RAM is the quiet gatekeeper of the AI PC era. Everyone obsesses over NPU TOPS, but it is memory that determines whether those AI models actually run — or stutter, or refuse to load at all. Microsoft drew the line at 16 GB for Copilot+ PCs, and unlike some marketing numbers, this one is grounded in how AI workloads physically work.

The reason is architectural. An AI model is a giant file of numbers (weights) that must sit in memory while the feature runs. Windows’ on-device models are gigabytes in size, and they share RAM with everything else your laptop is doing. Drop below the floor and the system starts paging to disk — at which point your “AI PC” feels slower than a normal laptop. For the full hardware picture, see our Copilot+ PC requirements ; for the buying context, our AI laptop buying guide .

Here is how to think about RAM for AI PCs, honestly.

Why this happens

  • Resident models: On-device AI keeps multi-gigabyte models loaded in RAM continuously.
  • Shared memory: The NPU, CPU, and GPU all draw from the same system RAM pool on most laptops.
  • Multitasking reality: AI features run alongside browsers, Office, and calls — not instead of them.
  • Future growth: Models get larger over time; today’s comfortable margin is tomorrow’s minimum.
  • Soldered decisions: Most AI laptops solder RAM, making the purchase choice permanent.

How to fix it

Work through these in order — diagnose first, then act.

Fix 1: Understand the 16 GB floor

Why Microsoft chose this number, and why it is real.

  1. Windows itself needs roughly 4 GB for comfortable operation with modern apps open.
  2. A browser with real tabs open easily claims another 2-4 GB.
  3. On-device AI models (the ones behind Recall, live translation) need several GB resident.
  4. Add headroom for spikes and you land at 16 GB as the genuine minimum for AI features plus normal work.
  5. At 8 GB, enabling AI features means constant paging — the machine feels worse, not smarter.
  6. This is why the Copilot+ spec mandates 16 GB: below it, the features cannot deliver their promised experience.

Fix 2: Decide if you need 32 GB

The step up is real but narrow in audience.

  1. You need 32 GB if: you run local LLMs (7B+ parameter models), Stable Diffusion, or heavy ML development.
  2. You need 32 GB if: you edit 4K video or huge Photoshop files while AI features run in the background.
  3. You do not need 32 GB for: Recall, Studio Effects, Live Captions, Copilot chats, or any built-in Windows AI feature — 16 GB handles all of these.
  4. Price the difference: if 32 GB costs little extra on your chosen model, it is cheap future-proofing.
  5. If it costs a lot, skip it — put the money toward a better screen or larger SSD instead.
  6. Our do I need an NPU laptop guide pairs with this: match memory to your actual AI workload, not anxiety.

Fix 3: Check whether your RAM is upgradeable

Before worrying, see if the question is even permanent.

  1. Find your laptop’s service manual on the manufacturer’s support site.
  2. Look for SODIMM slots versus “soldered” / “onboard” memory.
  3. Two SODIMM slots: fully upgradeable — buy RAM later if needed, no stress now.
  4. One soldered + one slot (common on some models): partially upgradeable.
  5. Fully soldered (most thin-and-lights): the purchase-time choice is forever — choose carefully now.
  6. If upgradeable and you are at 8 GB, a 16 GB kit is one of the cheapest meaningful upgrades in computing.

Fix 4: Measure your actual usage before deciding

Data beats speculation. See what you really consume.

  1. Open Task Manager > Performance > Memory during your heaviest normal workload.
  2. Note the “In use” figure and the percentage. Sustained usage above 80% means you are tight.
  3. Now enable the AI features you use (Studio Effects on a call, Live Captions) and watch the number climb.
  4. If 16 GB keeps you under ~75% at your heaviest, you are comfortable.
  5. If you are buying and cannot test, assume your usage will grow — software bloat is the one constant.
  6. Students and office workers: 16 GB. Creators and local-AI tinkerers: 32 GB. That covers nearly everyone.

Fix 5: Optimize what you have

If you are stuck at 8 GB or a tight 16 GB, make every gigabyte count.

  1. In Task Manager > Startup apps, disable everything you do not need at boot — each one eats resident RAM.
  2. Browser discipline matters most: each tab is 100-300 MB. Use tab sleep features and close what you are not reading.
  3. Disable AI features you do not use — Recall snapshots and background AI services consume memory even idle.
  4. In Settings > Apps, check background app permissions and set unused apps to Never.
  5. Consider a clean Windows install if the machine came bloated — OEM bloatware is a RAM tax.
  6. These optimizations recover gigabytes, but they cannot conjure what soldered 8 GB lacks for real AI work.

Fix 6: Buy right next time

RAM regrets last the life of the laptop. Avoid them.

  1. Default to 16 GB for any AI-capable laptop purchase in 2026 — treat 8 GB configurations as non-starters.
  2. Choose 32 GB if you identified a real need in Fix 2, or if the price gap is small.
  3. Prefer models with at least one SODIMM slot if upgradability matters to you — check before buying, not after.
  4. Do not let a low price on an 8 GB “AI laptop” tempt you — the NPU cannot compensate for starved memory.
  5. When comparing two models, RAM differences outweigh small NPU TOPS differences for real-world smoothness.
  6. And if the budget forces a choice between 32 GB RAM and a bigger NPU, take the RAM — it benefits every single task, AI or otherwise. For the full value question, see is a Copilot+ PC worth it .

Closing

RAM for AI PCs is refreshingly uncomplicated: 16 GB is the real floor, 32 GB is for defined heavy workloads, and 8 GB is a trap no NPU can rescue. Because most AI laptops solder their memory, this is a buy-once decision — so make it with your actual usage in mind, not the sticker. Get the memory right and the NPU has room to do its job; get it wrong and even 50 TOPS will feel sluggish. On Snapdragon X laptops specifically, RAM pressure can also masquerade as app failures — see app not working on Snapdragon X laptop if programs misbehave.

Frequently asked questions

Is 8 GB of RAM enough for an AI laptop?
No — not for AI features. Microsoft’s Copilot+ certification requires 16 GB because on-device AI models need memory to live in. An 8 GB machine can browse and stream, but AI features will stutter or refuse to run.
Does AI use RAM differently from normal apps?
Yes. AI models load large weight files into memory and keep them resident for instant response. Unlike a document you open and close, an AI model sits in RAM the whole time its feature is active, which is why the floor is higher.
Is 32 GB overkill for most people?
For built-in Windows AI features, yes — 16 GB is plenty. 32 GB matters only if you run local LLMs, image generation, or heavy creative AI workloads yourself. Buy 32 GB for a purpose, not for peace of mind.
Can I upgrade RAM later if 16 GB is not enough?
Only if your laptop has SODIMM slots rather than soldered memory. Most thin-and-light AI laptops solder RAM, making the purchase-time choice permanent. Check the service manual before buying if upgradability matters to you.