RAM is the spec most buyers get wrong. People obsess over the CPU, argue about the GPU, and then quietly accept whatever memory configuration the laptop shipped with — usually the cheapest one.

That was a survivable mistake five years ago. It isn't anymore. Two things changed in 2026: software got hungrier (local AI tools now run on ordinary laptops), and memory got expensive (AI data centers are eating the world's DRAM supply). Buying too little RAM today means living with it for the life of the machine, because topping it up later costs far more than it used to.

Here's what you actually need, tier by tier.

1. What RAM Actually Does (30-Second Version)

Random access memory is your computer's working desk. Your SSD is the filing cabinet — big, permanent, comparatively slow. RAM is the surface you spread things out on while you work: the operating system, every open browser tab, your IDE, that Photoshop file, the model weights your local chatbot is running on.

When the desk fills up, the system starts shuffling data back and forth to the SSD to make room. That shuffling is what you feel as stutter, beachballs, and tab reloads. More RAM doesn't make your computer faster in a straight line — it just delays the point where it falls off a cliff.

Two consequences worth internalizing:

  • Unused RAM isn't wasted, but over-buying is. Modern operating systems cache aggressively, so spare capacity does get used. Still, buying 64GB to browse the web is money that belongs in a faster SSD or GPU.
  • RAM clears when you power off. It's volatile storage. Nothing you have "open" is actually saved until it's written to disk.

2. How to Check How Much RAM You Have

Before you buy anything, find out what you're working with.

Windows 11 / Windows 10: Settings → System → About. Look under Device specifications for Installed RAM. For live usage, open Task Manager (Ctrl + Shift + Esc) → Performance → Memory. That view also shows your speed in MT/s and how many slots are occupied — critical if you're planning an upgrade.

macOS: Apple menu → About This Mac. Apple silicon uses unified memory shared between the CPU and GPU, which matters enormously for local AI (more on that below). Note that it's soldered — what you buy is what you keep.

Linux: Run free -h in a terminal for a human-readable summary, or sudo dmidecode --type memory for module-level detail including speed and empty slots.

The number that actually matters: open everything you'd normally have running on a busy day, then check memory usage. If you're consistently above 80%, you're already constrained.

3. The 2026 Complication: RAM Got Expensive

This is the part most buying guides published before 2026 completely miss.AI infrastructure has consumed the memory market. Samsung, SK Hynix and Micron — who between them make effectively all the world's DRAM — reallocated wafer capacity toward high-bandwidth memory for AI accelerators, because the margins are dramatically better. Consumer DDR4 and DDR5 got squeezed out of the production queue.

The result was a price shock with no modern precedent. Contract prices for conventional DRAM rose roughly 50–55% quarter-over-quarter entering 2026, then accelerated further through Q2. Retail followed: 32GB DDR5 kits that sold for under $90 in early 2025 were commanding several hundred dollars a year later, and even ageing DDR4 doubled. Micron retired its consumer-facing Crucial brand in February 2026 to focus on data-center products — a fairly blunt signal about where priorities sit. Analysts expect the crunch to ease only as new fabrication capacity comes online, likely 2027 at the earliest.

Two practical takeaways:

  1. Buy the RAM you need at purchase time, not later. On a soldered-memory laptop this was always true. In 2026 it's true for desktops too, because the upgrade you defer will cost more than the upgrade you make today.
  2. Don't panic-buy capacity you'll never touch. The price environment cuts both ways — overspending on 64GB you don't need is a genuinely expensive mistake right now.

4. How Much RAM Do You Need? Tier by Tier

8GB — The Floor, and It's Cracking

Eight gigabytes still technically works for a narrow use case: a handful of browser tabs, email, documents, streaming video. That's it. Windows itself claims 4–6GB before you open anything, which leaves very little room to work.

If you're buying new in 2026, treat 8GB as a red flag rather than a budget option. It's the configuration you'll regret within eighteen months.

16GB — The Sensible Standard

This is the real baseline now, and it's where most people should land. Sixteen gigabytes handles heavy multitasking, thirty-plus browser tabs, video calls with everything else still running, Office workloads, photo editing, and the great majority of modern games — most 2026 AAA titles list 16GB as minimum spec.

It's also the practical entry point for AI features. Copilot+ PCs require 16GB, and on-device processing needs headroom even when a dedicated NPU is doing the inference work.

If you're a student, a remote worker, or a normal person who wants a laptop that stays pleasant for four years, buy 16GB.


32GB — Power Users, Creators, and Anyone Serious About Local AI

Thirty-two gigabytes is where you go if your work is the reason you bought the machine. Video editing, 3D and design software, large codebases with a couple of Docker containers running, virtual machines, streaming while gaming, or running local language models alongside everything else.

For local AI specifically, 32GB has become the sensible floor rather than a luxury. After the OS takes its cut you're left with roughly 26GB of usable headroom — enough for an 8–14B parameter model plus the browser, editor and pipeline tooling around it.

64GB — Genuinely Heavy Workloads

Sixty-four gigabytes stops being overkill in three situations: 4K and 8K video work where you want to keep editing while renders churn, large-scale data work in memory, and running big models with CPU offloading.

