Another domestic AI chip company has made its debut on the public stage!

On January 25, the Hong Kong Stock Exchange announced that Axera had successfully cleared the listing hearing and released its prospectus.

Amidst the rapid adoption of AI applications, Axera's performance has entered a high-speed growth trajectory. Over the past three years, the company's revenues have soared from 50.23 million yuan to 230 million yuan, and then to 473 million yuan, marking an impressive over 8-fold surge within two years.

This remarkable growth stems from two key drivers: on one hand, it benefits from the scale expansion through mergers and acquisitions; on the other hand, it reflects the soaring demand for AI inference in real-world applications. Currently, Axera's AI inference chips have achieved large-scale deployment across three major scenarios: visual terminal computing, intelligent vehicles, and edge AI inference.

From an industry perspective, this performance positions the company among the top tier.

Based on 2024 shipment volumes, Axera ranks first in the mid-to-high-end vision-side AI inference chip market, boasting computing power of no less than 1 TOPS. Meanwhile, the company stands as China's second-largest domestic supplier of intelligent driving SoCs and the third-largest domestic supplier of edge AI inference chips.

Next, let's delve into the business and underlying logic of this domestic AI inference chip leader with Silicon-Based Lord.

/ 01 /

Inference Computing Power Shifts to Edge and Terminal: The Battle for Dominance

To comprehend Axera, one must first grasp the two primary forms of AI chips: Training and Inference.

Think of training as "further education at a university." Training an AI model requires processing massive amounts of data and iteratively refining parameters. This demands extremely high computing power density and bandwidth, enabling hundreds or thousands of graphics cards to work in unison like a well-coordinated army. This domain is firmly dominated by NVIDIA.

Inference, on the other hand, represents the model's "real-world application" after graduation.

Every time an AI application is triggered (such as facial recognition for door access or assisted driving brakes), inference comes into play. It doesn't pursue infinite stacking of single-time computing power but demands lightning-fast response (low latency) and extremely low cost (high energy efficiency).

Traditionally, inference was often performed in the cloud. However, with the explosion of AI devices, the limitations of cloud-based inference became apparent: slow data return, privacy concerns, and high bandwidth costs.

Consequently, computing power began to shift downward, following two main paths: edge AI inference chips and terminal-side AI inference chips.

Terminal-side AI chips are straightforward; they are typically deployed in devices like smartphones, wearables, and AI glasses.

Edge AI inference chips, in contrast, are positioned closer to data sources, such as edge servers, gateways, or base stations, and handle real-time, localized inference tasks. Many decisions need to be made within milliseconds, and the data itself is not suitable for cloud processing.

This shift has created a vast incremental market. In 2024, the global market size for AI inference chips reached approximately 606.7 billion yuan, with edge and terminal-side chips rapidly expanding their share, accounting for over 50%.

This trend is particularly pronounced in China. A significant number of AI applications originate not from internet platforms but from physical scenarios like urban governance, industrial systems, transportation, and automobiles.

These scenarios share common characteristics: inference needs to run continuously but doesn't require infinitely scalable computing power. This has carved out a relatively independent market space for terminal-side and edge inference chips.

Axera's business focuses precisely on AI inference chips for edge and terminal scenarios.

From a revenue composition perspective, the company's income primarily stems from terminal computing products, edge AI inference products, and intelligent vehicle products.

Among them, terminal computing products for security and industrial vision scenarios constitute the company's core revenue source.

In 2024, these products generated 447 million yuan in revenue, accounting for 94.5% of the company's total revenue.

In terms of specific forms, terminal computing products are primarily vision perception system-level chips. More intuitively, they serve as the "brain" of cameras.

Among all AI inference tasks, visual data remains the primary input source, accounting for approximately 80% of all perception data. However, it's also the most challenging type of data to process: large data volumes, complex structures, variable shooting environments (such as nighttime, backlighting, strong noise), and strict real-time response requirements.

This places higher demands on computing power organization, memory bandwidth, and energy efficiency for vision-side AI inference chips.

Based on performance, vision-side AI inference chips can generally be divided into low-end and mid-to-high-end categories. Low-end chips typically have computing power below 1 TOPS, while mid-to-high-end chips exceed this threshold and can support more complex visual inference tasks. In 2024, the global market size for mid-to-high-end vision-side AI inference chips was approximately 2.4 billion yuan.

In this niche market, Axera has established a clear advantage. Based on mid-to-high-end vision-side AI inference chips with computing power exceeding 1 TOPS, the company achieved a global market share of 24.1% in 2024, ranking first.

In addition to terminal computing products, Axera has also simultaneously deployed edge AI inference SoCs and intelligent vehicle SoCs.

Among them, edge AI inference SoCs are primarily deployed in edge servers, AI boxes, and all-in-one machines. Based on 2024 shipment volumes, the company ranks third in the domestic edge AI inference chip market.

The intelligent vehicle business focuses on L2 and L2+ ADAS scenarios, with related products achieving mass production of multiple automotive-grade chips and obtaining designated support from multiple Tier 1 suppliers.

/ 02 /

Axera's Two Major "Trump Cards": Computing Efficiently and Seeing Clearly

If the logic for cloud-based AI chips is "might makes right," then on the edge and terminal sides, the logic completely changes. The factors measuring the value of AI chips are only two: computing efficiency and environmental perception.

First, let's talk about computing efficiency. Modern AI inference chips must walk a tightrope between millisecond-level response speeds and milliwatt-level power consumption budgets. Without efficient computing capabilities, perceived data quickly becomes a "burden" rather than an "asset."

Traditional general-purpose processors are ill-suited here; CPUs are too slow to handle trillion-level parallel computations, while GPUs are too expensive and have slow heat dissipation.

