After the founder's departure, two leadership changes, and fluctuating strategies, ILUVATAR COREX is finally going public

"The 'Four Little Dragons' of Domestic GPUs" are finally set to gather in the capital market.

Following the listings of Moore Threads, Muxi Co., Ltd., and BIREN TECH, Shanghai ILUVATAR COREX Semiconductor Co., Ltd. (hereinafter referred to as "ILUVATAR COREX") is finally about to make its capital debut.

On January 8, ILUVATAR COREX will be listed on the Hong Kong Stock Exchange under the stock code "9903."

With an issue price of HKD 144.6 per share, ILUVATAR COREX's fundraising amount is approximately HKD 3.5 billion, and the IPO market value will reach HKD 35.4 billion.

Compared to other "Four Little Dragons" of domestic GPUs, ILUVATAR COREX has distinct characteristics.

It is the most experienced "veteran" in this group, yet it is also the "pathfinder" facing the toughest journey.

Founded in 2015, ILUVATAR COREX started nearly half a cycle earlier than "newcomers" like Moore Threads. However, in the past decade, it has not enjoyed the stability of a first mover; instead, it has taken on a dramatic script: holding the title of "the first to achieve mass production of training general-purpose GPU chips," while also bearing the heavy footnotes of founder departures, two leadership changes, and strategic fluctuations.

Today, ILUVATAR COREX has evolved into a company co-governed by financial investors and professional managers, with notable investors such as Dazheng Capital and Sequoia Capital among its shareholders.

Stimulated by demand in the AI large model market, ILUVATAR COREX's revenue is expected to reach 540 million yuan in 2024, a year-on-year increase of over 80%.

With its upcoming listing on the Hong Kong stock market, whether ILUVATAR COREX can continue to ride the wave is under scrutiny.

Wavering Route

ILUVATAR COREX has traversed nearly a decade, adorned with multiple halos.

According to Frost & Sullivan, ILUVATAR COREX is the first Chinese chip design company to achieve mass production of inference general-purpose GPU chips, the first to achieve mass production of training general-purpose GPU chips, and the first to reach these milestones using advanced 7nm process technology.

Despite this, ILUVATAR COREX's valuation level is far lower than that of domestic GPU manufacturers like Moore Threads and Muxi Co., Ltd.

Based on the last round of financing before the IPO, ILUVATAR COREX's valuation is 12 billion yuan; whereas the valuations of Moore Threads and Muxi Co., Ltd. before their IPOs have both exceeded 20 billion yuan.

The valuation gap is attributed to the multiple setbacks ILUVATAR COREX has experienced over the past decade, including founder departures, strategic focus adjustments, and changes in the core management team.

In 2015, former Oracle executive Li Yunpeng founded ILUVATAR COREX with a vision for AI computing, proclaiming, "We would rather become a company with systemic capabilities like Google."

In 2019, under Li Yunpeng's leadership, ILUVATAR COREX launched its first AI chip, "Iluvatar CoreX I," an edge AI inference chip that supports visual intelligence algorithms such as detection, classification, and recognition.

From this, it can be seen that ILUVATAR COREX, during Li Yunpeng's leadership, focused on the application of AI chips in end products.

However, less than a year after the chip's launch, founder Li Yunpeng exited the management team of ILUVATAR COREX.

Subsequently, Diao Shijing, who previously served as the director of the Electronic Information Department of the Ministry of Industry and Information Technology, took over as chairman, forcefully reversing ILUVATAR COREX's operational direction During the two years led by Diao Shijing, TianShu ZhiXin completed a key transition to the general GPU track.

In 2021, TianShu ZhiXin launched China's first general GPU product, the "First Generation Training Series TianHua Gen 1," and the following year introduced the first and second generation inference series, ZhiKai Gen 1 and ZhiKai Gen 1X.

However, in 2022, Diao Shijing stepped down from the management of TianShu ZhiXin for personal reasons.

In this turbulent time, the investment representative, Gai Lujiang, who had already been on the board, stepped forward to take over this troubled company.

Unlike Li Yunpeng and Diao Shijing, Gai Lujiang does not have a technical background; his previous experience has been primarily in finance and investment.

TianShu ZhiXin also acknowledged in its prospectus: "Unlike typical founder-led companies, since the start of our track record, the ownership of our company has been held by our employees through multiple special purpose entities and diversified passive financial investors."

