven in the real business of a national-scale internet company is more persuasive than any test report. This is also the precondition for the revenue explosion: 1.120 billion yuan in the first half of 2026, up 279% year on year, already exceeding the full year of 2025; the company expects 2.3 billion to 3.0 billion yuan for the first three quarters, up 325.78% to 455.36% year on year. For a company founded eight years ago that has posted losses for five consecutive years, this proving ground and showcase is a scarce resource.