AI Ate Your RAM: Memory Prices Are Up 500% and Local AI Just Got the Bill

A year ago, RAM was the part of a PC build nobody argued about. You wanted 32GB of DDR5-6000, you paid about $120, and you moved on with your life. It was the cheapest meaningful upgrade in computing.

This month, that same kit costs $392.

Tom's Hardware has been tracking this all year, and their latest index is the kind of chart you squint at because you assume there's a data error. There isn't. Memory prices are up 500% in twelve months. Some parts have hit ten times their all-time lows. The trigger, per their reporting: AI datacenter demand, with tariffs sprinkled on top like asbestos frosting.

If you're a normal person building a gaming PC, this is annoying. If you're one of the growing number of people running AI models locally — on a home rig, a Mac, a workstation — this is a direct hit to the one component that actually matters. Because here's the thing nobody says out loud: local AI is a memory business. Not FLOPS. Memory. And the memory market just went supernova.

The numbers, because you should see them

Let's start with the retail side. A mainstream 32GB DDR5-6000 kit sat around $110–140 in late 2025. As of August 2026, it's $392. Not a typo, not a scalper — that's the tracked street price.

Retail follows contract prices, and contract prices have gone vertical. TrendForce's data for Q1 2026 shows PC DRAM — the DDR4 and DDR5 that goes into desktops and laptops — up 105–110% in a single quarter. That's the steepest quarterly jump ever recorded for the category. Conventional DRAM across all segments climbed 90–95%. NAND flash, the stuff inside your SSD, rose 55–60% in the same window, and enterprise SSDs climbed another 53–58% on top of previous hikes.

Read that again: the price of memory doubled in one quarter, and that was just Q1. It's kept climbing since.

Hacker News put the Tom's Hardware piece on the front page this week, where it collected 500+ points and a comment section that reads like a wake for the DIY PC. One recurring theme: people who bought 64GB of DDR5 in 2024 for cheap are now sitting on hardware that appreciates like a used Toyota. Your RAM has better resale value than your car.

Why this is happening (it's the AI you were promised)

The mechanism is almost insultingly simple.

Samsung, SK Hynix, and Micron make DRAM on fabrication lines. Some of those lines make the DDR5 sticks you and I buy. Others make high-bandwidth memory — HBM — the stacked memory modules that go into Nvidia's AI accelerators. When demand for HBM goes parabolic and Nvidia is willing to pay effectively any price, the memory makers do the rational thing: they retool consumer lines for HBM, because the margins aren't in the same universe.

Every fab line converted to HBM is a fab line no longer making the memory in your laptop. Meanwhile AI servers themselves are also stuffed with conventional DRAM — a single serious AI server carries terabytes of it — so datacenters are eating the consumer supply from both ends.

Tim Cook, in a Wall Street Journal interview in June, said the quiet part loudly: "There's less supply at a time when consumers want devices and the memory guys are passing along huge price increases." He called the situation "unsustainable" and confirmed price hikes for Apple products are "unavoidable."

Fortune called it RAM-ageddon. The name stuck because it's accurate.

The Apple tax arrives

Apple's response played out in three acts this summer.

Act one, May: Apple quietly raised the Mac mini's starting price by deleting the base storage configuration. Same sticker price on the surviving SKU, higher floor. Classic.

Act two, June 17: Cook does the WSJ interview, uses the word "unavoidable," and specifically calls out HBM allocations strangling consumer DRAM supply.

Act three, June 25: MacBook and iPad prices go up across the line. Not one flagship product — the whole shelf.

There's a detail buried in all this that's genuinely funny in a grim way. Apple's most affordable Mac notebook right now — the MacBook Neo, otherwise a lovely little machine — ships with 8GB of RAM. Eight. In 2026. When even midrange phones carry more. Apple didn't design that product in a vacuum; they designed it for a world where memory is the scarce input, and 8GB is what the margins allow. The memory crisis isn't just raising prices. It's reshaping what products get made.

If you've been waiting to buy a 64GB Mac for local models, understand what's happened: you're now shopping in a market where the machine you want is a memory-strategy decision, repriced twice in three months.

The part that stings if you run local models

Here's where I get opinionated, because this is the angle everyone's missing.

