The “AI PC” generation is defined by the NPU (Neural Processing Unit) — a low-power accelerator that runs small models continuously (background agents, always-on voice, Copilot+ features) without draining the battery. Microsoft’s Copilot+ bar is a 40+ TOPS NPU.
Honest caveat: for serious LLM work (8B+ models), the NPU matters less than CPU/GPU/RAM. The GPU and unified memory do the heavy lifting; the NPU is for always-on, low-power workloads. All model picks lead with the newest releases (Aug 2026).
| Spec | Value |
|---|---|
| NPU | 48 TOPS (NPU4) |
| CPU | 8 cores (4 P-core + 4 E-core), ~5.1 GHz max boost |
| GPU | 8 Xe2 (Battlemage) cores |
| Memory | 16–32 GB LPDDR5X-8533 (on-package) |
| Total platform AI | ~48+ TOPS |
| Process | TSMC N3B |
| Availability | Laptops since Sept 2024–2025 |
| Price (Aug 2026) | ~$600–1,400 — 2024–25 stock is deeply discounted (budget units ~$600, mid-tier ~$900); premium configs $1,400+ |
Best for: Battery-efficient Copilot+ laptops; the first Intel chips to clear the 40 TOPS bar.
Recommended AI models (newest first):
| Spec | Value |
|---|---|
| NPU | 50 TOPS (XDNA2) |
| CPU | HX 370: 12 cores (4 Zen 5 + 8 Zen 5c); HX 365: 10 cores |
| GPU | Radeon 890M / 880M (RDNA 3.5, 16/12 CU) |
| Memory | LPDDR5X-7500 / DDR5 |
| Price (Aug 2026) | ~$900–1,500 (HX 370 mid-tier ~$900–1,200; premium/creator configs $1,400+) |
Best for: 2025’s best all-round AI laptop platform — the most powerful NPU of its generation plus a strong iGPU that also accelerates LLMs.
Recommended AI models (newest first):
| Spec | Value |
|---|---|
| NPU | 45 TOPS (Hexagon) |
| CPU | Up to 12 Oryon cores, up to 4.6 GHz |
| GPU | Adreno X1 |
| Memory | LPDDR5X-8448 |
| Process | 4 nm |
| Availability | Laptops since mid-2024 |
| Price (Aug 2026) | ~$600–1,000 mainstream (Qualcomm lists ‘from $599’; Dell from ~$1,000); premium $1,400–1,700 (Surface Laptop 7 ~$1,690) |
Best for: Outstanding NPU efficiency and battery life; the original Copilot+ platform. Arm-based, so x86 app compatibility needs checking (Windows on Arm has matured a lot by 2026).
Recommended AI models (newest first):
| Spec | Value |
|---|---|
| NPU | ~50 TOPS |
| CPU | Up to 18 cores (8 P + 8 E + 2 LP E), Panther Cove cores |
| GPU | Xe3 (Celestial), up to 12 cores |
| Process | Intel 18A |
| Availability | Laptops from late 2025–2026 (CES 2026 wave) |
| Price (Aug 2026) | ~$999–2,400 (mainstream ~$1,100–1,500; high-end >$2,000) |
Best for: The 2026 Intel answer — combines the fast P-cores with a bigger NPU; leads in single-thread perf among 2026 AI chips.
Recommended AI models (newest first):
| Spec | Value |
|---|---|
| NPU | 60 TOPS (XDNA2 refresh) |
| CPU | Zen 5 (up to 12 cores) |
| GPU | RDNA (up to 16 CU) |
| Availability | Laptops + first Copilot+ desktops, early 2026 |
| Price (Aug 2026) | ~$750–1,500 (mainstream ~$900–1,200) |
Best for: Highest NPU throughput in the AMD lineup; also powers the desktop PCs covered in ai-desktops.md.
Recommended AI models (newest first):
| Spec | Value |
|---|---|
| NPU | 80 TOPS (Hexagon) — Elite Extreme up to 80–85 TOPS |
| CPU | Elite: up to 18 Oryon v3 cores (12 Prime + 6 Perf), up to 5.0 GHz; Plus: 10 cores |
| GPU | Adreno X2-90 (~1.85 GHz) |
| Process | 3 nm |
| Availability | Laptops from late 2025–2026 (CES 2026 wave) |
| Price (Aug 2026) | ~$1,080–2,050 (Surface 13.8″ from $1,399; premium 16″ ~$1,600–1,700; HP first X2 Elite from ~$2,050; deals to ~$1,080) |
Best for: The fastest consumer NPU in 2026 — the Extreme config leads the Copilot+ class, and battery life remains class-leading.
Recommended AI models (newest first):
| Chip (gen) | NPU TOPS | Year | Newest NPU sweet-spot models | 8B Q4 speed (iGPU/CPU) | Laptop street price (Aug 2026) |
|---|---|---|---|---|---|
| Lunar Lake | 48 | 2024–25 | Ministral 3 3B, Qwen3.5/Gemma 4 sub-12B, Voxtral | 15–25 tok/s | ~$600–1,400 |
| Strix Point | 50 | 2024–25 | Ministral 3 3B/8B, Qwen3.5/Gemma 4 sub-12B | 25–40 tok/s | ~$900–1,500 |
| Snapdragon X Elite | 45 | 2024–25 | Ministral 3 3B, Qwen3.5 sub-12B | 15–25 tok/s | ~$600–1,700 |
| Panther Lake | ~50 | 2025–26 | Ministral 3 3B/8B | 20–35 tok/s | ~$999–2,400 |
| Ryzen AI 400 | 60 | 2026 | Ministral 3 8B, dual workloads | 25–40 tok/s | ~$750–1,500 |
| Snapdragon X2 | 80–85 | 2026 | up to 8B on NPU | 25–35 tok/s | ~$1,080–2,050 |
Rule of thumb: if your goal is local LLMs, prioritize RAM (16 GB minimum, 32 GB ideal) and memory bandwidth over NPU TOPS — and load the newest small models (Ministral 3, Qwen3.5, Gemma 4). If your goal is always-on AI assistants (voice, transcription, background agents), the NPU number is what matters — that’s where Snapdragon X2 and Ryzen AI 400 shine.