The DRAM Shortage Outlasts 2026: Who Actually Pays the AI Memory Premium

Conventional DRAM contract prices rose between 93% and 98% in a single quarter in early 2026, according to TrendForce. Micron turned that into an 87.0% gross margin. The DRAM shortage of 2026 is usually told as a supplier windfall, but every dollar of that margin is a cost somebody downstream had to carry. This post follows the memory premium from the fab to the laptop shelf.

Short answer. In 2026 the DRAM premium was paid in three layers. Hyperscalers absorbed it as higher capital spending. Nvidia absorbed a few points of gross margin and has said it will reprice. PC and phone buyers paid through higher prices, smaller storage and fewer units shipped. Micron and its peers kept the difference.

The DRAM shortage in 2026, measured in contract prices

Contract prices are what PC makers, phone makers and server builders pay memory suppliers under quarterly or longer agreements. They move more slowly than spot prices, which makes them the better measure of what your suppliers actually paid. TrendForce publishes the most widely used quarterly estimates.

The sequence for 2026 runs like this. In February, TrendForce raised its 1Q26 call for conventional DRAM to a rise of 90% to 95%, up from 55% to 60% a month earlier. Its June release then reported the outcome: contract prices rose approximately 93% to 98% in the quarter, and DRAM industry revenue rose 81% to about $97 billion.

The same June release expected another 58% to 63% for 2Q26. In July, TrendForce put 3Q26 at 13% to 18%. On 30 September it put 4Q26 at 10% to 15%, with NAND Flash up 15% to 20%.

The slowdown is still a rise

Headlines about "moderating" increases miss the base effect. A 10% rise on a price that has already tripled is a larger dollar increase than a 30% rise on the 2025 price. Chaining TrendForce's calls at their midpoints, a contract that cost 100 at the end of 2025 costs about 408 by the end of 2026.

That index is my calculation, not TrendForce's, and only the first quarter in it is a reported result. The other three are TrendForce's calls at the time. I would still plan on it, because every revision this year has gone up rather than down.

What Micron's earnings say about AI demand and pricing power

Micron reports on a fiscal year ending in late August or early September, so its quarters sit about a month behind the calendar. Read its four 2026 results releases in sequence and the pricing story is visible without any commentary.

Gross margin from 45.7% to 87.0% in four quarters

Non-GAAP gross margin was 56.8% in the quarter to 27 November 2025, 74.9% to 26 February, 84.9% to 28 May and 87.0% to 3 September 2026. That sequence comes from Micron's fiscal fourth quarter release. The year-ago quarter was 45.7%. Revenue went from $11.3 billion to $54.2 billion over the same year.

A margin that high in a manufacturing business means price, not volume, did most of the work. Micron's own guidance for the quarter to late November 2026 is $61.5 billion of revenue at about 86.25% non-GAAP gross margin. That is the first quarter in the sequence where the margin is guided flat to slightly down, which matters for the counter-case later.

The margin line also explains the stock story, which this site covered in why memory stocks outran Nvidia in 2026. This post asks the other question: whose costs moved to fund it.

Take-or-pay contracts move the risk downstream

On its September earnings call, Micron said it had signed 26 Strategic Customer Agreements. Chief executive Sanjay Mehrotra described them as multi-year take-or-pay contracts and estimated they cover over 35% of Micron's revenue through 2030.

Take-or-pay means the customer pays for committed volume whether it needs the chips or not. In a shortage, buyers sign these to secure supply. The effect is that a share of the downturn risk, whenever it comes, now sits with the buyer rather than the supplier. That is the most underappreciated transfer in the whole memory story.

Mehrotra also told analysts Micron expects memory and storage supply-demand conditions to be much tighter in 2027 and 2028 than they were in 2026, per the Investing.com transcript of the call. That is the basis for the claim that the shortage outlasts 2026. It is a supplier's forecast about its own market, and you should weight it accordingly.

Quarter by quarter: memory prices against downstream prices

The table below lines up the three things you need on one page. Those are what memory cost, what Micron earned on it, and what the next company in the chain did about it. Calendar and fiscal quarters do not align exactly, so read each row as a period, not a matched pair.

Only the 1Q26 price change is a reported result. 2Q26 to 4Q26 are TrendForce's calls when published. GPU partner, phone maker and Lenovo moves come from press reporting, not company price lists. Nvidia's published consumer MSRPs were not changed in any source found for this post.

