AXERA (00600) has delivered a strong financial performance during the peak earnings season, reporting a first-half 2026 revenue of 402 million yuan, a 181.8% increase year-over-year, with gross margins significantly improving to 29%. All three business lines—edge AI inference, smart automotive, and terminal computing—are growing in tandem, with the edge AI segment achieving a 251.9% year-on-year increase, establishing itself as a key growth driver. Founded less than a decade ago, AXERA has become one of the fastest-growing edge AI chip companies in China. According to CIC data, AXERA holds a 12.2% market share, ranking third in China's edge AI inference chip industry, and leads globally with a 24.1% share in the mid-to-high-end visual edge chip segment.
Behind this rapid growth, AXERA has chosen a distinct path from most competitors: it has developed a custom NPU processor for AI, breaking through existing NPU computational bottlenecks, and built a product matrix with gradient computing power ranging up to thousands of TOPS. This unified architecture can be reused and migrated across multiple downstream fields, including embodied intelligence, smart driving, and vision, covering a wide variety of application scenarios. The company's proprietary AI processor technology forms the "hardcore" core of its technical foundation.
In the AI computing landscape, the choice of hardware architecture is not a simple binary opposition but is driven by the precise division of computing scenarios and physical constraints. General-purpose GPUs remain the primary choice for cloud-based large model training and ultra-complex computing due to their powerful computing flexibility and universal parallel processing capabilities. However, at the edge and endpoint, located near data sources, NPUs—custom-designed for neural network computing—offer irreplaceable energy efficiency and cost-effectiveness, constrained by power consumption, heat dissipation, BOM costs, and millisecond-level real-time response requirements. The AI computing architecture is now accelerating its shift from centralized cloud computing to distributed cloud-edge-device collaboration. As large models become more widespread, pure cloud inference faces challenges such as network latency, data privacy compliance, and high costs for continuous token calls. Transferring high-frequency, low-latency, privacy-sensitive inference tasks to edge nodes and terminal devices, creating a collaborative system of "cloud training/complex decision-making + edge local efficient inference," has become industry consensus. In this trend, the high efficiency and low power consumption of NPUs are becoming irreplaceable.
From on-device AI chips to inference accelerator cards, edge computing boxes, and inference servers, NPU architecture has become the mainstream choice for low-power scenarios. AXERA's self-developed NPU, the "Axiom NPU," is designed to solve the power and performance bottlenecks of edge inference. It eliminates redundant designs in traditional chips, natively supporting Transformer large models and mixed-precision computing to fully resolve data transfer bottlenecks. With the AX8850 as an example, it achieves a measured energy efficiency ratio up to 10 times that of traditional architectures, processing nearly 200 frames per second per watt. More importantly, it breaks the "difficult to reuse" barrier of traditional dedicated chips—based on the same technical foundation, it can be used for both low-power, small-scale terminals and expanded to high-performance chips for advanced smart driving and high-T-level edge computing. This platform-based reusability, covering a full range of scenarios, significantly reduces duplicate R&D and hardware manufacturing costs, enabling the company to efficiently and rapidly deploy physical AI applications across various scenarios. Currently, the company's product matrix, covering the full spectrum of computing power, has been deployed across thousands of industries, transforming AI into real productivity for the public good.
Edge AI is deployed on servers, gateways, or base stations close to data sources to perform real-time local inference, balancing the high performance of the cloud with the low latency and data security of the endpoint. Driven by stricter privacy regulations, sensitivity to network latency, and high bandwidth costs, AI deployment is shifting from centralized cloud inference to distributed intelligence. Uploading raw data to the cloud increases privacy compliance risks, while edge inference naturally ensures "data stays within the domain." Autonomous driving and other scenarios require millisecond-level responses, which local inference delivers by eliminating network round trips, ensuring high real-time performance and reliability. Additionally, the token costs for cloud inference are rising rapidly, while marginal inference costs are significantly lower, making AI service business models more sustainable. Based on this keen insight, edge AI has become a strategic focus for AXERA. Since 2025, the company has positioned itself ahead of the curve in three key large-model deployment nodes: edge private clouds, edge servers, and intelligent agent terminals. It is actively introducing secure, controllable, and cost-effective AI inference computing products for edge applications to meet the surging demand for computing power.
In July 2026, AXERA announced the establishment of a wholly-owned subsidiary, "AXIS Computing," to further deepen its strategic deployment in edge AI and accelerate the commercialization of related products. It is understood that AXIS Computing will work closely with its parent company, transforming the company's underlying chip capabilities into standardized solutions like computing cards and core boards. By delivering system-level products, it will lower the development threshold for customers and enable rapid application deployment. At the latest WAIC conference, the company unveiled its Yuanxi series of high-computing-power products for the first time. This series of AI inference cards boasts over 1000 TOPS of computing power, large memory capacity, and high bandwidth, designed for high-concurrency scenarios such as server clusters, multi-channel video analysis, and private knowledge bases. It efficiently supports high token consumption tasks like long-text inference and high-definition image analysis, significantly reducing reliance on cloud computing. For low-power, lightweight, and standardized scenarios, the AX8850 computing card supports "plug-and-play," helping small and medium-sized enterprises significantly reduce daily AI application costs. It has been deployed in large volumes in vertical scenarios like smart education and smart industry, with shipment volumes growing substantially during the reporting period. From consumer-grade lightweight to edge-level high computing power, AXERA has built a complete computing product spectrum. Compared to the GPU route, the NPU approach offers significant economic advantages in edge inference, which will become a competitive moat for AXERA across all scenarios.
