AI News

Jimei Consulting Releases "AI Industry Series Report: VPU Chips (2026)": Edge Inference Explodes, VPU Leaps to the "Third Pole" of Computing Power Systems

Jiwei.comUpdated Nov 16 20:58

AI summary

Jimei Consulting has released the "AI Industry Series Report: VPU Chips (2026)," systematically reviewing the technological evolution, market landscape, and domestic substitution path of VPUs. According to IDC data, in the first quarter of 2026, global shipments of device-side and edge AI chips grew 45% year-over-year, nearly four times the concurrent growth rate of cloud AI chips (12%); the Houmo Intelligence M50 chip, with 160 TOPS@INT8 computing power and 10W typical power consumption, has already secured more than 150 product design wins within less than a year of mass production.

Body · AI translation

As AI inference workloads accelerate their migration from the cloud to the edge, a "third computing power mainline" running parallel to cloud training and cloud inference is taking shape. According to IDC data, in the first quarter of 2026, global shipments of on-device and edge AI chips grew 45% year-on-year, nearly four times the concurrent growth rate of cloud AI chips (12%), and the VPU (vision processing unit), which carries visual perception and local inference, is riding the steepest growth curve of this cycle.

The industry space for VPUs is rooted in the AI-driven semiconductor super upcycle. According to WSTS statistics, the global semiconductor market reached US$791.7 billion in 2025, up 25.6% year-on-year, and is expected to reach US$1,655.0 billion in 2026 and US$2,142.0 billion in 2027. In 2025, global consumer IPC shipments already exceeded 192 million units; combined with terminals such as AI phones and AI PCs, the installed base of devices capable of carrying VPU capabilities is measured in billions of units.

Recently, JW Insights officially released the "AI Industry Series Report: VPU Chips (2026)," which is organized into four major chapters—industry overview, industry chain analysis, representative companies, and future outlook—and systematically reviews the technological evolution, market landscape, and localization path of VPUs.

Below are selected contents from the report:

Market size: a super upcycle led by edge-side growth

The expansion of the edge-side market is particularly evident in China's terminal data: according to IDC data, China's AI phone shipments in 2025 were about 118 million units, accounting for 40.7% of China's smartphone market, and are expected to reach 147 million units in 2026, with penetration rising to 53.0%; according to Gartner forecasts, global AI PC shipments will reach 143 million units in 2026, accounting for 55% of total PC shipments.

Figure 1 Global semiconductor market size and forecast (Source: WSTS, compiled by JW Insights)

Figure 2 China AI phone shipments and forecast (Source: IDC, compiled by JW Insights)

Growth logic: rising volume and price, with edge inference becoming the main battlefield

The incremental growth of the VPU market in this cycle is driven by both "volume growth" and "price increases." On the volume side, according to IDC and Gartner forecasts, by 2027 more than half of generative AI model inference will be completed on the edge side. On the price side, the average price of flagship SoCs has increased by about 30%-50% compared with the non-AI era; Houmo AI's M50 chip, mass-produced based on the SRAM-CIM route, defines a new benchmark for on-device large-model inference with 160 TOPS@INT8 computing power and 10W typical power consumption, and has already secured more than 150 product design wins within less than a year of mass production.

Figure 3 Comparison of year-on-year shipment growth of global on-device/edge and cloud AI chips in the first quarter of 2026 (Source: IDC, compiled by JW Insights)

Competitive landscape: the domestic camp enters a period of concentrated performance realization

In the first half of 2026, domestic visual AI chip manufacturers fully delivered on performance: Rockchip's revenue was 2.877 billion yuan, up 40.60% year-on-year; SigmaStar's revenue was 2.658 billion yuan, up 89.43% year-on-year; Fullhan Micro's revenue was 1.463 billion yuan, up 112.66% year-on-year; Horizon Robotics' revenue was 2.055 billion yuan, up 31% year-on-year; and AXERA's revenue was 402 million yuan, up 181.8% year-on-year. Horizon Robotics' Journey series chips have cumulatively exceeded 15 million units in mass production, with a 50.02% share of the ADAS market for Chinese independent-brand passenger vehicles; in the upstream IP segment, VeriSilicon's NPU IP has been adopted by 91 customers, and cumulative global shipments of AI chips integrating this IP have reached the hundreds-of-millions level.

Table 1 Operating indicators of major Chinese visual/edge AI chip manufacturers in the first half of 2026

(Source: compiled by JW Insights)

Figure 4 VPU industry chain map (Source: IDC, compiled by JW Insights)

Demand evolution: robotics as the largest incremental driver to take over

The report judges that the VPU substitution pace follows a three-wave structure: "security first, phones and PCs next, robotics taking over." Robotics represents the new demand with the largest marginal change: according to IDC statistics, global humanoid robot shipments were close to 18,000 units in 2025 and are expected to exceed 510,000 units by 2030, corresponding to a compound annual growth rate of nearly 95%; embodied intelligence places higher demands on synchronized visual perception across multiple body parts, making the VPU the core computing power carrier for robots' "eyes" and "cerebellum."

Figure 5 Global humanoid robot shipments and forecast (Source: IDC, compiled by JW Insights)

Policy and localization: a key window period opens

From December 2025 to January 2026, the Ministry of Industry and Information Technology and seven other departments issued the "Implementation Opinions on the 'AI + Manufacturing' Special Action," and edge inference chips were included for the first time in the chip research catalog of a national special action. The report predicts that in 2027, edge-side generative AI inference will account for more than half, and domestic edge inference platforms will enter large-scale commercial use; from 2028 to 2030, with the ramp-up of humanoid robots and the spread of edge private clouds, the domestic VPU industry chain is expected to obtain a global share in this largest incremental computing power market that matches its design capabilities.

Overall, the domestic VPU force has moved from "security substitutability" into a stage of "multi-scenario global competition." Memory supply constraints determine the short-term pace, while energy efficiency and software ecosystems determine the long-term landscape; manufacturers with complete capabilities spanning "perception algorithms + inference architecture + development environment" will occupy the high end of the value chain.

At present, the "AI Industry Series Report: VPU Chips (2026)" has officially launched on the JW Insights official website and APP. You are welcome to log in to the JW Insights official website or the JW Insights APP and click the "JW Insights Reports" section on the homepage to place an order.

To view more reports, please click

Source

Jiwei.com