Neural Processing Units (NPU) Market (2026-2036)

The growth of this market is mainly driven by the AI PC upgrade cycle and NPU performance requirements, on-device generative AI in smartphones, the rise of domestic data center NPUs in China, and the spread of NPUs into vehicles, cameras, robots, and industrial devices. However, fragmented software and low NPU utilization, memory bandwidth and cost constraints, export controls and manufacturing limits for Chinese NPUs, and power and silicon area budgets restrain the growth of this market.

Furthermore, NPU intellectual property licensing, sovereign and national NPU programs, and NPUs in automotive, robotics, and microcontrollers are expected to offer growth opportunities for the stakeholders in this market. However, rapid changes in AI model architectures, the gap between advertised TOPS and real performance, competition from GPUs and CPU AI extensions, and scaling challenges for independent NPU developers remain major challenges impacting the growth of this market. Additionally, the NPU performance race in PCs, heterogeneous scheduling across NPUs, GPUs, and CPUs, rack-scale NPU supernodes, and low-precision inference are prominent trends in this market.

The Neural Processing Units (NPU) Market comprises processors and processor blocks designed specifically to accelerate neural network computation, typically using large arrays of multiply-accumulate units, dedicated on-chip memory, and support for low-precision data types. The market covers integrated NPUs within system-on-chips for smartphones, PCs, tablets, wearables, extended reality devices, vehicles, cameras, and microcontrollers; discrete edge NPUs sold as standalone chips or modules; and data center NPUs, such as Huawei's Ascend, Cambricon's MLU, and NPUs from Korean developers Rebellions and FuriosaAI, used for AI training and inference. It also covers NPU intellectual property licensed to chip designers by companies such as Arm, Cadence, Synopsys, CEVA, and VeriSilicon. Market value is measured at the NPU-attributable value of integrated NPUs, consistent with edge AI chip markets, the full value of discrete and data center NPUs, and NPU IP licensing and royalty revenue. GPUs and hyperscaler custom ASICs such as TPUs are excluded and analyzed in AI inference chip markets. The ecosystem spans NPU designers, IP providers, foundries, device and vehicle OEMs, cloud providers, and software developers.

NPUs have spread from niche accelerators to standard components of consumer devices. Apple introduced its Neural Engine in 2017, and smartphone platforms from Apple, Qualcomm, MediaTek, Samsung, Google, and Huawei now include NPUs that run camera processing, voice recognition, and generative AI features on device. In PCs, Intel introduced its first NPU-equipped Core Ultra processors in December 2023, Microsoft's Copilot+ PC specification introduced in 2024 requires NPUs of at least 40 trillion operations per second, and processors such as Qualcomm's Snapdragon X Elite, rated at 45 TOPS, Intel's Lunar Lake, rated at 48 TOPS, and AMD's Ryzen AI 300 series, rated at 50 TOPS, raised NPU performance rapidly in 2024. Microsoft ended support for Windows 10 in October 2025, encouraging PC upgrades to NPU-equipped systems.

In China, data center NPUs are growing rapidly as U.S. export controls restrict access to advanced GPUs. Cambricon reported 2025 revenue of RMB 6.497 billion, up 453%, and net profit of RMB 2.059 billion, its first profitable year, after first-half 2025 revenue rose more than fortyfold, and its first-quarter 2026 revenue rose about 160% to RMB 2.885 billion. Huawei announced in September 2025 its Atlas 950 SuperPoD, combining 8,192 Ascend NPUs, for the fourth quarter of 2026, and the Atlas 960 SuperPoD, with 15,488 Ascend NPUs, for the fourth quarter of 2027. U.S. restrictions on NVIDIA's H20 chip in 2025, and Chinese guidance discouraging purchases of some U.S. chips, have accelerated this shift.

NPUs are also expanding into vehicles, cameras, robots, industrial systems, and microcontrollers, and NPU IP licensing is growing as more chipmakers integrate AI engines. At the same time, the market faces challenges: many applications still run AI on GPUs or CPUs because NPU software is fragmented, memory bandwidth limits performance, TOPS ratings can overstate real-world capability, and Chinese NPUs are constrained by limited access to advanced manufacturing and high-bandwidth memory. Korean NPU developers are pursuing domestic and international opportunities, with FuriosaAI reportedly declining an acquisition offer from Meta in 2025 and securing LG as a customer for its RNGD inference NPU.