RAM Prices Just Doubled in a Quarter, and the AI Boom Reached Your Invoice
DRAM contract prices rose 90-95% as datacentres bought the memory first.
On 2 February 2026 the analyst firm TrendForce revised a forecast it had published only weeks before, and the size of the revision is the story. In early January the firm expected contract prices for DRAM — the working memory that sits in phones, laptops and servers alike — to rise 55 to 60 percent during the first quarter of 2026. By the start of February that estimate had been raised to 90 to 95 percent for the same quarter. NAND flash, the memory inside solid-state drives, moved from an expected 33 to 38 percent up to 55 to 60 percent.
Contract prices are what the large buyers pay, not what you see on a shop shelf, so they arrive in the retail world with a delay. But a near-doubling in a single quarter does not stay hidden for long. TrendForce expects PC DRAM to roughly double from the holiday quarter, and forecasts a rise of around 90 percent for the LPDDR4x and LPDDR5x memory soldered into notebooks and smartphones — increases it called the steepest in the history of those parts.
Two pressures collided. AI hyperscalers and cloud providers have been buying memory at a pace the supply chain was not built for, and on top of that, PC shipments in the fourth quarter of 2025 came in higher than expected, drawing down the stock that was left.
The part of the AI build-out nobody put on a slide
The mechanism is worth understanding, because it explains why a datacentre purchase changes the price of a laptop. A large language model answering a prompt holds its state in a key-value cache — in effect the model's short-term memory for that conversation. As The Register explains it, that cache sits in the accelerator's high-bandwidth memory while a session is active, moves out to slower system memory once the session goes idle, and in many cases ends up on a storage tier after that. Keeping it around saves a great deal of recomputation and makes the assistant feel responsive. It also consumes ordinary system memory and SSDs at a scale the supply chain had not planned for. The demand comes from inference — models serving users — and not from training runs alone. TrendForce noted that since late 2025 the leading North American cloud providers had been increasing procurement sharply, "resulting in a surge of enterprise SSD orders".
The reason it reaches consumer devices is that the parts are not separate markets. Nvidia's most powerful rack systems each carry 54 terabytes of LPDDR5x — the same class of memory that goes into a thin notebook or a phone. Memory that goes into a rack is memory that is not available for a laptop.
There is a timing quirk that explains why prebuilt machines looked calm while standalone memory kits tripled in price. Big system builders — The Register names Dell and HP — buy memory roughly twelve months ahead of the demand they expect, so they have been shipping systems assembled from older, cheaper stock. Once that stock runs down and restocking happens at current prices, system prices follow.
Nor is relief close. Vendors now have the capital to build new fabs, but bringing a plant of that kind into production is a multi-year exercise. Prices are expected to peak later in 2026 and to stay high through 2028, with no return to earlier levels expected for years.
What this means if you are building something
Right-size your instances now, not at renewal. Memory is a materially more expensive part of a server than it was a year ago. Plenty of small applications run on instances chosen years ago by rounding up, and that habit is no longer cheap. Measure actual usage before the next invoice, not after.
Open hosting renewals early. If a contract comes up for renewal in the next two quarters, start the conversation while you still have time to move. A provider that bought ahead has some cushion, as the system builders did, but that cushion is finite and we would not assume mid-2026 quotes will look like 2025 ones.
Treat a memory-hungry architecture as a budget decision. Caching whole datasets in RAM, running a heavy search index, keeping generous idle worker pools — these were once free-ish defaults. They now have a price attached that compounds every month. That does not make them wrong; it makes them a choice someone should approve on purpose.
Do not wait for the dip on hardware. If a client planned to refresh laptops or buy a build machine in 2026, the published forecast points upward through the year and high through 2028. Waiting is not obviously a strategy, though nobody should promise a specific price either.
Our reading: this is not a reason to redesign anything. It is a reason to reopen the spreadsheet. Most of the projects we run would absorb the increase without anyone noticing; the ones that will feel it are the memory-heavy ones where nobody has looked at the instance size since launch. That review costs an afternoon and it is worth having done before the invoice arrives.
Sources