The tech press is covering this as a PC gaming story. Frames per dollar, build guides, "should you upgrade now or wait." Fine. But there's a whole population — call it the local AI crowd — for whom this is a structural problem, and almost nobody is talking to them.

If you run open-weight models, you know the actual bottleneck isn't compute. A quantized 30B model runs happily on a gaming GPU. The constraint is memory: system RAM to hold the model, VRAM or unified memory to run it fast, and fast NVMe to load it — because model weights are now 30, 60, 100+ gigabyte files. You don't need a supercomputer. You need a lot of the exact three things that just got most expensive: DRAM, VRAM-adjacent supply, and NAND.

This year has been a golden one for open weights. Qwen, DeepSeek, Meta's return with Muse Glimmer — genuinely capable models, MIT and Apache licensed, with token prices on API providers so low they're rounding errors. The software side of local AI has never been better.

And in the same year, the hardware side of local AI got 3–5x more expensive. The door and the cover charge both moved. The models are free; the room they live in now costs like a nice GPU used to.

There's an irony here worth naming: the same AI boom that made open models abundant is consuming the memory supply that made running them at home affordable. Datacenters eat the fabs, the fabs stop making your RAM, and the exit ramp from cloud AI gets tolled. You can rent tokens from the cloud for pennies, but the gear to opt out? Luxury goods now.

What to actually do

Practical advice, in order of usefulness:

If you were going to buy memory, the window is behind you — but it's still behind-er next year. Analysts across TrendForce and Counterpoint see this crunch persisting into 2027. New fab capacity takes years. Waiting for a 2026 Black Friday deal is a mistake; the deals are gone because retailers have no margin cushion left. Contract prices hit consumers within weeks now.

Buy used, and buy DDR4 without shame. The legacy DDR4 market is rallying too, but a used 64GB DDR4 kit from an upgraded workstation is still dramatically cheaper per gigabyte than new DDR5, and for inference — where memory bandwidth matters less than capacity — it's a legitimately rational buy. Somebody's decommissioned server is your discount.

Lean on quantization. A Q4-quantized 30B model fits comfortably in 24GB and retains most of its quality. The era of "download the full-precision weights because storage is cheap" is over; storage isn't cheap anymore either, since SSDs are doubling on the NAND crunch. Smaller, quantized, more of them — that's the local stack now.

Go hybrid. This is the honest answer most people land on: run a small capable model locally for the everyday stuff — quick edits, automation, file wrangling — and rent the big frontier weights by the token when you genuinely need them. With open-model APIs priced like they are, a hybrid setup costs less per month than the price increase on a single RAM kit. That math is not subtle.

Don't panic-buy a GPU "for the VRAM" this month. Consumer GPU pricing has its own turbulence ahead as HBM and GDDR allocation tightens, but buying hardware in a panic during a supply crisis is how you pay the crisis premium. Plan the build, buy pieces opportunistically, used market included.

When does it end?

Memory is cyclical. Everyone in the industry knows the script: shortage, price spike, capacity buildout, glut, collapse, hangover. It happened in 2017–19, it happened in 2021–23, it will happen again. Samsung, SK Hynix, and Micron are all building, and when the new capacity lands — most estimates say 2027 — prices should normalize with the usual overshoot, and RAM will be cheap again and we'll all pretend this never happened.

Two things make this cycle different. First, the demand side isn't consumers, who eventually get priced out — it's datacenter capex, which so far has treated price increases as a rounding error. When your customer doesn't flinch at 5x, the shortage lasts longer than the textbooks predict. Second, tariffs are layered on top, which doesn't cause the shortage but does tax the cure.

So: 2027 relief, maybe. Emphasis on maybe.

In the meantime, the local AI crowd gets to be the ones who pay for the party. Your DDR5 is the AI boom's catering bill, and you weren't even invited to the event.

If you want to use open-source models — GLM, DeepSeek, Qwen, the whole shelf — without mortgaging a memory upgrade, that's literally what we built CopperRiver for: a desktop assistant that does the browsing, terminal work, and file wrangling on top of open models, from $9 a month. A full year of it still costs less than a third of the price hike on one RAM kit. We did the math so your wallet doesn't have to.