Learning Objectives
- Understand what Enflame CloudBlazer is and where it sits in the AI accelerator landscape
- Explain the role of Enflame's in-house DTU architecture and its training versus inference lines
- Evaluate Enflame's strategic position as a Tencent-backed domestic alternative to NVIDIA in China
⚠️Warning
Niche availability — read this first. Enflame is a Chinese chipmaker, and CloudBlazer accelerators are sold primarily to domestic Chinese customers; they are not generally available to buyers outside China and have little independent benchmark coverage in Western markets. Many specifications below come from Enflame's own materials and partner announcements. Treat this page as an orientation to a strategically important but regionally scoped product, not a hands-on review.
What Is Enflame CloudBlazer?
Enflame CloudBlazer (sold in China under the name Yunsui) is a family of data-center accelerators for artificial intelligence, made by Enflame — a Shanghai semiconductor company founded in 2018 and known in China as Suiyuan Technology. CloudBlazer cards are Enflame's bid to give Chinese data centers a homegrown alternative to NVIDIA's GPUs for the heavy compute behind modern AI.
The lineup splits into two jobs. Training accelerators are built to crunch through the enormous datasets used to build models from scratch, while inference accelerators are tuned to run already-trained models efficiently in production. Both are powered by Enflame's own chip architecture rather than a licensed design, which is central to the company's pitch: a fully domestic stack at a time when US export controls limit China's access to the most advanced foreign hardware.
💡Key Concept
DTU — the Deep Thinking Unit. Enflame's accelerators are built around a custom processor it calls the DTU, or Deep Thinking Unit. Like NVIDIA's GPUs or Google's TPUs, the DTU is a parallel processor designed specifically for the matrix math at the heart of neural networks. Enflame designs the DTU in-house and pairs it with high-bandwidth memory and a high-speed interconnect so multiple cards can work together on a single large model.
Training and Inference Lines
Enflame's first CloudBlazer training accelerator, the T-series, paired the DTU with high-bandwidth memory for large-scale cluster training in data centers. The company later added an inference-focused i-series — for example the CloudBlazer i20, built on a newer generation of the DTU core — aimed at serving trained models at lower cost per query.
The split mirrors how the wider industry now thinks about AI silicon: training is a compute-heavy, build-the-model job, while inference is a high-volume, run-the-model job where cost and efficiency matter most.
| Line | Primary job | Example |
|---|---|---|
| CloudBlazer T-series | Large-scale model training in data centers | First-generation DTU training cards |
| CloudBlazer i-series | Serving trained models in production | CloudBlazer i20 (newer DTU core) |
Availability and the Tencent Relationship
Enflame's story is inseparable from Tencent. The internet giant is Enflame's largest shareholder, holding roughly 20 percent of the company, and is also its dominant customer, accounting for the large majority of recent revenue. That relationship guarantees demand but concentrates risk in a single buyer.
In mid-2026 Enflame won approval for an initial public offering on Shanghai's STAR Market — the Shanghai Stock Exchange's technology-focused board — where it plans to raise about 6 billion yuan, roughly $888 million, to fund its next chip generations. It is widely described as the last of China's so-called "four little dragons" of AI chips to head toward public markets, alongside peers racing to supply domestic accelerators as Chinese buyers seek alternatives to NVIDIA.
Pricing
- Sold to data-center and cloud customers
- Primarily in China
- Pricing not public
- Server makers and system integrators
- Volume and configuration based
- Contact Enflame
Like other data-center accelerators, CloudBlazer is not sold at a public list price. It reaches customers through enterprise agreements and server-maker partners, with terms set by volume and configuration. Enflame's positioning emphasizes domestic supply and total cost of ownership rather than headline performance against NVIDIA's flagships.
Strengths
- Fully domestic stack: Enflame designs the DTU in-house, giving Chinese customers a homegrown alternative as export controls restrict access to advanced foreign hardware
- Strong strategic backer: Tencent's ownership stake and large purchasing commitment provide both capital and guaranteed demand
- Training and inference coverage: Separate T-series and i-series lines address both ends of the AI workload spectrum
- Policy tailwind: China's national push for semiconductor self-sufficiency directs both demand and capital toward domestic designers like Enflame
- Public-market validation: A STAR Market listing brings fresh funding and visibility for the next chip generations
Limitations & Considerations
- Regionally scoped: CloudBlazer is aimed at the Chinese market and is not generally available to buyers elsewhere, with limited independent benchmark coverage abroad
- Small share against a giant: Enflame holds only a low-single-digit share of China's domestic accelerator market, where NVIDIA still dominates
- Heavy customer concentration: Reliance on Tencent for the bulk of revenue is a structural risk if that relationship changes
- Unprofitable so far: The company has accumulated significant losses while investing in research and development, typical of a young chip designer
- Software-ecosystem gap: Accelerators live or die on framework and tooling support; matching the maturity of NVIDIA's software stack is a long-term challenge
Best Use Cases
| Scenario | Why CloudBlazer |
|---|---|
| Chinese data centers seeking domestic supply | Homegrown accelerators that sidestep foreign export controls |
| Large-scale model training in China | CloudBlazer T-series targets cluster training workloads |
| Cost-sensitive inference at scale | CloudBlazer i-series focuses on serving models efficiently |
| Tencent-aligned cloud and AI workloads | Tight integration with Enflame's largest customer and shareholder |
When to consider alternatives:
- Buyers outside China, or anyone needing the deepest software ecosystem today, will find NVIDIA's data-center GPUs and CUDA stack the default choice
- Workloads that demand the highest verified training throughput should look to established high-bandwidth-memory accelerators with public benchmarks
Related Tools
Enflame sits alongside other AI-infrastructure and accelerator products in the directory. Readers comparing options may also want to look at NVIDIA Vera CPU, Groq Cloud, Cerebras Inference, and Intel Crescent Island — each taking a different approach to the cost, memory, and throughput trade-offs of AI compute.
Key Takeaways
- Enflame CloudBlazer is a family of data-center AI accelerators from Shanghai chipmaker Enflame, built on its in-house DTU (Deep Thinking Unit) architecture
- The lineup spans a training T-series and an inference i-series, positioned as a domestic Chinese alternative to NVIDIA's data-center GPUs
- Tencent is both Enflame's largest shareholder (about 20 percent) and its dominant customer, supplying the bulk of revenue
- In mid-2026 Enflame won approval for a Shanghai STAR Market IPO to raise about 6 billion yuan (roughly $888 million), the last of China's "four little dragons" of AI chips to go public
- Enflame remains a small, regionally scoped player holding a low-single-digit domestic share — best understood as a strategic bet on homegrown AI silicon rather than a near-term challenger to NVIDIA