This is why more and more edge and terminal-side AI chips are incorporating dedicated NPUs (Neural Processing Units).

By optimizing for neural network computing patterns, NPUs can complete inference tasks with lower power consumption, reducing memory access and heat dissipation pressure, not only improving real-time performance but also extending the overall operational lifespan of devices.

Based on this insight, Axera did not follow the general-purpose computing architecture from the outset but instead chose to focus on NPUs, redesigning chips around inference workloads and launching the Axera Neutron mixed-precision NPU.

If NVIDIA GPUs excel in versatility, capable of handling various complex tasks, then the Axera Neutron is more like a set of carefully crafted tools for specific scenarios.

According to the prospectus, the advantage of the Axera Neutron NPU lies in its ability to dynamically adjust computing precision based on task complexity.

For example, when facing simple tasks, it automatically switches to INT4 or INT8 to reduce power consumption and improve speed; when facing complex tasks, it switches to INT16 to ensure precision.

The brilliance of this architecture lies in its dynamic adjustment of computing precision, significantly reducing computing redundancy and unnecessary data movement. This directly breaks through the common "memory wall" and "data wall" bottlenecks in AI computing.

The results are astonishing. Under the same chip area, the throughput per watt (energy efficiency) of the Axera Neutron NPU is 10 times higher than that of traditional GPU architectures.

In addition to computing efficiency, environmental perception is another major feature of Axera.

In edge scenarios, AI chips must directly confront the noisy real world. Visual signals account for 80% of perception input. If the input images are blurry or noisy, even the most powerful backend models will be "blind."

In such cases, the value of image signal processors (ISPs) becomes evident.

The role of ISPs is simple: to "clean" the raw data collected by cameras (denoising, color calibration, optimizing dynamic range) before AI inference.

However, traditional ISPs rely on fixed rules and often struggle with extremely low light or complex lighting conditions. According to the prospectus, Axera's breakthrough lies in introducing AI into ISPs, launching the "Axera Proton AI-ISP."

Unlike traditional ISPs that rely on fixed rules, this AI-ISP incorporates AI models into the image signal processing chain, performing pixel-level optimization of key links to achieve denoising, high dynamic range enhancement, and low-light environment imaging optimization.

Even in pitch-black nighttime conditions, it can restore clear, full-color images.

In application scenarios highly sensitive to imaging quality, this integrated design that deeply fuses "perception" and "computing" not only improves the quality of data transmission but also reduces the bandwidth demand for data return to the cloud, lowering the overall system latency.

/ 03 /

Why Is Revenue Up 10 Times but Still Losing Money? Dissecting Axera's Financials

Financially, Axera also faces growing pains.

Although revenue has surged, the company is still operating at a loss. From 2022 to 2024, the company's adjusted net losses were 444 million yuan, 542 million yuan, and 628 million yuan, respectively.

In addition to high R&D investment, relatively low gross margins are another major reason for the losses. Over the past three years, Axera's gross margins were 25.9%, 25.7%, and 21%, respectively.

In the cost structure, wafers have always been the company's most significant and challenging expenditure item.

Wafer prices are not solely determined by the company but are highly influenced by the supply chain environment and process technology.

Around 2020, impacted by the pandemic, upstream capacity tightened, wafer supply became scarce, and prices rose significantly. These changes continued to affect chip companies' cost structures in subsequent years, thereby impacting their profitability.

On the demand side, changes in average selling prices (ASPs) also reflect the differentiation among Axera's various product lines.

In 2022, the ASP of terminal computing products was 48.18 yuan per unit; it dropped significantly to 4.49 yuan in 2023; then rebounded to 5.70 yuan in 2024. By the first nine months of 2024, the ASP stabilized at 6.03 yuan and remained largely flat at 6.04 yuan in the same period of 2025.

This decline did not stem from weakened product competitiveness but was primarily due to changes in product mix.

In 2023, the company completed the acquisition of Huatu's related business, leading to a substantial increase in low-priced product shipments, which pulled down the overall average price. As a result, the ASP data for 2022 and subsequent years are not directly comparable in structure.

As the product mix gradually adjusted and digested, the proportion of mid-to-high-end chips increased starting in 2024, and the ASP of terminal computing products began to return to a relatively stable range.

Compared to terminal computing products, intelligent vehicle and edge AI inference products face more intense market competition.

By the first nine months of 2024, the ASP of intelligent vehicle products dropped from 78.70 yuan per unit in the same period of the previous year to 63.30 yuan.

The company explained that this decline was primarily related to increased industry competition. In the intelligent driving chip market, as the number of participants increased and product generations updated more rapidly, price pressure gradually emerged, leading to a temporary decline in ASP.

The pricing trend for edge AI inference products is more representative of the industry.

In 2023, the ASP of these products was 240.39 yuan per unit; it dropped to 185.48 yuan in 2024; by the first nine months of 2024, it was 192.32 yuan and further decreased to 176.43 yuan in the same period of 2025.

Behind the price decline is not weakened demand but changes in customer mix brought about by market expansion. As application scenarios expanded and the number of customers increased, the company placed greater emphasis on scale and penetration in its pricing strategy, seeking shipment volume growth through more competitive prices.

Overall, Axera is still in a stage of scale expansion. While revenue is growing rapidly, R&D investment, wafer costs, and price competition are collectively suppressing profitability. The ASP changes across different product lines also reflect the company's balancing act between expansion and pricing.

Going forward, the company's key focus is not on short-term profitability but on whether it can gradually improve gross margins through product mix optimization and cost control while continuing to scale up shipments. This will determine the long-term sustainability of its growth.