The Hong Kong stock market is on the verge of welcoming another domestic GPU enterprise!

On December 19th, the official website of the Hong Kong Stock Exchange revealed that Iluvatar CoreX had successfully passed its listing hearing and released its prospectus.

On this highly competitive domestic hard technology track, Iluvatar CoreX has achieved remarkable acceleration: Over the past three years, the company's revenues have soared to 189 million yuan, 289 million yuan, and 540 million yuan, respectively.

This performance has catapulted it to the forefront of the industry. Based on 2024 revenue figures, Iluvatar CoreX ranks third in the domestic general-purpose GPU market, holding a 9.8% market share, trailing only Hygon Information and Moore Threads.

However, focusing solely on revenue scale may overlook another equally significant transformation: Iluvatar CoreX's business model is also undergoing an evolution.

In the past, Iluvatar CoreX resembled a traditional GPU product company, with its primary revenue stream coming from direct sales of general-purpose GPUs. In 2023, revenue from general-purpose GPUs still accounted for over 90%.

By 2024, the revenue share from AI computing power solutions had surged to 30.8%, becoming a new growth engine. This indicates the company's shift from 'selling chips' to 'delivering computing power.'

This transition is not unique to Iluvatar CoreX but reflects a common reality faced by domestic GPU manufacturers during the commercialization phase.

To some extent, adopting a more comprehensive delivery model to compensate for the lack of a fully established ecosystem may be the most pragmatic path for domestic GPUs to achieve scale.

/ 01 /

Third-Largest Domestic General-Purpose GPU Manufacturer Achieves 186% Revenue Growth in 24 Years

From a revenue perspective, Iluvatar CoreX has demonstrated rapid growth.

In 2022, Iluvatar CoreX's revenue stood at 189 million yuan. By 2024, this figure had skyrocketed to 540 million yuan, representing a 186% increase. In the first half of 2025, revenue reached 324 million yuan, up 64.47% year-on-year.

In terms of revenue scale, within the domestic general-purpose GPU market, Iluvatar CoreX secured a 9.8% share, ranking third in the industry, trailing only Hygon Information and Moore Threads. (A represents NVIDIA, B represents AMD)

Unlike Moore Threads, which emphasizes 'full-function GPUs,' Iluvatar CoreX focuses more on general-purpose GPUs (GPGPUs), prioritizing AI training and AI inference—the most practical and essential demands.

The earliest GPUs were not designed for 'computing.' Their sole purpose was graphics rendering: swiftly transforming geometric models, textures, and lighting into pixels on the screen. To achieve this, GPUs were equipped with numerous parallel units, excelling at repetitive calculations of the same operation—exactly what graphics pipelines require.

Later, researchers gradually realized that this parallel architecture 'born for graphics' was also well-suited for another type of problem: large-scale matrix and vector operations. Thus, people began to 'conveniently' extract GPUs from graphics pipelines and utilize them for non-graphics calculations.

To distinguish between these two entirely different usage scenarios, academia and industry introduced a new term: GPGPU (General-Purpose Computing on Graphics Processing Units).

This choice directly aligns with Iluvatar CoreX's product matrix, which is divided into two clear main lines:

One is the training general-purpose GPU centered around the 'Tiankai' series;

The other is the inference general-purpose GPU represented by the 'Zhikai' series.

Among them, the Tiankai series is Iluvatar CoreX's flagship product and one of the first domestically produced training general-purpose GPUs to enter mass production.

Its design logic is not derived from graphics but from workload requirements: targeting the characteristics of large throughput, varying precision demands, and multi-card parallelism in large model training, it directly optimizes computing units.

In terms of product timing, Tiankai Gen 1 achieved mass production in 2021, Gen 2 entered mass production in the fourth quarter of 2023 and quickly became the main revenue contributor, and in the third quarter of 2024, the company released Tiankai Gen 3, with plans to enter mass production in the first quarter of 2026.

From shipment data, the training series shipped 7,700 units in 2022, declined to 7,000 units in 2023, and remained stable in 2024. However, in the first half of 2025, shipments reached 6,200 units, significantly higher than the same period last year.

The Zhikai series, an inference chip, is another core product of Iluvatar CoreX.

Released at the end of 2022, the Zhikai series is positioned as China's first general-purpose GPU designed for inference.

Unlike training cards, the core metrics for inference cards are low latency, high throughput, and energy efficiency. Therefore, at the architectural level, the Zhikai series has made targeted optimizations to integer computing units and data pathways and extensively introduced quantization techniques to enhance actual performance in end-user applications.

Even to adapt to more edge scenarios (such as vending machines and industrial computers), Iluvatar CoreX has launched the Zhikai Gen 1X with a power consumption of only 75W. It retains full video processing capabilities but features an extreme design in energy consumption.

Benefiting from the explosion of AI applications, the Zhikai series has seen even stronger growth momentum than the Tiankai series. Shipments of the Zhikai series surged from a mere 38 units in 2022 to 9,800 units in 2024. In the first half of 2025, shipments reached 9,500 units, approaching the full-year level of the previous year.

This dual-wheel-driven strategy is transforming Iluvatar CoreX's revenue structure.

As of 2024, training chips (Tiankai) remained the absolute revenue pillar, accounting for as high as 99.3%. However, by the end of last year, the share of training chips had dropped to 49.9%, while the share of inference chips (Zhikai) had climbed from 0 to 18.6%.

