AI News

Jiwei Consulting Releases "AI Industry Series Report: NPU Chips (2026)": Domestic Share Exceeds Half, NPUs Take Center Stage in Computing Power

Jiwei.comUpdated Nov 16 20:58

AI summary

Jiwei Consulting released the "AI Industry Series Report: NPU Chips (2026)," noting that in the first half of 2026, China's data center AI accelerator shipments reached 2.008 million units, with the domestic camp's share significantly exceeding half, and 9 of 11 domestic chip manufacturers saw year-on-year shipment growth. IDC data shows that NVIDIA's share in China is about 34%, followed closely by Huawei at about 28%, while T-Head, Cambricon, Kunlunxin, and Hygon account for about 7%, 5%, 4%, and 3%, respectively. The report holds that NPU competition has escalated from single-chip performance to system engineering competition encompassing "chips + interconnect + complete systems + software stacks."

Body · AI translation

The competitive balance in China's AI computing power market is tilting. According to IDC data, in the first half of 2026, China's data center AI accelerator shipments reached 2.008 million units, with the domestic camp's overall share already significantly above half, and 9 of 11 domestic chip manufacturers saw year-on-year shipment growth—the NPU (neural processing unit) carrying this round of substitution is moving from incremental penetration to stock replacement.

NPU is a processor specially designed to accelerate neural network computing, optimizing AI workloads with large-scale parallel matrix and tensor operations, with energy efficiency significantly better than CPU and GPU, covering two major markets: cloud AI accelerators and edge-side NPUs. According to Gartner forecasts, global total AI spending in 2026 will reach $2.59 trillion, up 47% year-on-year; according to IDC forecasts, global AI infrastructure spending in the same period will reach $497 billion, up about 56% year-on-year, and AI accelerators, where NPUs sit, are the hardware segment with the highest share.

Recently, JW Insights officially released the "AI Industry Series Report: NPU Chips (2026)," featuring four major chapters: industry overview, industry chain analysis, representative enterprise analysis, and future outlook, systematically reviewing the NPU's technological evolution, market landscape, and domestic substitution path.

The following are selected contents of the report:

Market size: historic expansion, China's growth rate near 90%

In the global market, according to TrendForce forecasts, global AI server shipments in 2026 will grow nearly 31% year-on-year, and the combined capital expenditure of nine major CSPs—Google, Amazon, Meta, Microsoft, Oracle, ByteDance, Tencent, Alibaba, and Baidu—will exceed $886.7 billion, up about 90% year-on-year. In the Chinese market, according to IDC forecasts, China's intelligent computing power scale in 2026 will reach 2,576.5 EFLOPS, up 87.9% year-on-year, and will reach 10,524.3 EFLOPS by 2030, providing a continuously expanding demand base for domestic NPUs.

Figure 1 Forecast of global AI-related spending scale in 2026 (Source: Gartner, IDC, compiled by JW Insights)

Figure 2 Forecast of China's intelligent computing power scale (Source: IDC, compiled by JW Insights)

Volume and cost: inference and edge devices drive on two wheels, memory and packaging prices rise

On the volume side, generative AI is shifting from training-led to inference- and agent-led, with edge devices also scaling up simultaneously: according to Gartner data, global AI PC shipments in 2026 are expected to reach 143 million units, with a penetration rate of about 55%, a sharp jump from 77.8 million units in 2025; GenAI phone shipments also reach the scale of hundreds of millions of units, and edge-side NPUs have become standard components of terminal SoCs. On the price side, NPU performance release is highly dependent on memory and packaging. According to TrendForce forecasts, the blended average selling price of HBM in 2027 may rise 121% year-on-year, and CoWoS advanced packaging quotes may be raised by up to about 20%, with the NPU's share of overall device value continuing to increase.

Figure 3 Forecast of global AI PC shipments (Source: Gartner, compiled by JW Insights)

Competitive landscape: overseas in-house development scales up, domestic echelon takes shape

Overseas cloud providers' in-house NPUs are scaling up across the board: Microsoft's Maia 200 has entered production, and Maia 300 will enter mass production in 2027 and expand to gigawatt-level deployment; some international investment banks expect that in 2027, custom ASIC shipments will exceed GPUs for the first time. In the domestic market, according to IDC data, in the first half of 2026, Nvidia's share of China's AI accelerator market was about 34%, with shipments declining significantly year-on-year; Huawei followed closely with about 28%, narrowing the gap to single-digit percentage points, while Pingtouge, Cambricon, Kunlunxin, and Hygon held about 7%, 5%, 4%, and 3%, respectively.

Table 1 Major NPU/AI chip manufacturers' financial and operating performance in 2026

Second-quarter revenue of $96.2 billion (+106%), data center $89 billion (+117%)

Blackwell platform scales up, rack-level solutions dominate the market

Cambricon

Revenue in the first three quarters of 9.526 billion yuan (+106.76%), net profit attributable to parent of 3.715 billion yuan (+131.49%)

Siyuan 690 enters mass production, mainstream large models quickly adapted

Moore Threads

First-half revenue of 1.736 billion yuan (+147.42%), losses significantly narrowed

Commercialization of full-function GPUs and intelligent computing clusters accelerates, planning a Hong Kong listing

Hygon Information

Revenue in the first three quarters of 14.955 billion yuan (+57.58%), net profit attributable to parent of 2.847 billion yuan (+45.12%)

DCU deep computing accelerators broadly adapted to large models

(Source: compiled by JW Insights)

Figure 4 China's AI accelerator shipment share in the first half of 2026 (Source: IDC, compiled by JW Insights)

Domestic substitution: full-stack capability decides the winner, three-stage path clear

Huawei's Ascend 910C supernode deployments have exceeded 1,000 sets, and self-developed HBM has been introduced in new-generation products; Alibaba Pingtouge's Zhenwu series iterates once a year, with cumulative shipments of hundreds of thousands of units and more than 400 customers served; Baidu Kunlunxin's P800 has completed delivery of multiple 10,000-card clusters, and the company has initiated coaching for a STAR Market listing. On the policy side, the 15th Five-Year Plan proposes the orderly deployment of 10,000-card-level and above 100,000-card-level intelligent computing clusters; domestic substitution is advancing in echelons along the path of "internet players first, intelligent computing centers scaling up, industry and edge devices following."

Figure 5 NPU industry chain map (Source: IDC, TrendForce, compiled by JW Insights)

The report judges that NPU competition has escalated from single-chip performance comparison to system engineering competition of "chip + interconnect + complete machine + software stack," with inference and agents becoming the primary driving force; domestic NPUs are accelerating simultaneously in share, ecosystem, and capitalization, and are expected to form a competitive landscape of substantive parallel running with international manufacturers across training, inference, and edge-side full scenarios.

At present, the "AI Industry Series Report: NPU Chips (2026)" has been officially launched on the iJiwei official website and APP. Welcome to log in to the iJiwei official website or Jiwei APP and click the "Jiwei Reports" section on the homepage to place an order.

To view more reports, please click here

Source

Jiwei.com