Two patterns stand out. First, downstream price moves lag memory by about a quarter, because contracts reset quarterly and finished goods carry old inventory. Second, the companies closest to the chip moved first and the companies closest to the consumer moved last, which is the order of bargaining power.

GPU makers: Nvidia absorbs first, then reprices

Every AI accelerator carries high-bandwidth memory, or HBM, stacked next to the processor. HBM is DRAM built in a more complex way, and it competes with conventional DRAM for the same wafer capacity. That shared capacity is why an AI server boom raises the price of laptop memory.

Data centre GPUs: a margin dip, then a price rise

Nvidia's gross margin was 75.0% in the quarter to 26 July 2026, per its results release. On the call, chief financial officer Colette Kress said memory pricing had exceeded the company's prior expectations and was heading higher into next year, according to Barchart's account of the call. Nvidia guided a margin trough of 71% to 72% in its fiscal fourth quarter, recovering to 72% to 73% in fiscal 2028 as price increases already agreed with customers take effect.

That is a textbook absorb-then-pass-through pattern. The Motley Fool reported that Nvidia's purchase commitments rose from $119 billion to $279 billion, primarily for memory. My read is that Nvidia is the one buyer large enough to secure supply in bulk, so it carries the timing risk for a few quarters and then sends the bill to hyperscalers. Those figures come from press accounts of the call and filing, not documents I opened, which is noted in the references.

Consumer graphics cards: the channel moves before the list price

Gaming GPUs use GDDR memory, which is a smaller market and gets less supplier priority. Here the increases arrived through board partners and regional channels rather than official list prices. TrendForce reported in August that RTX 50 series prices in Korea could rise as much as 30%, citing GDDR7 and TSMC wafer costs.

The quiet route matters. When the list price holds and the shelf price rises, the buyer cannot easily see who took the margin. I would treat any consumer GPU price you see quoted against MSRP in 2026 as a floor, not an estimate.

Hyperscalers pay through capex, and the HBM shortage extends the bill

The largest single absorber of AI hardware costs is the cloud builder. Unlike a PC maker, a hyperscaler cannot raise the shelf price of a server it already owns. It pays the memory premium into its capital budget and recovers it over years of depreciation.

Meta was explicit. Its first quarter 2026 release raised full-year capital expenditure to $125 billion to $145 billion, from $115 billion to $135 billion. The stated reasons were higher component pricing and, to a lesser extent, more data centre costs, per Meta's results release. Microsoft's April commentary, as reported by The Motley Fool, put about $25 billion of its calendar 2026 capex down to higher component pricing.

That figure was later restated alongside a lease accounting change, so treat it as an order of magnitude. The broader capex picture is in the analysis of 2026 hyperscaler capex and its missing payback math.

HBM in 2027: the next bill is already priced

On 29 September, TrendForce raised its 2027 HBM outlook and now expects the blended average selling price to rise 121% year on year. It expects the HBM market to remain supply-constrained throughout 2027. Micron said on its call that it has agreed the vast majority of its calendar 2027 HBM supply at significantly higher prices.

The most telling line in the TrendForce release is that GPU and ASIC makers are considering less HBM per device, driven by limited supply and quickly rising system costs rather than weak demand. When buyers start designing memory out of their products, the premium has reached the point where it changes engineering decisions.

A hyperscaler that pays more per accelerator either earns less per unit of compute or charges more for it. Either way, it feeds into what you pay for inference, a dynamic covered in how inference costs set AI product margins and in how GPU depreciation schedules shape reported earnings.

Device buyers pay last, and they pay in specs as well as price

The end of the chain is where the bargaining power runs out. A PC maker buying memory by the quarter has neither Nvidia's scale nor a hyperscaler's balance sheet. It has three levers: raise prices, cut specifications, or accept lower margins. In 2026 it pulled all three.

PCs: higher revenue, fewer units, thinner margins

HP's fiscal third quarter shows all three levers at once. Personal Systems revenue rose 18% to $11.8 billion while total units fell 16%, and segment operating margin slipped to 4.6% from 5.4%, per HP's results release. Prices went up, volumes went down, and the margin still shrank.

