Local AI hardware

AI PCs — NPU laptops vs GPU desktops vs Mac

Honest trade-offs between Copilot+ NPUs, discrete-GPU towers, and Apple silicon for local open-weight models — with links to GPUs, Macs, and cloud inference.

GPU catalog Cloud GPU inference Mac buying guide Open-weight models

The planning rule

Pick the largest model you need to run locally, then buy the memory architecture that can hold it — NPU marketing comes after that.

Reviewed against Microsoft, Qualcomm, AMD, NVIDIA, Framework and Apple first-party material on October 2, 2026. Capacity uses the site's GPU/unified-memory planning arithmetic, not measured fit or benchmark rankings. Announced future configurations are identified separately. Current retail prices are unverified; Amazon links are product-class searches.

Amazon UK search links are provided for current availability; no affiliate programme is currently configured, and no live price is claimed here. Affiliate disclosure.

Form factors

Four AI PC paths — and what each is actually for

NPU TOPS and GPU TFLOPS are not interchangeable scores. Memory capacity is the first filter.

Copilot+ / NPU laptop

Compute: CPU + integrated graphics + dedicated NPU (vendor-quoted TOPS)
Memory: Shared system memory; Microsoft's Copilot+ minimum is 16GB RAM and 256GB storage
Local AI: Supports device-specific Windows AI features; open-weight inference depends on memory, model and supported CPU/GPU/NPU runtime
Gaming: Depends on the exact graphics hardware, game and settings
Portability: High — battery-powered daily driver

Search Amazon UK

Discrete-GPU desktop

Compute: Dedicated GPU VRAM + CPU — CUDA/ROCm/oneAPI depending on card
Memory: Separate system RAM and GPU VRAM; CPU offload and multi-GPU support depend on the runtime
Local AI: A configurable option for GPU inference when the selected card and runtime support the model
Gaming: Can also run games; performance depends on the exact card, game and settings
Portability: None — tower or SFF on a desk

Search Amazon UK

Apple Silicon Mac

Compute: Unified memory shared by CPU, GPU, and Neural Engine
Memory: Shared CPU/GPU memory pool; model, quantisation, context, and other apps determine practical capacity
Local AI: Large unified-memory configurations can accommodate some medium models; actual fit and speed depend on model, context, runtime, and configuration
Gaming: Check macOS availability and performance for the games you use
Portability: High on MacBook; Mac mini / Studio are desk-bound

Search Amazon UK

Compact GPU workstation / mini PC

Compute: Integrated graphics with shared memory or a discrete GPU, depending on the system
Memory: Shared system memory or separate GPU VRAM; capacity and upgrade options depend on the exact configuration
Local AI: An option for local inference with a smaller desk footprint; assess the model, memory and backend
Gaming: Depends on the processor/GPU, power limits, game and settings
Portability: Low — small desk footprint, not a laptop

Search Amazon UK

AI PC form factor comparison
Attribute	Copilot+ / NPU laptop	Discrete-GPU desktop	Apple Silicon Mac	Compact GPU workstation / mini PC
Compute	CPU + integrated graphics + dedicated NPU (vendor-quoted TOPS)	Dedicated GPU VRAM + CPU — CUDA/ROCm/oneAPI depending on card	Unified memory shared by CPU, GPU, and Neural Engine	Integrated graphics with shared memory or a discrete GPU, depending on the system
Memory	Shared system memory; Microsoft's Copilot+ minimum is 16GB RAM and 256GB storage	Separate system RAM and GPU VRAM; CPU offload and multi-GPU support depend on the runtime	Shared CPU/GPU memory pool; model, quantisation, context, and other apps determine practical capacity	Shared system memory or separate GPU VRAM; capacity and upgrade options depend on the exact configuration
Local AI	Supports device-specific Windows AI features; open-weight inference depends on memory, model and supported CPU/GPU/NPU runtime	A configurable option for GPU inference when the selected card and runtime support the model	Large unified-memory configurations can accommodate some medium models; actual fit and speed depend on model, context, runtime, and configuration	An option for local inference with a smaller desk footprint; assess the model, memory and backend
Gaming	Depends on the exact graphics hardware, game and settings	Can also run games; performance depends on the exact card, game and settings	Check macOS availability and performance for the games you use	Depends on the processor/GPU, power limits, game and settings
Portability	High — battery-powered daily driver	None — tower or SFF on a desk	High on MacBook; Mac mini / Studio are desk-bound	Low — small desk footprint, not a laptop
Price check	Search Amazon UK (opens in new tab)	Search Amazon UK (opens in new tab)	Search Amazon UK (opens in new tab)	Search Amazon UK (opens in new tab)

Comparison notes last reviewed October 2, 2026. Figures describe product classes, not benchmark winners.

NPU vs GPU — different jobs

Copilot+ NPUs accelerate vendor-defined on-device tasks such as live captions and background blur. NPU TOPS do not establish that a system can load a particular open-weight model: this site estimates a 32B Q4_K_M weight set at about 19.4GB before its planning allowances. Compare memory, supported runtime, and model settings rather than NPU TOPS with GPU TFLOPS.

