Founder’s departure, two CEO changes, strategic wavering—Tianshu Zhixin is finally going public.

```

The "Four Little Dragons of Domestic GPU" are finally gathering in the capital market.

Following the successive listings of Moore Threads, Muxi, and Biren Technology, Shanghai Iluvatar CoreX Semiconductor Co., Ltd. ("Iluvatar CoreX") is finally set to make its own capital market debut.

On January 8th, Iluvatar CoreX will be listed on the Hong Kong Stock Exchange with the stock code "9903".

Based on the issue price of HKD 144.6 per share, Iluvatar CoreX will raise about HKD 3.5 billion, with an IPO valuation reaching HKD 35.4 billion.

Compared to other members of the "Four Little Dragons of Domestic GPU," Iluvatar CoreX has distinctly unique characteristics.

It is the most senior "veteran" in this group, but it has also faced the toughest journey as a "trailblazer."

Founded in 2015, Iluvatar CoreX started almost half a cycle ahead of later competitors like Moore Threads. Yet, over the past ten years, it hasn't enjoyed the stability of a first mover; instead, its story is full of drama: crowned as "China’s first company to mass-produce general-purpose GPU chips for training," but also burdened with the founder's departure, two changes in leadership, and drastic strategic swings.

Today, Iluvatar CoreX has evolved into a company jointly run by financial investors and professional managers, with major shareholders including well-known investors like Centurium Capital and Sequoia.

Driven by demand in the large AI model market, Iluvatar CoreX's revenue in 2024 reached 540 million yuan, a year-on-year increase of more than 80%.

With this listing on the Hong Kong market, whether Iluvatar CoreX can continue to ride the momentum is under close scrutiny.

Wavering Strategy

Iluvatar CoreX has experienced nearly a decade, carrying several distinctions.

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

Even so, Iluvatar CoreX's valuation is much lower than domestic GPU makers such as Moore Threads and Muxi.

Based on the last round of financing before its IPO, Iluvatar CoreX is valued at 12 billion yuan; while Moore Threads and Muxi both exceeded 20 billion yuan before their IPOs.

Behind this valuation gap is a decade of founder departure, strategic shifts, and management changes for Iluvatar CoreX.

In 2015, former Oracle executive Li Yunpeng founded Iluvatar CoreX with the vision of AI computing, declaring "We aspire to be a company with system capabilities like Google."

In 2019, under Li Yunpeng's leadership, Iluvatar CoreX launched its first AI chip, "Iluvatar CoreX I," designed for edge AI inference, supporting detection, classification, recognition, and other visual intelligence algorithms.

This shows that under Li Yunpeng’s leadership, Iluvatar CoreX focused on the application of AI chips in end products.

But less than a year after this chip was launched, Li Yunpeng as founder left Iluvatar CoreX’s management.

Afterwards, Diao Shijing, former Director of the Electronic Information Department of the Ministry of Industry and Information Technology, assumed the role of Chairman, forcefully steering Iluvatar CoreX in a new direction.

During Diao Shijing’s two years in charge, Iluvatar CoreX made a key pivot towards the general-purpose GPU track.

In 2021, Iluvatar CoreX launched China’s first general-purpose GPU product, the “first-generation training series Tiangai Gen 1,” and the next year released the first and second-generation inference series Zhikai Gen 1 and Zhikai Gen 1X.

But in 2022, due to personal reasons, Diao Shijing exited the management team.

In the ensuing turbulence, board representative investor Gai Lijiang stepped to the forefront, taking over this much-troubled company.

Unlike Li Yunpeng and Diao Shijing, Gai Lijiang does not have a technical background; his career has mainly focused on finance and investment.

Iluvatar CoreX also admitted in its prospectus: "Unlike a typical founder-led company, since our historical record period began, the company's ownership has always been held by employees through various special purpose vehicles and diverse passive financial investors."

Currently, the largest shareholder of Iluvatar CoreX is Shanghai Shuqi, which holds 23.61% equity and is the general partner of Shanghai Xishi, Shanghai Yishi, Shanghai Sishi, Shanghai Nashi, Shanghai Yueshi, Shanghai Yuanshi, and Shanghai Qiongyu (hereafter called “the holding platform”).

