Analysis · Memory

Why is RAM so expensive right now? The AI memory shortage, and who is collecting the money

7 August 2026 · 12 min read

The short version: memory makers moved their production lines onto high-bandwidth memory for AI accelerators, because that is where the money is. The DDR5 in a desktop and the SSD in a laptop are made on the same lines and now compete for what is left. Contract prices rose 58-63% in one quarter, then again in the next. Your RAM did not get more expensive because retailers are gouging. It got more expensive because a data centre outbid you.

The second half of this piece is the part almost nobody writes: the four listed companies collecting that money have all fallen sharply anyway - two of them by more than 45% from their highs. If you came here thinking "memory is scarce, so memory stocks must be a buy", the last three weeks have already tested that idea and the answer was uncomfortable.

What actually happened

An AI accelerator needs high-bandwidth memory - HBM - stacked next to the GPU die. HBM is made from DRAM, on DRAM lines, using DRAM wafer capacity. Every wafer that becomes HBM is a wafer that does not become the DDR5 stick in a gaming PC.

The same squeeze happened one shelf down in storage. AI inference workloads consume enormous amounts of fast NAND flash, and NAND capacity was redirected towards enterprise SSDs and away from the consumer parts that go into laptops, phones and games consoles.

Two things then made it worse than an ordinary shortage:

The supply cannot respond quickly. A memory fab is a multi-year, multi-billion-dollar build. Nothing announced in 2026 produces meaningful volume in 2026.
The buyers on the other side are not price-sensitive. A hyperscaler weighing memory cost against the revenue on an AI cluster will pay almost anything. A person building a PC will not. When those two bid for the same wafer, it is not a close contest.

The prices, in order

These are TrendForce's contract price forecasts - the prices manufacturers pay, which flow through to retail with a lag of a month or two.

Quarter	Conventional DRAM	NAND flash
Q2 2026	+58% to +63%	+70% to +75%
Q3 2026	+13% to +18%	+10% to +15%

Compounding those two quarters at the midpoints gives roughly +85% for DRAM and +95% for NAND in six months. That arithmetic is ours, applied to TrendForce's published ranges, and it is the honest way to read a quarter-on-quarter series - the increases stack, they do not replace each other.

Note what the second row is really saying. Q3 looks like relief only against a Q2 that was extraordinary. Prices are still rising, just less violently, and TrendForce is explicit about why: consumer buyers of PCs and smartphones have hit their affordability limit. Demand is not being satisfied. It is being priced out.

Why you pay more than a hyperscaler does

This is the part that explains the anger in every PC-building forum, and it is rarely spelled out.

Several large US cloud providers signed multi-year long-term agreements with the memory makers. Those contracts function as price caps: TrendForce notes they restrict suppliers from raising prices for those clients. Which means the increases have to land somewhere else.

The customers with the most buying power locked in their price. Everyone without a contract absorbs the rise on their behalf.

So the retail buyer, the small system integrator and the OEM without leverage are not merely paying a shortage premium. They are paying the portion of the increase that the largest customers contractually cannot be charged. The person hit hardest by the AI build-out is the one with the least to do with it.

What it costs you, specifically

Gartner put numbers on the consumer side in February 2026, and they have aged well:

Combined DRAM and SSD prices up 130% by the end of 2026
PC prices up 17% and smartphone prices up 13% against 2025 levels
PC shipments down 10.4%, smartphone shipments down 8.4% across 2026
Memory rising to 23% of a PC's total bill of materials, from 16% in 2025
The sub-$500 PC segment disappearing entirely by 2028
People holding devices longer - lifetimes extending 15% for business buyers and 20% for consumers

Gartner's Ranjit Atwal described it as "the lowest level of device shipments witnessed in over a decade". A component shortage has become a consumer-electronics contraction.

Who is collecting the money

Four companies on our board sell into this directly. They are not interchangeable, and the differences matter more than usual right now.