Currently, the largest shareholder of TianShu ZhiXin is Shanghai Shuqi, which holds a 23.61% stake and is the general partner of Shanghai XiShi, Shanghai YiShi, Shanghai SuShi, Shanghai NaShi, Shanghai YueShi, Shanghai YuanShi, and Shanghai QiongYu (hereinafter referred to as "holding platform").

Shanghai Shuqi exercises voting rights on behalf of the holding platform and must act according to the decisions of the management committee.

The management committee is mainly composed of the executive directors and senior management members of TianShu ZhiXin, responsible for overseeing and managing the company's daily operations.

It is this decade of ups and downs that has shaped the unique survival landscape of TianShu ZhiXin today.

"Training + Inference" Dual Focus

TianShu ZhiXin's general GPU products are mainly divided into two lines: AI training and inference.

In terms of AI training, as the first domestically mass-produced general GPU product, TianShu ZhiXin's TianHua series is specifically designed for AI model training, which is the main source of revenue.

In the first half of 2024 and 2025, the training GPU products represented by the TianHua series are expected to generate revenues of 269 million yuan and 190 million yuan, respectively, with each accounting for about 50% of total revenue.

Training GPUs are also a high-margin source for TianShu ZhiXin, with a gross margin of 60.2% in 2024, exceeding inference products by 13.5 percentage points.

The ability of training GPU products to carry the revenue banner is a prominent feature that distinguishes TianShu ZhiXin from currently listed domestic GPU manufacturers.

Moore Threads' main revenue comes from AI computing products, including training smart computing cards, inference cards, super node servers, and KuaE smart computing clusters, while MuXi Co., Ltd. focuses on integrated training and inference products.

However, due to increasing competition and other factors, this series has seen price reductions. In the first half of 2025, the unit price of training products was 30,400 yuan, a nearly 25% decline year-on-year.

According to TianShu ZhiXin, this is mainly because it actively lowered the price of TianHua Gen1 to accelerate sales.

In 2023, TianShu ZhiXin launched the higher-performance TianHua Gen2, and it is expected to launch the TianHua Gen4, a general GPU product tailored for computing center demand with a focus on advanced training, in the second quarter of 2026. In the second quarter of 2027, it plans to launch the TianHua Gen5, optimized for advanced and cost-effective deployments suitable for various computing scenarios Complementing TianShu is the ZhiKai series, which is primarily designed for inference.

Benefiting from the explosive demand for model inference, the revenue growth of inference products, mainly from the ZhiKai series, has been significant. In the first half of 2025, it generated 87 million yuan, a year-on-year increase of over three times.

The market is quite optimistic about the future demand for inference.

The main reason is that after the large-scale implementation of generative AI, token consumption has surged, driving the explosion of inference demand.

According to Morgan Stanley's supply chain survey, the focus of AI computing in China is rapidly shifting from the training end to the inference end. Currently, the available computing power in the market mainly comes from NVIDIA's 5090-level consumer graphics cards, early Hopper architecture products, and some domestic AI acceleration chips. In this combination, some local chip suppliers are accelerating the iteration pace of inference-specific chips through more mature 6nm and 7nm nodes.

However, TianShu's inference GPU product gross margin is experiencing a sharp decline, dropping to 32% in the first half of 2025, a year-on-year decrease of over 20 percentage points.

"This is mainly due to our reduction in the prices of ZhiKai Gen 1 and ZhiKai Gen 1X to accelerate inventory sales," TianShu explained.

TianShu's new generation ZhiKai Gen2 and Gen3u inference GPUs are set to launch and go into mass production in the fourth quarter of 2025 and the second quarter of 2026, respectively.

The competition in the inference market is becoming increasingly fierce. For example, the latest S5000 inference card launched by Moore Threads achieves a single card prefill throughput of over 4000 tokens/s, compared to approximately 6500 tokens/s for the NVIDIA H100 in the same scenario.

In the span of ten years, the management has changed twice, from Li Yunpeng's technical passion, to Diao Shijing's course correction, and now to Gai Lujian's capital support, TianShu's turnaround is accompanied by a dramatic shift in strategic focus.

As the focus of AI computing shifts from the training end to the inference end, whether TianShu can break free from the fate of "swaying" and effectively leverage the full-stack layout of "training + inference" to break through the encirclement is drawing attention