More than half of the PCs sold in 2026 ship with an NPU. How many of those NPUs do any real work is a different question, and the one worth answering before a fleet refresh.
The data: the institutions are not measuring the same thing
| Institution | 2026 global AI PC forecast | Penetration |
|---|---|---|
| Gartner (May 2026) | 143.1 million units | 54.7% (first time above half) |
| Goldman Sachs | 150 million units | 59% |
| Counterpoint | "More than half of global shipments" | Above 50% |
| Canalys | — | 43.5% (new PCs with a 40+ TOPS NPU) |
| Morgan Stanley | — | About 28% (60% by 2028) |
Optimists say 55% for 2026, pessimists 28%, nearly a factor of two. The gap is definitional: count every machine with an NPU and penetration looks high; count only Copilot+ PCs, with a 40+ TOPS NPU and the AI features actually on, and it falls. Morgan Stanley uses that stricter definition.
China's numbers are more consistent:
- IDC puts China's 2026 AI PC shipments near 22 million units at 52% penetration, up 146.5% year on year, with a 58.7% compound annual growth rate through 2029.
- Canalys tracks China's penetration at 8% in 2023, 15% in 2024, 34% in 2025, 52% in 2026 and 73% in 2028.
- Lenovo shipped 2.53 million AI PCs in China in the first quarter of 2026, 55% of its domestic PC sales.
- China shipped 14.62 million AI PCs in 2025, up more than 150%, over 140 percentage points ahead of the overall market.
There is also a counterintuitive macro backdrop: IDC expects China's PC shipments to slip slightly in 2026 on the memory shortage, while market value still rises 1.6%. Fewer machines, more money, and AI PCs are the only structural growth area.
The hardware bar: 40 TOPS is a ticket, not a capability
Microsoft draws the Copilot+ line at 40 TOPS. Here is where the 2026 platforms sit.
| Platform | NPU | Architecture | Notes |
|---|---|---|---|
| Qualcomm Snapdragon X Elite | 45 TOPS | ARM / Oryon | Best battery life; x86 via translation |
| Intel Lunar Lake (Core Ultra 200V) | 48 TOPS | x86 | Best office experience |
| Intel Panther Lake (3rd-gen Core Ultra) | 50 TOPS NPU5; 180 TOPS platform | x86 / Intel 18A | January 2026; claims 27 hours |
| AMD Ryzen AI 300 (Strix Point) | 50 TOPS | x86 | Strong multi-core, balanced price |
| Qualcomm Snapdragon X2 Elite | 80 TOPS | ARM | Thin-and-light leader |
| NVIDIA N1X | 180+ TOPS | ARM / Tensor | Runs larger models locally |
| Apple M4 | 38 TOPS | ARM | No Windows, outside Copilot+ |
In short: 40 TOPS is the ticket, 80 TOPS the new flagship, and 180+ TOPS is where a local model of real size becomes possible.
What the NPU does: useful work and idle silicon
Ordered by actual usage, according to industry survey data:
| Function | Usage | Practical value |
|---|---|---|
| Real-time voice transcription and call summaries | Most-used feature, 48% adoption | Meetings transcribe themselves; live subtitles |
| Photo editing and generative erase | Second, used by 42% of smartphone users | Useful, not essential; PCs trail phones |
| Video call Studio Effects | Blur, eye correction, auto framing | Practical; 65–80% less power than a GPU |
| Natural language file search | New in the second half of 2026 | Clear value for sales, admin and legal desks |
| Local large-model inference | 7B–8B quantized, 18.5–28 tokens/s on a mobile NPU | Needs 16–24GB and a setup-minded user |
| Copilot chat | Most queries still go to the cloud | Low NPU involvement; the emptiest claim in AI PC marketing |
The conclusion is blunt. The NPU earns its keep on sensory work such as voice, images and video, at low power, which is why batteries last longer and calls run smoother. Running a large model locally works, but for most office users it sits idle.
What actually drives on-device AI
Why are companies taking local AI seriously? The pain is specific:
- Data cannot leave the building. Business documents, confidential R&D and customer records cannot go to a third-party server, and for government, manufacturing, finance and healthcare that is a hard rule. Industry surveys put the share of IT security leaders who insist on local processing at 76%.
- Per-token API costs compound. Teams that lean on AI can run monthly bills above RMB 10,000; over three years that exceeds the cost of a batch of well-configured machines.
- Some sites have no network. Factory floors, field work, classified intranets and restricted regions cannot use a cloud API at all.
The conditions have arrived. Quantization, distillation and sparsification make models above ten billion parameters practical for documents, code and images, and open-source tools let an ordinary technical user finish a local install. Gartner sizes edge AI hardware at US$38.5 billion; local inference uses 65–80% less power than a cloud GPU and can cut bandwidth and egress costs by 60–85%. It mirrors mobile chips before the smartphone boom: mature hardware, usable software, real demand.
How to decide: three cases
Case 1: You need local AI, so buy high
Any one trigger puts you here: data that cannot leave the premises, a cloud API bill above RMB 5,000 a month, or offline AI in the field or on a classified network.
- An NPU of 50 TOPS or more (Core Ultra 200V or 300, Ryzen AI 300, or the 80 TOPS Snapdragon X2 Elite).
- 32GB to start, 64GB to run models locally: one 8B model under concurrent use needs a 16–24GB unified memory baseline, plus the system and business software.
- A 1TB SSD, since model files run 4–15GB each.
- Memory is the wild card: the memory alone in a 64GB build now costs RMB 7,000. Budget from real quotes.
Case 2: Ordinary office work, so skip the premium
OA, Excel, Kingdee, browsers, email: none of it touches an NPU. The previous generation of non-AI machines (Core Ultra 100 series, Ryzen 7040/8040) is the best value, and the same specification costs 15–25% less. The exception is a machine that must last more than three years: Windows 11's roadmap is tied to the NPU, and older systems will lose features such as Studio Effects.
Case 3: The old machine still runs, so upgrade
An eight-year-old i5-4590 with 4GB of memory and a mechanical disk, given a 240GB SSD and 8GB more memory, went from a 127-second boot to 19 seconds, and OA and Kingdee entry from laggy to instant. Thirty-two machines cost RMB 12,000, or RMB 380 each, against RMB 144,000 new. If the upgrade costs under 15% of a replacement and buys another three years, upgrade.
Not worth upgrading: a dual-core low-voltage CPU, a DDR2 platform, a board with hidden faults, or work already dependent on local AI and large design software.
Three ways to avoid a bad buy
- Do not pay for a TOPS number. 50 and 80 TOPS are indistinguishable in office work; the gap only appears with a local model.
- Test software on ARM. Snapdragon X battery life is real, but x86 software runs through translation, and old ERP clients, dongle drivers and industrial packages can break. Run a POC with your own software.
- Do not buy the first batch. The first three to six months bring BIOS bugs, immature drivers and unstable supply. Wait for the second batch, when prices soften and faults are known.
ACCPC handles enterprise PC procurement and volume builds: role-by-role needs assessment, branded machines through channel partners (Lenovo, Dell, HP and ASUS, channel pricing from 10 units), custom assembly, batch deployment with master image, cloning, preinstalled software and asset tagging, old-machine upgrades and buyback, and a three-year on-site warranty. We work on site across Guangzhou through the weekend and wholesale computers and components. Call 020-39029800.