Micron (MU) - the broadest exposure

Micron sells HBM, conventional DRAM and NAND, so it captures the shortage from every angle. In fiscal Q3 2026, reported 24 June 2026, revenue was $41.46 billion, up 346% year on year and up 74% on the prior quarter, at a GAAP gross margin of 84.6%. It guided the following quarter to $50.0 billion at roughly 86% gross margin.

One number worth sitting with: its Core Data Center and Cloud Memory segments together produced $25.29 billion of that $41.46 billion. By our arithmetic, 61% of Micron's revenue is now data centre. This is not a PC-components company that happens to serve AI. It is an AI company that also sells PC components.

SanDisk (SNDK) - the purest NAND read

SanDisk reported fiscal Q4 on 5 August 2026: revenue $8.965 billion, up 372% year on year and 51% sequentially, at an 84.6% gross margin, with GAAP diluted earnings of $43.97 per share. Full-year revenue was $20.25 billion, up 175%, and the data centre line grew 437% to $5.15 billion. Guidance for the current quarter is $10.30-10.80 billion.

Those are not normal semiconductor numbers. They are what happens when a commodity with fixed short-run supply meets a buyer who cannot substitute.

Western Digital (WDC) and Seagate (STX) - the other kind of storage

Both now sell hard drives into AI data centres, where cost per terabyte still beats flash at scale. Western Digital reported fiscal Q4 on 5 August 2026 with revenue of $3.75 billion, up 44%, a 54.1% GAAP gross margin, and long-term debt reduced to zero from $2.485 billion. Seagate reported on 28 July 2026 with revenue of $3.6 billion, up 48.5%, a 52.7% non-GAAP gross margin, and guided to non-GAAP earnings of $7.30 per share.

Hold the gross margins side by side, because the gap is the whole story: roughly 85% at Micron and SanDisk, roughly 53% at Western Digital and Seagate. DRAM and NAND pricing moves violently with spot supply, so the upside lands almost entirely in margin. Hard drive pricing runs on longer contracts and moves in steps. The flash names give you more of the shortage; they will also give you more of whatever comes after it. The WDC vs SNDK comparison and MU vs STX are the clearest way to see that split on live numbers.

Now the uncomfortable part: the stocks already fell

Every figure above is a record. Here is where those four shares actually sit, as at the close on 7 August 2026:

Name	Price	5-day	From 52-week high
Micron	$877.57	+6.6%	-30.1%
SanDisk	$1,212.21	-0.2%	-48.5%
Western Digital	$434.30	-20.3%	-45.7%
Seagate	$812.76	-5.1%	-29.0%

Western Digital is the cleanest illustration. It beat: revenue up 44%, earnings per share more than doubled, debt cleared. Our board still has it down 20.3% over five sessions.

The reason appears to be the shape of the guidance rather than its level. Non-GAAP gross margin came in at 54.4% and the company guided to 55-56% - a step of roughly 110 basis points at the midpoint, by our arithmetic. After several quarters of far larger jumps, that is deceleration in the rate of improvement, on a stock that had already tripled. Investors were no longer paying for growth; they were paying for acceleration, and acceleration is a much harder thing to keep delivering.

Meanwhile capital went somewhere else entirely. Over the same five sessions our board has NVIDIA up 11.6% and within about 5% of its 52-week high, and CoreWeave up 26.3%. Money did not leave the AI trade this week. It rotated out of the companies that supply the build-out and back into the ones doing it. You can watch that on the daily movers page, or split by supply-chain segment on the segments board.

Retail noticed. Our own data shows it

We count posts on X carrying each ticker's cashtag every day, directly against the X API. On Friday 7 August, Micron ranked third of our 34 names with 1,764 posts and SanDisk fifth with 1,155 - both ahead of TSMC, a company more than twenty times Micron's size, which drew 199. Western Digital, barely discussed by retail a year ago, sat fifteenth.

Memory is no longer a niche corner of the semiconductor market that specialists follow. It is, this month, one of the most crowded trades retail has. Worth remembering what that measures and what it does not: post counts describe attention, not approval, and they carry no relationship at all with the direction of a day's move.