When cloud inference is the honest answer

If the target model exceeds your local capacity, compare a compatible cloud GPU or API with an upgraded local configuration. Include usage, purchase cost, electricity, hosting charges and data requirements; no option is universally cheaper. Provider/model availability and terms need separate checks.

Software stack matters as much as silicon

Check whether the selected llama.cpp, Ollama, vLLM, MLX or vendor runtime supports your model architecture, quantization and device. Model loading and performance depend on backend, context, cache and offload settings. Compare measured results for the exact configuration rather than ranking devices by TFLOPS.

Best for

Who each path suits

Match the hardware class to the workload — NPU laptops, discrete GPUs, and Macs solve different jobs. Each card ends with the catch to check before buying.

Copilot+ / NPU laptop

Windows laptops with Copilot+ features
Meetings, notes, and lightweight on-device AI
Buyers who will rent cloud inference for big models

Watch out

Microsoft's Copilot+ NPU floor is 40+ TOPS. That badge does not establish local-model capacity or inference speed. Qualcomm lists up to 85 TOPS for Snapdragon X2 Elite, with SKU-dependent 80/85 figures; its captured table and footnote disagree for X2E-96-100, so check the exact specification. NPU TOPS, GPU TFLOPS and measured tokens per second are different quantities. Windows features vary by device, region and update. Confirm memory capacity, runtime support and the exact model pack before buying; no current retailer price is verified here.

Discrete-GPU desktop

Running 7B–70B open models locally (VRAM permitting)
Gamers who also want local inference
Upgradeable buyers who can swap GPUs later

Watch out

A model kept entirely on one GPU must share its VRAM with cache and runtime allocations. CPU offload and multi-GPU inference can change memory requirements and performance. Verify the exact GPU, driver/backend, power supply, chassis and cooling. Laptop and desktop GPUs with the same tier name have different specifications. Compare actual local costs with cloud usage; no price or performance winner is established here.

Apple Silicon Mac

Developers already on macOS
Quantised models whose estimated memory fits the chosen unified-memory configuration; capacity arithmetic is not a runtime or performance guarantee
Quiet desk inference without building a tower

Watch out

Choose unified-memory capacity when purchasing; it is shared with the OS and applications. Apple's current Studio specifications list M5 Max up to 128GB and M5 Ultra up to 512GB. Apple announced the 512GB configuration for late October 2026; it is a future configuration as of October 2. M5 Max bandwidth varies with GPU configuration (460 or 614GB/s). The site's 75% available-memory rule is a planning assumption, not Apple's guaranteed GPU allocation. Check the exact model pack and Metal/MLX/runtime support; memory capacity does not prove speed or value.

Compact GPU workstation / mini PC

Homelab and inference beside the main desk
Buyers who cannot host a full tower
Secondary inference node while a laptop is the daily driver

Watch out

Read the exact system specification; compact systems do not share one memory or upgrade architecture. Framework lists 32/64/128/192GB Desktop configurations and Ryzen AI Max+ PRO 495. AMD specifies up to 192GB system memory and up to 160GB GPU allocation for PRO 495; its memory bus/rate imply about 273GB/s theoretical peak. This site's 75% rule instead budgets 144GB from 192GB, and neither number certifies a model's fit. Framework's 163GB DeepSeek V4.1 Flash Q2 reference uses specified runtime/context conditions; it does not establish full GPU residency within AMD's 160GB allocation. NVIDIA announced Windows RTX Spark systems with a 20-core Grace CPU and up to 128GB for October 2026; this is distinct from DGX Spark's GB10 CPU specification. Confirm shipping availability, sustained-load behavior and purchase price.

Common questions

AI PC buying questions

Answered from the verified figures on this page rather than general guidance.

Is a Copilot+ PC enough to run local LLMs?

The badge establishes eligibility for specified Windows AI features, not a universal local-LLM capability. Some quantized open-weight models can run on compatible CPU, integrated graphics or NPU backends. Check the exact memory configuration, model pack, runtime and context; the minimum specification does not establish large-model fit.

Mac or PC for local AI in 2026?

Choose from the exact model pack, memory requirements, compatible runtime and workload. Macs offer shared unified memory; PCs include integrated-memory and discrete-GPU configurations with different upgrade options. Compare current system prices and measured performance for the selected configuration. No universal cost or speed winner is established here.

How much RAM or VRAM do I need?

Use the GPU catalog as a rough estimate: parameter count × effective quantisation bits ÷ 8 plus the site's planning allowances. Its 32B Q4_K_M estimate is about 22.3GB. This is not a model-specific KV-cache or runtime guarantee; 16GB systems remain useful for smaller quantised models but will not fit every target.

Should I build a PC or buy prebuilt for AI?

Compare the exact GPU and memory configuration, power supply, cooling, available slots, warranty and total price. A system-RAM headline does not establish GPU VRAM capacity. Building and buying prebuilt both require checking the intended model and runtime.

Keep exploring

Go deeper on the stack

VRAM fit, unified memory, and rental inference each answer a different constraint.

GPU VRAM fit chart Mac unified memory Rent cloud GPUs Build a GPU desktop Prebuilt gaming PCs Open-weight catalog GPU finder tool Laptop finder