Shanghai Shuqi represents the holding platform in exercising voting rights, and must act in accordance with decisions of the management committee.

This management committee is primarily made up of Iluvatar CoreX’s executive directors and senior management, responsible for overseeing and managing the company’s daily operations.

It is exactly the ups and downs of these ten years that have shaped the unique business landscape of Iluvatar CoreX today.

"Training + Inference" Dual Strategy

Iluvatar CoreX’s general-purpose GPU products are mainly divided into two lines: AI training and inference.

On the AI training side, as the first mass-produced general-purpose GPU product in China, the Tiangai series from Iluvatar CoreX is designed for AI model training and is the main source of income.

In 2024 and the first half of 2025, training GPU products led by the Tiangai series brought in revenue of 269 million yuan and 190 million yuan respectively, accounting for around 50% of total revenue.

Training GPUs also deliver high gross profit margins, with a margin of 60.2% in 2024, 13.5 percentage points higher than inference products.

The dominance of training GPUs in revenue distinguishes Iluvatar CoreX from other listed domestic GPU firms.

Moore Threads’ main income comes from AI computing products including training smart cards, inference cards, supernode servers, and Kua'e AI computing clusters. Muxi mainly focuses on integrated training-inference products.

However, due to intensifying competition and other factors, prices of the series have declined. In the first half of 2025, the unit price of training products was 30,400 yuan, down nearly a quarter year-over-year.

Iluvatar CoreX explained this was mainly because they proactively reduced prices for Tiangai Gen1 to accelerate sales.

In 2023, Iluvatar CoreX launched the higher-performance Tiangai Gen2. Gen4, a general-purpose GPU tailored for computing center needs with advanced training capabilities, is scheduled for Q2 2026, while Gen5 optimized for advanced and cost-effective deployments across various scenarios is planned for Q2 2027.

Complementing the Tiangai series is the Zhikai series, mainly designed for inference.

Benefiting from surging demand for model inference, revenue from Zhikai series-based inference products jumped sharply, reaching 87 million yuan in the first half of 2025—up more than threefold year-on-year.

The market holds high expectations for further demand in the inference sector.

Mainly because, after the large-scale rollout of generative AI, token consumption has surged, driving explosive growth in inference demand.

Morgan Stanley’s supply chain research shows that China’s focus in AI computing is rapidly shifting from training to inference. Currently, computing power in the market mainly comes from Nvidia’s 5090-class consumer graphics cards, early Hopper architecture products, and some domestic AI accelerator chips. With this mix, some local chip suppliers are speeding up the update cycle for dedicated inference chips using mature 6nm and 7nm process nodes.

However, gross profit margins for Iluvatar CoreX’s inference GPUs are plummeting, hitting 32% in the first half of 2025, down more than 20 percentage points year-on-year.

“Mainly because we reduced prices for Zhikai Gen1 and Zhikai Gen1X to speed up inventory sales,” the company explained.

Iluvatar CoreX's next-generation Zhikai Gen2 and Gen3u inference GPUs will be released and mass-produced in Q4 2025 and Q2 2026 respectively.

Competition in inference is intensifying. For example, Moore Threads' newest S5000 inference card achieves single-card prefill throughput exceeding 4,000 tokens/s, compared to about 6,500 tokens/s for Nvidia H100 in similar scenarios.

Ten years, two management shifts; from Li Yunpeng’s technical vision, to Diao Shijing’s strategic pivot, to Gai Lijiang’s capital guidance—every transition has brought fierce swings in strategy for Iluvatar CoreX.

As the AI computing focus shifts from training to inference, whether Iluvatar CoreX can break free from its “wavering” fate and capitalize on its full-stack "training + inference" strategy is closely watched.

Risk Warning and DisclaimerThe market carries risks, and investment should be approached with caution. This article does not constitute personal investment advice, nor does it take into account individual users' special investment objectives, financial situations, or needs. Users should consider whether any opinions, views, or conclusions in this article are suited to their own circumstances. If you invest based on this, you do so at your own risk. ```