#	Name	Posts	Market sentiment
1	NVIDIA NVDA	14,210	37
2	IREN Limited IREN	3,938	10
3	Micron Technology MU	3,009	45
4	Marvell Technology MRVL	2,467	9
5	Nebius Group NBIS	1,524	27

Posts on X carrying each ticker's cashtag, counted by this site over 27 August 2026. Retweets excluded. See all 34 names →

When does it end?

The honest answer is that nobody knows, and anyone giving you a confident date is guessing. What can be said is what the mechanism requires.

It ends when supply grows faster than demand, and supply is measured in fab construction time. TrendForce puts RDIMM bit supply growth at just 15-20% year on year, below expected server CPU shipment growth, and describes server DRAM as undersupplied with shortages anticipated through 2027. Samsung and SK hynix have both been reported as warning that tightness extends into 2027 and beyond, with new capacity not arriving in volume until later - that reporting is worth reading, but it is reporting, not a filing, and should be weighed accordingly.

Two things would genuinely signal a turn, and neither has happened:

Contract prices falling, not merely rising more slowly. Q3's deceleration is not that. No published forecast has DRAM or NAND contract prices declining during 2026.
Demand destruction reaching the data centre. Consumers have already been priced out - that is what Gartner's shipment declines are. The shortage only really breaks if hyperscalers start refusing prices too, and there is no sign of that yet.

If you are buying RAM for a personal build, the uncomfortable summary is that waiting has not paid this year, and no forecast currently points to relief inside 2026. That is not advice to rush; it is the absence of a good option.

What I actually think

The consumer shortage and the memory trade are running on two different clocks, and conflating them is the mistake I would most want a reader to avoid.

The shortage is early. Prices are still climbing, supply cannot respond for years, and the segment being squeezed hardest - ordinary consumer DDR5 and NAND - has no advocate in the room. If you are asking "will my next laptop cost more", the answer is almost certainly yes, into 2027.

The trade is not early. It is somewhere in the middle, and it has stopped rewarding the thing that is easy to see. Every one of these four companies is printing records, and all four are down between 29% and 49% from their highs. That is the market telling you it stopped paying for the level of earnings some time ago and started paying for the second derivative - whether margin expansion is speeding up or slowing down. Western Digital beat, guided margins higher, and fell a fifth in a week on the shape of that improvement.

So "RAM is expensive, therefore buy memory stocks" is not a thesis. It is the thesis that has just been tested in public, and it lost. The genuine question is narrower and much harder: does pricing power outlast the customer's ability to pay? TrendForce flagging consumer affordability limits in Q3 is the first crack in that, and it is worth more attention than any of the record revenue numbers above.

What I would watch, in order: whether Q4 contract price forecasts hold above flat; whether hyperscaler long-term agreements get renewed at higher prices when they roll; and whether the gross margin step at Micron's late-September report is larger or smaller than the one Western Digital just guided to. That last one is the same test that broke WDC, applied to the biggest name in the group.

One thing I am not going to pretend: this is a cyclical industry with a long history of turning brutally, and every previous memory upcycle ended with the same companies losing money. This one is structurally better supported than most, because AI demand is contracted years forward rather than speculative. That makes it different in degree. It does not make it different in kind.

How the board reads it today

Sentiment on ChipSentiment is a behaviour score, not a valuation one. It reflects momentum, relative strength against the group, trend, volume, volatility and drawdown - nothing about earnings, margins or price-to-earnings. A company can post the best quarter in its history and score poorly, which is exactly what several of these names are doing right now. The full methodology is here.

All memory and storage stocks - live prices and sentiment →
Compare Micron and Western Digital side by side →
See which AI stocks X is posting about right now →
Today's board across all 34 names →

If the demand side of this is what interests you, the companies actually buying the memory are covered in our explainer on neoclouds, and the interconnect side of the same build-out in the Marvell piece.

