Here is the uncomfortable truth about AI laptops: the RAM is almost certainly soldered to the motherboard, and no screwdriver, YouTube tutorial, or repair shop can change its capacity. If you bought 16 GB, you own 16 GB forever. Shoppers coming from older laptops — where popping in a second SODIMM was a Saturday-afternoon upgrade — find this out at the worst possible moment: when the machine starts feeling tight and the “upgrade” they planned turns out to be impossible.

The neural processors in Copilot+ PCs demand memory bandwidth that socketed RAM physically cannot deliver, so fast LPDDR5X chips sit millimeters from the processor, soldered to the board. Knowing why it is soldered, how to check your model, and how to buy right the first time is the difference between a laptop that lasts five years and one that feels cramped in eighteen months. If you are still deciding how much memory to get, read our breakdown of how much RAM an AI PC really needs before you commit.

This guide covers six stages: confirming what your machine has, buying correctly, trimming memory waste, tuning virtual memory, offloading pressure to the cloud, and planning around a fixed ceiling.

Why this happens

  • The NPU needs bandwidth sockets cannot provide: Copilot+ NPUs and integrated GPUs feed on 100+ GB/s of memory bandwidth. LPDDR5X soldered next to the processor delivers it; SODIMM slots add trace length, latency, and signal limits that cap speeds far lower.
  • Unified memory architecture: CPU, GPU, and NPU all share one memory pool on AI laptops. Soldering keeps that pool fast and power-efficient, which is why battery life stays strong during AI workloads.
  • Thin designs leave no room: Soldered chips are flatter than slots plus modules. The thin-and-light chassis AI laptops compete in simply have no vertical space for upgradeable memory.
  • 16 GB is the Copilot+ floor, not the ceiling: Microsoft’s minimum gets you in the door, but local model work, creative apps, and heavy multitasking push past it — and you cannot add more later, so the purchase decision is permanent.
  • BIOS locks the configuration: Even if someone physically replaced the chips, the firmware memory training is tuned for the factory capacity. The board would not boot with a different size — upgradeability is closed at every layer.
  • Faster, not slower: Soldered LPDDR5X runs at speeds socketed DDR5 cannot touch while sipping power. You lose the upgrade slot but gain the performance the AI features were designed around.

How to fix it

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

Fix 1: Confirm whether your RAM is actually soldered

Check your specific model first. A small minority of AI laptops still ship one SODIMM slot alongside soldered memory, and knowing what you have sets the rest of the plan.

  1. Look up your exact model on the manufacturer’s spec page and find the memory line — it will say “soldered,” “onboard,” “LPDDR5X,” or list a SODIMM slot if one exists.
  2. Open Task Manager > Performance > Memory and read “Slots used” — “1 of 1” or a grayed-out slots field on an LPDDR machine means no upgrade path.
  3. Check the service manual if the maker publishes one; it shows whether a memory slot exists under the bottom panel before you ever pick up a screwdriver.
  4. If you do find a free slot, buy a matching SODIMM (same DDR generation and speed) and install it with the battery disconnected — the one genuinely upgradeable case.

Fix 2: Buy one tier above what you think you need

Since the decision is permanent, the buying strategy flips: today’s cheapest adequate configuration is tomorrow’s most expensive mistake, because memory pressure only grows.

  1. Treat 16 GB as the minimum for light use, 24 GB as the sweet spot, and 32 GB as the right call for creators, developers, and four-year keepers.
  2. Price the jump at purchase: the step from 16 to 32 GB usually costs far less than replacing the entire laptop a year early because memory ran out.
  3. Do not count on “the cloud will handle it” for memory — local AI features, large browser sessions, and creative apps consume RAM on the device no matter how good your internet is.
  4. If budget forces a choice between more RAM and more storage, choose RAM — storage can be extended with external SSDs, while memory cannot be extended at all.

Fix 3: Trim the memory waste you can control

You cannot add RAM, but you can stop leaking it. Most laptops waste gigabytes on unasked-for software, and reclaiming that headroom is free performance.

