“AI PC” has become a marketing label on almost every new laptop, which makes it hard to tell what hardware actually matters versus what’s just a sticker. Here’s a practical breakdown of what each component does for on-device AI workloads, and what’s genuinely worth paying for.
NPU: The New Piece of the Puzzle
An NPU (Neural Processing Unit) is a dedicated chip for running AI inference efficiently — it’s what lets a laptop run background tasks like live captioning, webcam background blur, or small local AI models without draining the battery the way a GPU would for the same workload. Microsoft’s Copilot+ PC certification requires an NPU rated at 40+ TOPS (trillion operations per second) specifically so Windows can offload its built-in AI features to it instead of the CPU or GPU. If you don’t use OS-level AI features or run local AI apps, a high-TOPS NPU mostly sits idle — it’s a real feature, but only valuable to a workload that actually uses it.
RAM: Usually the Actual Bottleneck
If you plan to run local large language models (via tools like Ollama or LM Studio), RAM capacity matters more than almost anything else, because the entire model typically has to fit in memory. A 7-billion-parameter model quantized to 4-bit needs roughly 4-5GB just for the model weights, before accounting for your OS and other apps — 16GB total system RAM is a realistic floor for comfortably running small-to-mid local models, and 32GB gives real headroom. This is the single most common upgrade people skip and then wonder why local AI tools are sluggish or crash.
CPU vs GPU vs NPU: Who Does What
- CPU: general-purpose, handles everything by default including AI workloads if nothing else is available — flexible but power-hungry and slower for AI-specific math.
- GPU: parallel processing designed for exactly the matrix math AI models need — fastest option for heavier local AI work (larger models, image generation), but draws significantly more power and is the main one most laptop AI apps actually target for acceleration.
- NPU: purpose-built for sustained, low-power AI inference — slower peak throughput than a GPU for heavy workloads, but far more power-efficient for the lighter, always-on AI features operating systems now ship with.
Does PCIe 5.0 / DDR5 Actually Matter Here?
DDR5 over DDR4 gives meaningfully higher memory bandwidth, which does help AI workloads that are memory-bandwidth-bound (loading model weights quickly matters for responsiveness). PCIe 5.0 mostly matters if you’re using a discrete GPU or extremely fast NVMe storage for large model files — for most laptop buyers it’s a smaller, secondary factor behind RAM capacity and the GPU/NPU itself.
A Practical Buying Checklist
- RAM first. 16GB minimum if you’ll run any local AI tools at all; 32GB if you want real headroom for bigger models or multitasking alongside them.
- Dedicated GPU with real VRAM if you plan to run larger local models or do AI image generation — 8GB+ VRAM is a reasonable starting point, more for anything serious.
- NPU is a “nice to have,” not the deciding factor unless you specifically rely on OS-level AI features that are built to use it.
- Fast NVMe storage (Gen 4 is plenty for most people) so loading multi-gigabyte model files doesn’t become the actual bottleneck.
Key Takeaways
- “AI PC” branding and raw NPU TOPS numbers matter far less than RAM capacity for most real local-AI use cases.
- A GPU still outperforms an NPU for heavier local AI work; the NPU’s advantage is power efficiency for lighter, always-on tasks.
- DDR5 and PCIe 5.0 are real but secondary factors — don’t let them outrank RAM capacity or GPU VRAM in a buying decision.
- Decide based on what you’ll actually run (which local models, how often) rather than the sticker on the box.