Information and analysis only, not investment advice, and it takes no account of your circumstances. Company figures are drawn from the results announcements linked below; price and mention data are from our own feed as at the close on 7 August 2026 and will have moved since. Percentage changes described as our arithmetic are calculated from the published figures cited. The author holds no position in any security mentioned.

Frequently asked questions

Why is RAM so expensive right now?

Memory makers shifted DRAM wafer capacity towards high-bandwidth memory for AI accelerators, and NAND capacity towards enterprise SSDs, leaving less production for the consumer DDR5 and flash that goes into PCs, phones and consoles. TrendForce put conventional DRAM contract prices up 58-63% in Q2 2026 and a further 13-18% in Q3, with NAND up 70-75% then 10-15%. Compounding those midpoints gives roughly 85% for DRAM and 95% for NAND across six months. Supply cannot respond quickly because new memory fabs take years to build.

When will RAM prices go down?

No published forecast has DRAM or NAND contract prices falling during 2026. Q3 2026 increases are smaller than Q2's, but that is slower growth, not a decline. TrendForce describes server DRAM as undersupplied with shortages anticipated through 2027, and puts RDIMM bit supply growth at only 15-20% year on year. Samsung and SK hynix have been reported as warning that tightness continues into 2027 and beyond. Meaningful relief requires new fab capacity, which takes years to come online.

Should I buy RAM now or wait?

That depends on how long you can wait, and there is no good option at present. Prices have risen every quarter through 2026 and no current forecast shows them declining before 2027. Waiting has not paid so far this year. Equally, prices are decelerating and consumer demand is visibly being destroyed, which is the condition that eventually ends shortages. This is information, not advice, and it takes no account of your circumstances.

Why is my laptop or phone more expensive in 2026?

Memory is now a much larger share of what a device costs to build. Gartner forecast in February 2026 that memory would rise to 23% of a PC's total bill of materials, up from 16% in 2025, pushing PC prices up 17% and smartphone prices up 13% against 2025 levels. It also forecast PC shipments falling 10.4% and smartphone shipments 8.4% across 2026, and expects the sub-$500 PC segment to disappear entirely by 2028.

Which stocks benefit from the memory shortage?

Four names on the ChipSentiment board sell into it directly. Micron (MU) has the broadest exposure across HBM, conventional DRAM and NAND. SanDisk (SNDK) is the purest read on NAND flash. Western Digital (WDC) and Seagate (STX) sell hard drives into AI data centres, where cost per terabyte still beats flash at scale. Gross margins split the group sharply: roughly 85% at Micron and SanDisk against roughly 53% at Western Digital and Seagate.

Why did Western Digital stock fall after good earnings?

Western Digital reported fiscal Q4 on 5 August 2026 with revenue up 44% to $3.75 billion, GAAP earnings per share of $8.21 against $0.67 a year earlier, and long-term debt reduced to zero. The ChipSentiment board still had it down 20.3% over the following five sessions. The likely reason is the shape of guidance rather than its level: non-GAAP gross margin of 54.4% guided to 55-56%, a step of about 110 basis points at the midpoint, which is slower improvement than prior quarters delivered on a stock that had already tripled.

Are memory stocks a buy because RAM is expensive?

That inference has just been tested in public and it did not hold. All four memory and storage names on our board posted record results and all four sit between 29% and 49% below their 52-week highs as at 7 August 2026. The market appears to have stopped paying for the level of earnings and started paying for whether margin expansion is accelerating or slowing. The harder question is whether pricing power outlasts customers' ability to pay, and TrendForce has already flagged consumer buyers hitting an affordability limit.

How much of Micron's business is now AI data centre?

In fiscal Q3 2026, reported 24 June 2026, Micron's Core Data Center business unit produced $11.52 billion and its Cloud Memory business unit $13.77 billion, a combined $25.29 billion out of total revenue of $41.46 billion. That is 61% of revenue by our arithmetic. Total revenue was up 346% year on year at a GAAP gross margin of 84.6%, and the company guided the following quarter to $50.0 billion at roughly 86% gross margin.