  1. Open Settings > Apps > Startup and disable everything that does not need to launch at boot — updaters, chat clients, and game launchers are the usual squatters.
  2. Uninstall trialware and duplicate utilities from Settings > Apps > Installed apps; each background service you remove returns memory to the shared pool the NPU draws from.
  3. Tame the browser: enable tab sleeping, keep only the extensions you use, and close the forty-tab “to read” pile — browsers are the biggest RAM consumer on most machines.
  4. Restart the laptop at least weekly; long uptimes accumulate memory leaks in background processes that only a fresh boot clears.

Fix 4: Tune virtual memory instead of fighting it

When RAM fills, Windows pages to the SSD — slower than RAM but far better than crashing, and less painful than its reputation on modern fast SSDs.

  1. Open System > About > Advanced system settings > Performance Settings > Advanced > Virtual memory and confirm it is system-managed on your fastest internal SSD — Microsoft’s Windows support hub documents the dialog if your build labels it differently.
  2. Leave it system-managed unless you have a specific reason — Windows sizes the pagefile dynamically, and manual caps usually cause more problems than they solve.
  3. Keep 15–20% of your SSD free; a nearly full drive chokes paging and accelerates wear.
  4. Understand the limit: paging keeps apps alive under pressure, but a machine constantly paging feels sluggish — it is a safety net, not a RAM upgrade.

Fix 5: Offload pressure with cloud and external storage

RAM pressure often comes from keeping everything local. Moving bulk data off the machine cuts the memory spent indexing, syncing, and caching files you rarely open.

  1. Turn on OneDrive Files On-Demand so cloud files show in Explorer without occupying local disk or sync memory until you open them.
  2. Move media libraries, archives, and old projects to an external SSD; keep only active work on the internal drive.
  3. If local AI model downloads keep failing or crawling on a tight machine, read our guide on Windows AI Foundry model download failures — low headroom is a common hidden cause.
  4. For genuinely heavy work — large model training, 8K timelines — pair the laptop with a desktop or cloud workstation instead of asking a thin ultrabook to be something it is not.

Fix 6: Plan your replacement cycle around the ceiling

A fixed-memory laptop has a predictable lifespan: the day your workload permanently exceeds RAM, its useful life ends. Planning for that date beats being surprised by it.

  1. Watch Task Manager > Performance > Memory during your heaviest typical session; if committed memory routinely exceeds physical RAM, you are living on paging.
  2. Note that some background features also consume memory quietly — if Click to Do’s send-to-email action misbehaves , for instance, troubleshoot the feature rather than assuming memory is the cause.
  3. For your next purchase, buy the memory tier for your year-three workload, not today’s.
  4. Sell or hand down the old machine while it holds value — resale offsets the upgrade.

Closing

Soldered RAM is the price of the NPU’s speed: the bandwidth AI features need only exists with chips fused to the board. You cannot change capacity afterward, so the game is played at purchase — buy one tier above your current need — and in daily habits: trim startup waste, tame the browser, keep SSD breathing room, and offload bulk data. Do that, and fixed memory feels like what it is: a fast, efficient design choice you planned around.

Frequently asked questions

Can soldered RAM ever be upgraded?
Not by any practical means. Soldered LPDDR chips require board-level rework with specialist equipment, and the BIOS will not recognize a different capacity anyway. For all real-world purposes, the RAM your AI laptop ships with is the RAM it keeps.
Why do AI laptops use soldered RAM instead of slots?
The NPU and integrated GPU need extreme memory bandwidth — LPDDR5X at 7500 MT/s and beyond — which only works with chips soldered close to the processor. Sockets add distance, latency, and power draw that would cripple AI performance.
Is 16 GB enough for a Copilot+ PC?
It meets the official minimum and handles everyday AI features fine. But 16 GB gets tight once you run local AI models, heavy multitasking, or creative apps alongside NPU workloads — which is why 24 GB or 32 GB is the smarter buy if you keep laptops for years.
Does soldered RAM perform worse than slot RAM?
The opposite. Soldered LPDDR5X runs at much higher speeds and lower voltage than socketed DDR5 SODIMMs, and the whole system — CPU, GPU, and NPU — shares it at full bandwidth. You trade upgradeability for speed.