Meanwhile, AXERA continues to increase its R&D investment to maintain its technological leadership and accelerate the mass production of next-generation products. According to the financial report, AXERA invested 516 million yuan in R&D in the first half of 2026, successfully completing the tape-out of several advanced process SoCs. It is understood that AXERA's next-generation edge AI chip will significantly increase computing power specifications, fully supporting the core nodes of large model deployment, and addressing industry pain points like long-context processing, efficient token generation, and real-time robot perception and decision-making. By using a unified instruction set across endpoints, edge, and data centers, AXERA's NPU architecture enables seamless migration of models and optimization experience between cloud, edge, and endpoint. This highly reusable underlying technology platform provides the company with significant R&D efficiency and commercial expansion advantages. Since 2026, the company has leveraged this architecture to rapidly enter multiple high-growth tracks, unveiling new high-computing-power products in emerging fields like embodied intelligence, advanced smart driving, and intelligent vision perception, accelerating the transformation of its underlying technology moat into real-world applications.
In the field of embodied intelligence, considered the ultimate form of physical AI, AXERA's latest embodied brain controller delivers 1500 TOPS of extreme computing power and ultra-flagship bandwidth approximately twice the industry-leading level. With AEC-Q100 grade comprehensive functional safety design, it ensures reliable device operation. Its open development system supports custom operator development and natively supports cutting-edge algorithms like world models and VLA, providing a high-performance, safe, and flexible intelligent computing hub for general-purpose robots. Combined with the already mass-deployed AX8910 dedicated vision chip for perception, the company's chip solution covers the entire chain from decision-making to perception in embodied intelligence. In addition to embodied robots, the smart driving sector, which also demands high bandwidth and cutting-edge algorithms like VLA and world models, is another core battleground for the commercialization of the company's high-computing-power chips. In the smart driving field, AXERA's advanced smart driving chip, the M97, successfully returned from tape-out and lit up in February 2026. During the first half of the year, engineering samples were delivered, and several leading automakers have initiated solution evaluations and selections. The chip optimizes the long-standing "bandwidth shortage" issue in mainstream domestic chips, doubling the bandwidth of flagship smart driving chips and maximizing effective computing power. Its single-chip computing power exceeds 700 TOPS, supporting L2+/L3/L4 level smart driving functions, and natively supports algorithm architectures like VLA and world models. While the M97 targets the mid-to-high-end smart driving market, the company's current automotive business growth is driven by its flagship product, the M57. To date, the M57 chip has been designated for standard models by several leading domestic OEMs, and overseas OEM projects are also progressing steadily. In the first half of the year, the company's smart automotive business revenue grew by 252.1% year-over-year, with SoC shipments reaching 420,000 units, 24 new mass-production vehicle model projects, and partnerships with 25 domestic and international OEMs, entering a fast lane of volume growth and globalization.
While edge AI and smart driving are growing rapidly, the company's terminal computing business, which forms the core of its revenue, has achieved steady breakthroughs under the trend of upgrading from "perception" to "perception + inference," firmly holding its industry-leading position and providing stable cash flow for new ventures. In the terminal business, as the core of next-generation visual terminal AI inference chips shifts from single "perception" to parallel "perception + local inference," the company's related high-end products are rapidly capturing market share. In the first half of the year, terminal computing business revenue grew by 169.5% year-over-year, with product sales volume increasing by over 100%. Among these, the "Black Light" series product sales growth exceeded 200%, significantly increasing market share. During the reporting period, the company's latest generation "Black Light" high-computing-power visual perception SoC, the AX615 series, saw rapid sales growth. This chip not only inherits the excellent "Black Light" video effects but also integrates a high-computing-power NPU, better meeting the needs of various on-device AI applications, and has been gradually deployed in emerging fields like embodied robots and industrial vision. The trend of AI computing power moving from the cloud to the physical world is an irreversible industrial trend. In the first half of 2026, AXERA successfully rode this trend, driving overall gross margin up to 29.0% through the structural upgrade of its terminal visual core and strong price pass-through capability. At the same time, both the edge AI and smart automotive businesses achieved over 200% year-over-year growth, validating the platform-based reuse advantage of the self-developed NPU architecture. With the establishment of its wholly-owned subsidiary, AXIS Computing, the company has further deepened its edge AI strategy, using standardized AI inference solutions to lower customer development barriers and accelerate the productization of AI applications. The R&D investment and forward inventory accumulation from the concentrated tape-out in the first half of the year are also securing production capacity and next-generation technology advantages for the company during the upcoming scale-up phase of edge AI. As a "shovel seller" in the AI wave, AXERA is well-positioned to be a key player worth watching, backed by its deep technical moat and forward-looking business layout.
Comments