Lenovo told investors in May that it raised PC prices to offset memory costs and shifted its mix toward premium machines. TrendForce's 4Q26 release adds the specification lever: PC brands are reducing SSD capacity in mainstream models to lower bill-of-materials costs. Your next fleet refresh may cost more and ship with less storage than the last one.

TrendForce's October notebook outlook expects global shipments to fall by a low single-digit percentage in 2027. If brands pass through costs fully, it says the decline could reach the high single digits. Its blunt conclusion is that brands are unlikely to maintain pricing, specifications and profitability simultaneously next year.

Phones: inventory buys time, not immunity

Smartphone makers are leaning on stock bought before the surge. TrendForce says brands are largely relying on existing eMMC and UFS inventories with very limited new procurement, and that rising costs are weighing on production. It put 2Q26 smartphone production at 275 million units, down 8% from a year earlier.

Caixin reported in March that Oppo, Vivo and Honor had raised prices on memory costs. When the cheap inventory runs out, the choice facing phone brands is the same one facing PC brands. I expect more of the adjustment to show up as lower storage tiers at the same price than as headline price rises, because that is easier to sell.

Where this argument is weakest

The case above is clean, and clean cases deserve suspicion. Three things could make it wrong.

Most of the price path is forecast. Only the 1Q26 contract price change in this post is a reported result. The 2Q26 to 4Q26 numbers are TrendForce's calls at the time, and actual outcomes could land on either side. The 408 index is a chain of midpoints, which compounds any error in each step.

Memory is the most cyclical market in semiconductors. Every prior shortage ended in a glut, usually faster than suppliers predicted. Micron is raising capital spending, its peers are adding clean rooms. Its own guidance for the quarter to November shows the gross margin flat to slightly down for the first time in a year. A plateau in supplier margin is often the first sign that buyers have regained some leverage. The take-or-pay contracts cut both ways: they prove buyers were desperate in 2026 and they lock in volume buyers may not want in 2028.

Several downstream figures rest on secondary reporting. Nvidia's margin guidance and purchase commitments, Microsoft's $25 billion component figure, the GPU partner price rises, Lenovo's pricing and the Chinese phone price rises all come from press accounts rather than documents I opened. They point the same way as the primary sources, which is why I used them, but none of them should carry a decision on its own. A consumer of this analysis who needs exact figures should go to the filings.

There is also a fair argument that the premium is mostly paid by shareholders of hyperscalers rather than by anyone else, because cloud prices have not risen in step. I cannot settle that with public data, because no hyperscaler discloses AI compute pricing by quarter.

What the memory premium means for your hardware and cloud budget

If you run a company that buys laptops, servers or cloud compute, the useful question is narrow. Where in your budget does this land, and when. The answer depends on how close each line item sits to the memory fab.

Laptops and desktops are the fastest and most direct exposure. Quotes are already shorter-dated, Dell's chief operating officer described repricing open quotes in January, and TrendForce expects cost pressure to persist into 2027. Buy what you need now, and check storage specifications on every quote, because a like-for-like model number may no longer mean like-for-like hardware.

On-premise servers are the second exposure, and the most negotiable. Server DRAM is undersupplied into 4Q26 per TrendForce, but some suppliers' increases lag the market because of long-term agreements. If you have a vendor with an existing agreement, use it before it resets.

Cloud and API costs are the slowest to move and the hardest to see. Hyperscalers have absorbed the premium so far, but the 2027 HBM price rise has to land somewhere. The pattern in why falling token prices still produce rising AI bills applies here too: unit prices can hold while your total bill rises through usage limits and plan changes.

My position is that a finance team should treat 2027 hardware as a cost-up year, not a cost-down year. Budget for at least the 2026 level plus the inflation already visible in TrendForce's 4Q26 call. If the cycle turns early, the surplus is a pleasant surprise. If you budget for relief and it does not arrive, you defer a refresh you need.

Frequently asked questions

Where to start this week

Pull your last three hardware quotes and your next planned refresh. For each, write down the memory and storage specification alongside the price, then ask your reseller for a requote today with the same specification stated explicitly. The gap between the two numbers is your personal measure of the DRAM premium, and it is more useful than any index in this post.

Then ask your cloud account manager one question in writing: what pricing commitments apply to your current instance types through 2027. A vague answer tells you the HBM bill has not landed yet. It does not tell you it will not.

Related reading