Why is RAM so expensive right now, and will it go down?

By Andrej Ruckij · November 9, 2026

TL;DR: RAM is expensive because AI data centers are eating the memory supply. The three big makers (Samsung, SK hynix, Micron) moved factory capacity to server memory and HBM, the stacked memory inside AI chips, because it pays better, and the biggest buyers locked up much of the next two years with long-term contracts. Ordinary RAM for laptops and phones gets what’s left. TrendForce counted contract prices up 45-50% in one quarter at the end of 2025 and forecast another 90-95% for the next one. The rises are slowing (TrendForce expects 10-15% this quarter), but the makers themselves say supply stays tight into 2027 and beyond, so I wouldn’t wait for prices to fall back. If you’re buying for a team, get 2027 quotes now and buy what you need, not more.

I saved a couple of posts about this over the weekend, and the more I read the more I realised it isn’t a gamer problem. It hits anyone buying laptops, phones, servers or GPU time next year. So here’s what’s actually going on, in plain words, with the numbers I could check.

It’s for anyone buying RAM or a new computer, and especially for anyone buying them for a team or renting GPUs for AI. There’s a short section for each at the end.

Why is RAM so expensive right now?

Most of the world’s DRAM comes from three companies: Samsung, SK hynix and Micron. Each one has a fixed number of factories, and every wafer that goes through them can become one kind of memory or another.

In 2026 the most profitable kind by far is memory for AI servers, especially HBM. HBM is the stacked, very fast memory that sits next to the processor in Nvidia and AMD accelerators. It also eats a lot of wafer. Micron has said HBM3E needs about three times the wafer supply of DDR5 for the same amount of memory (that’s Micron via press reports; I couldn’t open a Micron page that states it).

So the makers do the obvious thing. TrendForce, the research firm most of the industry quotes on memory prices, put it simply in March: “suppliers continue prioritizing server DRAM due to its superior profitability”, and they “have simultaneously reduced shipments to PC OEMs and module makers”. Samsung’s own Q3 investor deck says it “managed product-mix with focus on AI products amid limited capacity constraints”.

The second piece is contracts. SK hynix said in its October results that it has signed long-term agreements with 14 major customers. Those already cover more than half of what it can make in 2027 and 2028. Micron reported 26 strategic customer agreements in September, up from 16 in June, some running to 2031 (via the Korea Herald). Those buyers get supply and somewhat steadier prices. Everyone without a contract, which is most businesses and every consumer, buys from whatever is left, at whatever the price is that month.

That’s the whole mechanism. More of each wafer goes to AI, and the biggest buyers reserved their share years ahead.

What’s making RAM so expensive compared with a year ago?

The size of the jumps is what surprises people. These are TrendForce’s figures for conventional DRAM contract prices (what PC and phone makers pay, quarter on quarter):

Quarter	Change in contract price	Type
Q4 2025	+45-50%	reported
Q1 2026	+90-95% (PC DRAM over +100%)	forecast
Q2 2026	+58-63%	forecast
Q3 2026	+13-18%	forecast (made in July)
Q4 2026	+10-15%	forecast (made in September)

Most of those are forecasts made at the start of each quarter, and TrendForce’s own news site later relayed about 20% for Q4. If they played out anywhere close, contract memory now costs several times what it did in autumn 2025.

You can see it in what the companies that buy memory say:

Apple called memory pricing “unprecedented” on its November 2 earnings call, said December would bring “even higher memory costs”, and guided its holiday gross margin down to 45-46%.
HP grew its PC revenue 18% in the quarter to July while the number of units fell 16%. HP doesn’t blame memory for that, but it lists “increasing memory and storage costs” among its risks.
Raspberry Pi raised prices by $10 to $60 per board, depending on memory size, in February, saying the cost of some parts “more than doubled over the last quarter”.
Framework, the repairable-laptop maker, said on November 6 it had to raise prices again on 16GB, 32GB and 48GB memory.

If you priced a laptop or a RAM upgrade in 2025 and look again now, that’s why the number is different.

Why are laptops and phones short of memory?

Because makers serve servers first. TrendForce said in September that “the PC DRAM supply could decline in 2027 as more capacity shifts to server applications”. In July it expected smartphone makers to raise retail prices to cover the cost of phone memory.

The other thing to watch is quieter: specs. TrendForce wrote in October that brands could cut memory capacity to control costs, though it warned that would make products less competitive. So it’s worth checking whether a 2027 laptop at your usual price comes with less RAM. That matters more than it used to, because Microsoft requires 16GB of RAM for a Copilot+ PC, and anything with less isn’t one.

Does this make AI more expensive too?

Here it gets interesting, because most of us think of AI as something that keeps getting cheaper. The hardware underneath is going the other way.

GPU rental is the clearest example. Nebius lists an H100 at $4.50 an hour on its price page today. Uvik’s GPU price index says that was $3.85 before October 1, so about 17% more. AWS says its Capacity Block prices “are updated regularly based on trends in supply and demand”, and the next update is scheduled for January 2027. Neither company says memory is the reason, so I wouldn’t pin it on RAM alone. It’s the same story one level up, though: demand for AI hardware is ahead of supply.

