Story of Enflame's Investors: A Table, 20 Pages of PPT and the Long Halftime

In the winter of 2017, at No. 392 West Jianguo Road, Xuhui District, Shanghai.

This is an old villa hidden behind plane trees, which also serves as the office of DT Capital. The second-floor conference room is not large, with a long table placed in the center. Ye Weigang, Founding and Managing Partner of DT Capital, and his partners sat on one side of the table, while Zhao Lidong and Zhang Yalin sat on the other side.

The long table in the second-floor conference room of DT Capital, Photo source: photographed by *Jiazi Lightyear*

Zhao Lidong was born in 1966, and has worked at S3, Juniper Networks and AMD successively. During his seven years at AMD, he was responsible for multiple product planning and core IP R&D projects, and participated in the establishment of AMD China R&D Center. After returning to China in 2014, Zhao Lidong successively served as President of RDA Microelectronics and Vice President of Unisplendour Group.

Zhang Yalin, 12 years younger than Zhao Lidong, is a senior chip engineer. He joined AMD from Shanghai Qima Digital, and later served as Senior Chip Manager of AMD and Technical Director of China R&D Center. He was once responsible for the R&D and mass production of products such as the main chips of Microsoft Xbox One series, and also participated in the establishment and management of multiple R&D teams of AMD in Shanghai and Beijing.

In November 2017, these two veterans in the chip industry both left their former companies and decided to start their own business to develop domestic cloud-based AI chips.

Before walking into this old villa, they had met many investment institutions, but most of the conversations did not reach the final stage. Some investors even said harsh words: "With just the two of you, you still want to challenge NVIDIA?"

But right here, the situation took a turn for the better — they met DT Capital and Ye Weigang.

Ye Weigang once served as Senior Director of Business Development in the Marketing Department of Cadence, the global EDA (Electronic Design Automation) giant, and is very familiar with chip technology and the market. After returning to China in 2006, he entered the venture capital industry, and chips and semiconductors have always been his most familiar field.

Many years later, Ye Weigang can no longer remember what Zhao Lidong and Zhang Yalin wore that day, only that it was very cold in Shanghai that day. And the two people sitting opposite him also felt that their hearts had sunk to the bottom.

Zhao Lidong and Zhang Yalin had not established a company at that time, with no products and no revenue. What they brought to Ye Weigang was only a PPT of about 20 pages.

Ye Weigang asked Zhao Lidong to open the planned architecture diagram of the first chip in the PPT, and then began to ask questions: how to design the computing core? How to match the storage bandwidth? How to interconnect the chips? Which IPs are developed in-house and which are purchased externally? How to face the software ecosystem that CUDA has already formed? Where do the supply chain and tape-out resources come from?

"I opened their first chip structure diagram, looked at it layer by layer, and asked questions module by module." What Ye Weigang wanted to confirm was how many problems had been clearly thought out under this 20-page PPT.

After meeting that day, Ye Weigang and his partners quickly made a decision to invest in these two veterans in the chip industry.

Zhao Lidong and Zhang Yalin also gained more confidence. In March 2018, the next year after the Spring Festival, they established a new company — Enflame Technology.

Logo of Enflame Technology, Photo source: photographed by *Jiazi Lightyear*

DT Capital also participated in Enflame Technology's seed round financing as a co-lead investor. Over the past eight years, Enflame Technology has gone through a total of 10 rounds of financing: investment institutions such as DT Capital, ZhenFund, Yunhe Capital, Wuyuefeng Venture Capital, Tencent Investment, and Sunshine Ronghui Capital have successively appeared in Enflame Technology's list of shareholders.

Today, Enflame Technology is listed on the Sci-Tech Innovation Board of the Shanghai Stock Exchange, with an opening price of 410 yuan per share, an increase of 188.37% compared with the issue price of 142.18 yuan per share, and a total market value of 176.4 billion yuan.

In the narrative of the capital market, Enflame Technology is often listed alongside Moore Threads, Muxi, and Biren Technology, and is called one of the "Four Little Dragons of Domestic GPUs". However, in the eyes of Enflame Technology's investors, the "Four Little Dragons of Domestic GPUs" label also conceals the characteristics and differences between these companies.