That last one is the newest justification. When a model doesn't fit entirely in your GPU's VRAM, the overflow layers spill into system RAM — and community benchmarks consistently show 30–60% better effective throughput moving from 16GB to 64GB on offload-dependent workloads. If you're routinely running 70B-class models locally, this is your tier.

128GB and Above — Specialists Only

Aggressive offloading of very large mixture-of-experts models, serious virtualization labs, scientific computing. If you belong in this bracket you already know it, and you're not reading a general buying guide to find out.



6. System RAM vs VRAM: Don't Confuse Them

If you have a discrete graphics card, it carries its own dedicated video memory. That VRAM holds textures, frame buffers, and — increasingly — AI model weights.

For local AI, VRAM is usually the binding constraint. A rough rule: a model needs about 2GB of VRAM per billion parameters at FP16. Quantization cuts that sharply — Q8 roughly halves it, Q4 roughly quarters it — which is why a 13B model that would need 28GB at full precision fits comfortably on a 16GB card at Q8. Budget an extra 10–20% on top for the KV cache as your context window fills.

System RAM handles everything else: the OS, your applications, the retrieval pipeline and vector store, and any model layers that don't fit on the GPU. If you have a large GPU, prioritize VRAM. If you don't, system RAM is what determines what you can actually run.

Apple silicon sidesteps the distinction with unified memory — the CPU and GPU share one pool. It's why a 32GB or 64GB MacBook punches well above its weight for local inference, and why the memory tier you pick at checkout matters so much on a machine you can never upgrade.

7. Should You Buy Now or Wait for Prices to Fall?

The honest answer: don't wait on a machine you need today.

Supply constraints are structural, not seasonal. New DRAM fabs take three to five years to build, and manufacturers are wary of overbuilding into what could turn out to be an AI demand bubble. Forecasts point to elevated pricing persisting into 2027 and possibly beyond. The rate of increase has moderated — Q3 2026 projections came in around 13–18% quarter-over-quarter, well down from the roughly 60% jumps earlier in the year — but that's a slowdown in the climb, not a reversal.

If you're buying a laptop, configure the memory you need at order time. If you're building a desktop, buy the full kit at once rather than planning to add sticks later; mixing modules across a price spike often means mismatched speeds anyway.

8. How to Get More Out of the RAM You Already Have

Not upgrading right now? A few things genuinely help:

  • Audit startup programs. Task Manager's Startup tab on Windows, Login Items on macOS. Most people have several apps loading at boot that they use once a month.
  • Treat browser tabs as processes, not bookmarks. Chromium-based browsers spawn a process per tab. Tab suspension extensions, or simply closing things, recover more memory than any "RAM booster" utility.
  • Check for a memory leak before blaming capacity. Sort by memory usage in Task Manager. One misbehaving application often accounts for most of the pressure.
  • Ignore RAM cleaner apps. They force the OS to dump caches it deliberately built. You get a lower number in a widget and slightly worse performance.
  • Make sure dual channel is actually enabled. A single 16GB stick is meaningfully slower than two 8GB sticks on most platforms. Task Manager shows your slot configuration.
  • Verify XMP/EXPO is on. Plenty of DDR5 systems run at conservative default speeds because the memory profile was never enabled in BIOS.

9. The Short Version

For most people in 2026, 16GB is the standard and 32GB is the sweet spot. Buy 16GB if you're a normal user who wants four good years out of the machine. Buy 32GB if you edit video, write code with containers running, game while streaming, or want to run local models without constantly closing things. Go to 64GB only if a specific workload — large-model offload, high-resolution video, virtualization — actually demands it.

And whatever tier you land on: buy it now, in the machine, rather than planning to add it later. That's the one piece of advice this year's memory market has made non-negotiable.

Frequently Asked Questions

Is 8GB of RAM still enough in 2026? Only for genuinely light use — a few tabs, documents, streaming. It's below the comfortable threshold for modern multitasking, and it disqualifies you from most on-device AI features. Don't buy it new.

Is 16GB enough for gaming? Yes, for the games themselves. Most 2026 AAA titles list 16GB as minimum and 32GB as recommended. If you stream, record, or keep Discord and a browser running alongside, 32GB is the better call.

Do I need 32GB for local LLMs? Not strictly — you can run a quantized 7–8B model on 16GB if it fits entirely in your GPU's VRAM and you close other applications. But 32GB is the practical floor for a setup you'd actually enjoy using, since it leaves room for the OS, tooling and modest CPU offload.

Does faster RAM matter, or just capacity? Capacity first, always — running out is far worse than running slow. Once you have enough, speed matters most for integrated graphics, CPU-bound gaming, and offloaded model inference, where memory bandwidth directly limits throughput.

Can I mix different RAM sizes and speeds? Physically yes, but the system typically runs everything at the slowest module's speed and timings, and you may lose dual-channel benefits. Buy matched kits where you can.

Will RAM prices come back down? Not soon. Analysts expect the shortage cycle to persist into 2027 or later, since new fabrication capacity takes years to come online and manufacturers are prioritizing higher-margin AI memory.

Sources: TrendForce DRAM contract pricing surveys (Q1–Q3 2026); CNBC reporting on the AI memory shortage (January 2026); Tom's Hardware memory price index and Q3 2026 pricing analysis; HP and Newegg 2026 memory buying guides; community benchmarks and hardware requirement analyses for local LLM deployment (2026).