Besides selling GPUs, Iluvatar CoreX has also taken on a more comprehensive task: selling computing power solutions.

The logic behind this is not complex. For many customers, the GPUs they purchase are just components. To turn them into model-running computing power, complex hardware integration and software debugging are required in between.

The specific delivery models mainly fall into two categories:

One is the general-purpose GPU server, akin to a 'plug-and-play' standardized product. It integrates a certain number of general-purpose GPU accelerators with a complete software stack to form pre-configured computing nodes, primarily used for enterprise-level AI workloads and rapid deployment of large language models.

The other is the general-purpose GPU computing cluster, suitable for scenarios with higher demands on computing power density and scheduling capabilities, such as large-scale training and high-throughput inference.

Compared to a single server, this is a more system-level solution that deeply integrates general-purpose GPU products with third-party servers, storage, and network infrastructure, with overall system tuning.

Due to the growth in AI computing power demand, this need was quickly reflected in performance.

In the first year of offering AI computing power solutions, namely 2023, the revenue share of this business reached 5.4%. In 2024, this figure rose to 30.8%.

/ 02 /

For Every 2 Yuan of Revenue Earned, 1 Yuan Goes to the Foundry

While revenue continues to grow, Iluvatar CoreX's loss scale is also expanding simultaneously.

From 2022 to 2024, the company's adjusted net losses were 433 million yuan, 610 million yuan, and 645 million yuan, respectively. In 2025, losses did not continue to widen; in the first half of the year, the adjusted net loss was approximately 300 million yuan, roughly in line with the same period last year.

Against the backdrop of the chip industry, this performance is not bad.

From 2022 to 2024, the company's gross profit margins were 59.4%, 49.5%, and 49.1%, respectively, maintaining a level close to 50%. For a general-purpose GPU company still in rapid product iteration and yet to achieve scale effects, this remains a relatively impressive level.

However, behind the gross profit structure lies one of the most realistic constraints in the chip industry: manufacturing costs are almost unavoidable.

Among operating costs, the largest expenditure comes from chip manufacturing. In 2023, the company paid 136 million yuan to wafer manufacturers, accounting for 47% of annual revenue.

In other words, for every 2 yuan of revenue earned by AI chip manufacturers, about 1 yuan needs to be paid to downstream foundries.

This is not an issue specific to individual companies but a widespread reality in the chip industry during the advanced process era.

Further dissecting the gross profit structure, we can see significant differences among different product lines.

Training chips generally have higher and more stable gross profit margins. Over the past few years, the gross profit margin of training general-purpose GPUs has remained around 60%: 60.2% at the end of 2024, and although it slightly declined to 58.2% in the first half of this year, it remained relatively stable.

In contrast, the gross profit margin of inference chips fluctuates more significantly. At the end of 2024, the gross profit margin of inference chips was 46.7%; by the first half of this year, it had slipped from 52.9% in the same period last year to 32%.

This decline did not stem from a sudden deterioration in cost but was more the result of proactive price adjustments.

The reason is simple: market competition has become fiercer.

The prospectus disclosed that since the end of 2024, new generations of inference products from multiple competitors have successively entered mass production, significantly increasing market supply.

Under such circumstances, to extend its competitive window in mainstream inference scenarios, the company proactively lowered prices to maintain cost-effectiveness advantages in large-scale deployments. Simultaneously, it reserved performance and price space for the next generation of Zhikai products.

As a result, in the first half of 2025, the average selling price of inference series products dropped from 11,400 yuan in the same period last year to 9,200 yuan, a nearly 20% decrease.

Compared to directly selling chips, the gross profit margin of AI computing power solutions is inherently lower and more volatile.

In 2024, the gross profit margin of this business was 31.7%; by the first half of this year, it had risen to 45.7%.

This change was more influenced by project structure.

The reason is that such solutions are typically delivered in the form of 'complete machines or entire systems,' with gross profit margins highly dependent on the model, quantity, and third-party hardware configuration of embedded chips. Differences in scale, performance, and deployment methods among customers directly affect the cost structure of projects, ultimately reflected in gross profit margin fluctuations.

/ 03 /

Summary

Whether it is the already listed Moore Threads or Iluvatar CoreX, preparing to list on the Hong Kong stock market, a common and counterintuitive trend can be observed in the prospectuses of these domestic GPU unicorns:

Domestic GPU companies find it difficult to simply 'sell chips' like NVIDIA.

In the process of scaling revenue, both companies have inadvertently moved towards a more comprehensive delivery model. Moore Threads' main revenue comes from cluster products, while 'AI computing power solutions' are becoming an important revenue pillar for Iluvatar CoreX.

This reveals the most authentic and brutal survival logic in the domestic GPU industry today:

Whether it is Moore Threads' cluster delivery or Iluvatar CoreX's computing power solutions, they are essentially using 'engineering certainty' to compensate for 'ecosystem shortcomings.'

This integrated model of doing dirty and tiring work, although lowering gross profit margins, addresses the most core pain points of Chinese B-end customers: out-of-the-box usability and compliance controllability.

This may seem like a dimensionality reduction in business models, but it is actually a necessary path for domestic chips to cross the survival gap at this stage. Only by first achieving scale through cumbersome system engineering can they potentially feed back into the ecosystem in the future and ultimately evolve from 'project-based' system companies into true 'standard product-based' chip giants.