If you rent GPUs yourself, you feel it already. If you only use AI through a subscription or an API, someone between you and the hardware is carrying that cost for now, and I don’t know how long that lasts. And I wouldn’t build a 2027 budget on the assumption that AI costs only ever go down.

Will RAM prices go down?

Not soon, if you believe the people who make it, and they have the best view of their own order books.

SK hynix expects demand to exceed supply “through 2030” (its Q3 release). Korean press (Chosun Biz, relayed by TrendForce) reported the company saying its DRAM and HBM are “effectively sold out through 2029” once customer demand under discussion is counted. That line isn’t in the official release, so take it as reported.
Samsung told investors 2027 supply will be tighter than 2026.
Micron’s CEO said conditions would be “much tighter” in fiscal 2027 and 2028 (via the Korea Herald).

Now the case against, because there is one.

The rises are slowing. From +90-95% in a quarter to a forecast +10-15% is a big change. TrendForce said in July that consumer buyers “are reaching their affordability limit”, and in HP’s case fewer PCs went out the door even as revenue grew, though HP hasn’t said why.

The supply response is huge. Paul Kedrosky, an investor who writes about tech cycles, argued in August that “the current shortage is seeding its reversal” and called the factory spending “the largest supply response in memory-market history”. He thinks a 2028-29 downturn is plausible. That’s one person’s opinion, and mostly behind a paywall, but memory has crashed before: TrendForce’s forecasts called for drops of nearly 20% in early 2019 and 13-18% in spring 2023, both after inventories piled up.

China is adding capacity. CXMT is expected to grow from about 320,000 to 420,000 wafers a month by 2027, but TrendForce’s relay says the extra output is mainly for China’s own market, so I wouldn’t count on it for Western prices next year.

What’s different this time: supplier inventories are “at historic lows” (TrendForce, September), not piling up, and so much is locked into multi-year contracts. So my guess is prices flatten in 2027 rather than fall, and a real drop is a 2028 question at the earliest. No memory maker has given a date, and I found no forecast I’d trust more than a year out.

What I’d do this week

If you’re buying RAM or a computer for yourself:

If you need it now, buy it now. Quarterly rises are smaller than at the start of the year, but they’re still rises, and Framework raised its memory prices again on November 6. If you want a machine for running AI locally, Framework’s own advice in September (it sells memory, so bear that in mind) was to “pick up the memory capacity now that matches the models and workloads that you want to run”.
If you can wait a year, wait, but don’t expect a bargain. Back in February Raspberry Pi called the situation “ultimately a temporary one”, and prices have kept rising since. Nobody who makes memory has said when it ends.
Don’t buy less RAM than you’ll need. If you want Copilot features on Windows, Microsoft requires 16GB for a Copilot+ PC.

If you buy hardware for a team:

List what you must replace in 2027. Laptops past warranty, a server that’s running out of RAM, phones on a refresh plan. Your asset list or last year’s invoices are enough to start.
Send that list to your supplier today and ask for a quote, and whether they’ll hold it and for how long. Prices are still rising each quarter, so if they’ll hold a price at all, that’s the simplest protection you have.
Don’t buy far beyond one refresh cycle. Framework says it will “bring prices back down as suppliers bring costs down”. Stockpiling three years of laptops locks you into today’s prices if Kedrosky turns out to be right.
Check the RAM on every spec sheet. If a quote comes back cheaper than expected, check the memory first.

If you rent GPUs:

Compare list prices across providers before you renew. Uvik’s index shows the same H100 from $3.20 to $12.29 an hour across 16 providers, with GPU clouds far cheaper than the big three at list price.
Price a commitment. Uvik found three-year commitments at AWS, Google Cloud and Azure cut H100 prices by 56-57% compared with on-demand rates. That only makes sense if you’re sure you’ll use the capacity for three years.
If you use AWS Capacity Blocks, decide whether to book before January. AWS charges “the prevailing rate at the time of purchase”, even if the block starts after its next scheduled price update in January 2027. It hasn’t said which way that update goes, but given everything above I’d lock today’s rate for blocks I’m sure to use.

If your team calls AI models through an API, ask whoever built the integration two things this week: which tasks could run on the provider’s cheaper, smaller model, and whether prompt caching is switched on for prompts you send again and again (glossary/prompt-caching explains how it works). Both cut what you pay per task, which matters more if prices ever rise.

What I couldn’t find

No memory maker has published a date for prices to fall. I couldn’t find a Micron page that states the HBM-to-DDR5 wafer ratio (only press reports), retail price history I could verify (PCPartPicker-style data), or any number on what small businesses pay compared with the big contract buyers. If you’ve got real quotes from this autumn, I’d like to see how far apart they are.

Related: glossary/llm for what these models actually run on.

Sources

https://t.me/aidaily_business