Neural processing units (NPUs) designed for artificial intelligence are no longer a single product category; they've fractured into distinct market tiers with vastly different capabilities and price points. The gap between the cheapest AI-labeled mini PCs and premium desktop machines has grown so wide that comparing them by raw specifications alone tells almost nothing about real-world performance. This fragmentation is reshaping how buyers should think about on-device AI hardware.

What's Driving the NPU Performance Gap?

The clearest example of this divide emerged in September 2026 when ASUS unveiled the Ascent QN10 mini PC, a desktop machine built around Qualcomm's Snapdragon X2 Elite processor and an 80 TOPS neural processing unit. That 80 TOPS figure, which measures the theoretical peak throughput for AI inference operations, sits at exactly double Microsoft's Copilot+ PC baseline of 40 TOPS. But the real story isn't about ASUS's machine; it's about what the competition looks like.

When you line up publicly confirmed NPU ratings across the current crop of AI-branded mini PCs and compact laptops, the spread becomes striking. The MINIX N304-AI mini PC delivers only 24 TOPS, while the Acer Swift Air 16, despite carrying Microsoft's Copilot+ branding, manages just 17 TOPS. Both fall below the 40 TOPS requirement they're supposedly designed to meet. The ASUS Ascent QN10's 80 TOPS doesn't just exceed that bar; it doubles it, creating a performance tier that separates premium machines from budget alternatives in a way that raw core counts never could.

Why Does TOPS Matter Less Than You'd Think?

Here's where the story gets complicated. TOPS measures peak throughput for AI inference, not real-world task completion time. A machine with 80 TOPS doesn't necessarily run AI workloads twice as fast as a 40 TOPS system, and it certainly doesn't guarantee that agentic AI applications (software that can work through multiple steps toward a goal rather than simply responding to a single prompt) will run smoothly at usable speeds. The ASUS Ascent QN10 clears Microsoft's certification bar with room to spare, but whether it can actually run today's most capable agentic models locally at practical speeds remains untested in public benchmarks.

This measurement gap matters because buyers shopping by TOPS rating alone are comparing headline numbers without understanding what those numbers mean in practice. Software optimization, memory bandwidth, the complexity of a particular task, and how much work stays on the device versus traveling to a cloud server all affect the actual experience. A 55 TOPS neural processor in MediaTek's new Dimensity CX C10 Max, designed for Google's upcoming Googlebook, may deliver a fundamentally different user experience than an 80 TOPS system, depending entirely on how Google and individual applications use that processing headroom.

How Are Companies Building Infrastructure Around NPUs?

The fragmentation extends beyond consumer hardware into enterprise infrastructure. South Korean companies Okestro and FuriosaAI signed a memorandum of understanding in September 2026 to cooperate on neural processing unit-based artificial intelligence infrastructure and platform technology. The partnership emerged from months of technical verification work, during which the two companies installed servers running FuriosaAI's second-generation NPU, called RNGD, in Okestro's AI infrastructure environment.

The collaboration reveals a critical challenge: NPUs are only useful if the software ecosystem and infrastructure platforms support them. Okestro expanded its Concerto AI inference operations platform, previously designed mainly for graphics processing units (GPUs), to cover FuriosaAI's RNGD NPU. This kind of integration work is invisible to consumers but essential for enterprises trying to build AI infrastructure that can efficiently use various accelerators depending on workload characteristics.

"With this MOU as a starting point, we will combine the two companies' technological competitiveness to build new AI infrastructure references in the public, financial and corporate markets," said Kim Min-jun, chairman of Okestro Group.

Kim Min-jun, Chairman of Okestro Group

The partnership plans to jointly design RNGD-based AI infrastructure tailored to the specific needs of public-sector, financial, and corporate customers, then link joint sales, marketing, and technical support to expand market adoption. Over the medium to long term, they will advance integrated operations and monitoring technology spanning models, inference runtimes, and AI accelerators, extending cooperation to heterogeneous AI infrastructure environments that can efficiently use various accelerators including GPUs and NPUs depending on workload characteristics.

What's Happening in the Premium Laptop Market?

Meanwhile, the premium laptop segment is taking a different approach. MediaTek's new Dimensity CX C10 Max, built on TSMC's 3-nanometer manufacturing process, combines an eight-core CPU, an 11-core GPU, a 55 TOPS neural processor, Wi-Fi 7, and an advanced image processor aimed at better video calls. The chip is destined for an upcoming premium, ultra-portable Googlebook starting at $899.

The Dimensity CX C10 Max's 55 TOPS neural processor sits between the budget and premium tiers, but the real innovation lies in the supporting hardware. The chip includes an advanced image signal processor that supports camera sensors up to 32 megapixels, with AI-assisted processing for HDR video, smoother video zoom, sustained focus, and what MediaTek describes as cinematic video. It can also drive two external 4K displays, a useful feature for anyone who wants a lightweight travel computer that can become a multi-monitor workstation once they reach a desk.

How Should Buyers Evaluate NPU Hardware?

The fragmentation of the NPU market means that TOPS ratings alone are insufficient for making purchasing decisions. Here are the key factors that matter beyond raw specifications:

• Software Ecosystem Support: An NPU is only useful if the operating system, applications, and cloud services you rely on actually use it. Google's Gemini Intelligence features and Microsoft's Copilot+ ecosystem are the primary drivers of NPU adoption in consumer devices, so compatibility with your preferred AI services matters more than raw throughput.
• Real-World Benchmarks: TOPS ratings measure theoretical peak performance, not practical task completion time. Until reviewers publish end-to-end benchmarks showing how long it actually takes to run agentic AI workloads on specific hardware, TOPS comparisons are incomplete. The ASUS Ascent QN10 has not yet been tested against comparable Intel or AMD mini PCs in standardized benchmarks like Cinebench or Geekbench.
• Supporting Hardware Quality: An NPU's effectiveness depends on memory bandwidth, cache size, and the quality of supporting components. The ASUS Ascent QN10 includes dual M.2 NVMe slots and 2.5 GbE Ethernet, both of which matter for the kind of local AI workloads the machine targets. The MediaTek Dimensity CX C10 Max includes an advanced image processor and Wi-Fi 7, features that extend the chip's usefulness beyond raw AI throughput.
• Thermal and Power Efficiency: A 55 TOPS NPU in a fanless laptop may deliver a different user experience than an 80 TOPS NPU in a machine with active cooling. MediaTek is explicitly designing the Dimensity CX C10 Max for thin, light, fanless computers that can run all day without becoming a space heater or requiring constant charging.

What Does This Mean for the Future of On-Device AI?

The NPU market is maturing in a way that mirrors the early days of GPU adoption. Just as graphics processors evolved from niche gaming hardware into essential infrastructure for AI training and inference, neural processing units are becoming standard components in consumer and enterprise devices. But unlike GPUs, which have a relatively unified software ecosystem, NPUs are fragmenting into multiple incompatible designs from Qualcomm, MediaTek, Apple, and others.

This fragmentation creates both opportunity and risk. For consumers, it means the market is still early enough that buying decisions should focus on software support and real-world performance rather than headline specifications. For enterprises, it means infrastructure partnerships like the Okestro and FuriosaAI collaboration are essential for building systems that can work across multiple NPU designs. The companies that win this market will be those that can bridge the gap between hardware capabilities and software reality, not those that simply chase the highest TOPS rating.