On the eve of Enflame Technology's listing, *Jiazi Lightyear* found several investors who bet on it at different stages, trying to piece together the real trajectory of this company through eight years of technology cycles from their memories.

1. The Hardest Road

The uniqueness of Enflame Technology actually lies in the chip architecture diagram that Ye Weigang saw in 2017.

At that time, the domestic AI chip startup boom had just emerged. Companies such as Cambricon, Horizon Robotics, and DeepPhi Technology successively entered the capital's field of vision, and many entrepreneurs chose to cut in from inference chips or end-side scenarios. In contrast, Zhao Lidong and Zhang Yalin targeted their first product in the cloud AI training field.

This is a market with higher thresholds and larger investment.

Cloud AI training chips need to process massive amounts of data and high-intensity parallel computing, and the computing power of a single chip is only the foundation. Startups also need to solve problems such as storage bandwidth, chip interconnection, software compilation, model adaptation, and large-scale clusters. In front of all latecomers is the software and hardware system formed by NVIDIA after years of accumulation.

Ye Weigang remembers that this was the first time he heard a domestic entrepreneurial team clearly propose to develop cloud-based large-computing-power training chips and directly enter the market occupied by NVIDIA. "It's certainly not easy to believe." he said.

However, Ye Weigang found that Zhao Lidong and Zhang Yalin had broken down the next steps in sufficient detail. They not only drew the chip architecture, but also listed the 20-person core team they planned to form, and thought about a series of product issues: what background is required for each position, and where to find talents; which modules must be self-developed and which IPs can be purchased; who the suppliers are, and whether they have cooperated with them in the past.

Before Enflame appeared, Ye Weigang had looked at many AI chip companies and also invested in end-side chip projects. At that table, what he needed to judge was no longer whether the AI chip had a market, but whether the two people in front of him could really make it, and who would use the chip after it was made?

In March 2018, Enflame Technology was established. Zhao Lidong served as CEO, and Zhang Yalin served as COO. Shortly after its establishment, the company completed the seed round financing, which was invested by institutions such as DT Capital, ZhenFund, Wuyuefeng Venture Capital, Yunhe Capital, and Shanghai Science & Technology Venture Capital.

2. Hot Start

At the beginning of its establishment, Enflame had less than 10 employees in total, and the office was borrowed by Zhao Lidong from a friend, which could be used for free for three months.

At the end of April 2018, Yao Leiwen, Managing Director of Tencent Investment, and Xiao Hongda, Investment Director, contacted Zhao Lidong.

"I can't remember how they found us." Zhao Lidong told *Jiazi Lightyear* in 2020 when recalling the initial contact with Tencent, "The first time was a video conference. I remember we talked for almost two hours, and then they brought people to our office in Shanghai."

In the contact between the two sides, what really moved Zhao Lidong was a sentence Xiao Hongda said at that time: "Tencent can help Enflame achieve a 'hot start'."

The "cold start" dilemma in the chip industry is very typical. A company can complete product definition according to public information and industry trends, spend several years making chips, only to find that the customer's real business is inconsistent with the previous assumption. At this time, adjusting the hardware architecture is often too late.

"Hot start" means that from the product definition and R&D stage, Enflame can access the real business needs of large Internet companies, and understand the requirements of training tasks, model frameworks, and data center environments for chips.

In August 2018, Tencent led a 340 million yuan Pre-A round financing for Enflame Technology. *Jiazi Lightyear* learned that the investment agreement at that time did not take Tencent's procurement as a prerequisite, and the business cooperation between the two sides was not exclusive. Tencent's investment department is responsible for judging the company's value, and the business department still needs to decide whether to adopt the product according to performance, stability and cost.

Zhang Yalin once told *Jiazi Lightyear* that the reason why Enflame Technology chose to "marry" Tencent is also because Tencent Investment's biggest feature is "very tolerant of invested enterprises", will not interfere with micro-management, and give founders a lot of space. Tencent will fully support and help all key decisions of Enflame.

Tencent's "hot start" provided great help for Enflame's initial R&D; and Tencent's relatively restrained relationship with Enflame provided Enflame with a real and strict test field.

For Enflame, Tencent is not only an early shareholder, but also the most important potential customer for its first product. But before the customer actually places an order, Enflame first needs to cross the most dangerous hurdle in chip entrepreneurship: tape-out.

3. Successful Tape-out

In the second half of 2018, Zheng Yan joined Sunshine Ronghui Capital, an insurance-funded private equity fund management company.

His doctoral thesis focused on topics related to chip design for testability (DFT) and fault tolerance, and he has a deep understanding of the technologies applied in the integrated circuit design stage. During that period, Zheng Yan looked at most of the domestic AI chip startup projects, and finally focused his attention on Enflame.

Zheng Yan, who is now the head of technology track investment at Sunshine Ronghui Capital, recalled that in early 2019, he was introduced to Zhao Lidong through a junior fellow apprentice at Tsinghua University. Because of his professional background, his due diligence on Enflame entered the interior of the R&D organization from the very beginning.

"At that time, Enflame's first chip had not yet completed RTL Freeze (Register Transfer Level code freeze)." Zheng Yan told *Jiazi Lightyear*.

This means that the chip function and logic design of Enflame at that time had not been finally finalized, and there were still multiple key links before being sent to the wafer fab for tape-out.

Zheng Yan and his team respectively interviewed the heads of Enflame's computing core, on-chip network, SoC, DFT and back-end design links. He cared not only about where each person used to work, but also about whether the various modules could be connected smoothly.

Chip development is a highly collaborative project. One or two star engineers can determine the upper limit of local design, but whether the final tape-out can be successful also depends on whether the architecture, front-end, back-end, verification, testing and supply chain teams can complete their respective tasks at the same time node.

For a cautious and conservative insurance-funded investment institution, betting when the chip has not completed tape-out and the commercial prospect is far from clear is not a conventional choice. What promoted this investment to land is not only Zheng Yan's professional judgment on Enflame's team capabilities and tape-out risks, but also a positive pilot of Sunshine Ronghui for long-term capital to enter strategic emerging industries.

The tape-out cost of an advanced chip can reach hundreds of millions of yuan. Once it fails, the company's loss is not only R&D funds, but also the time required for redesign, verification and production scheduling, which often takes half a year or even longer.

For the AI chip industry with rapid technology iteration, missing half a year sometimes means missing a generation of market.

But Enflame did not miss it.

In December 2019, Enflame Technology released its first-generation cloud AI training chip "Saisi" — this chip was later called Saisi 1.0 in the prospectus. At the same time, Enflame also launched the training acceleration card Yunsui T10 based on Saisi 1.0. In December 2020, Enflame launched its first-generation inference acceleration card Yunsui i10.

In December 2019, Zhao Lidong showed the Yunsui T10, Photo source: Enflame Technology

The judgment Zheng Yan formed in the due diligence was verified by the product for the first time.

After that, Enflame continued to iterate: in July 2021, it released the second-generation training chip Saisi 2.0 and Yunsui T20, T21; in December of the same year, it released the second-generation inference chip Saisi 2.5 and Yunsui i20; in 2024, it launched the third-generation inference chip Saisi 320 and Enflame S60; in July 2025, it released the fourth-generation training and inference integrated chip Saisi 400 and Enflame L600.

Enflame S60, Photo source: Enflame Technology

Enflame Technology's prospectus shows that by the end of 2025, the company has completed four generations of computing architectures and five cloud AI chips, and all five chips have achieved one-time successful tape-out.

In the chip industry, "all five chips are successfully taped out at one time" is a rare engineering record. But to understand the path Enflame chose, we first need to clarify the technical route it adopts.

From the perspective of industry-wide classification, there are roughly two mainstream routes for cloud AI acceleration chips. One is GPGPU (General-Purpose Graphics Processing Unit) represented by NVIDIA, which emphasizes general parallel computing capabilities, can cover multiple tasks such as AI computing, scientific computing and graphics rendering, and has formed a mature software and developer ecosystem relying on CUDA.

The other is a dedicated accelerator optimized for AI workloads. Google TPU, Huawei Ascend and Enflame's products are usually classified into this category. Most of these chips adopt DSA (Domain-Specific Architecture) and are