Memory prices are climbing in contract and retail markets, and the tech press has a name for it: “RAMageddon”. The cause is the AI data centre boom, which is now setting the price of the world’s high-end memory: AI-server DRAM costs roughly doubled in Q1 2026 alone, AI data centres are expected to consume up to 70% of all high-end memory this year, and hyperscaler capex is projected past US$1 trillion. The squeeze is part of the collateral damage hitting tech prices, and it climbs three rungs: chips, then devices, then your budget.
What is “RAMageddon”, and why did AI-server memory costs roughly double in a single quarter?
“RAMageddon” is the name given to the 2026 memory price surge: AI-server DRAM costs roughly doubled in Q1 alone, according to Deloitte, which expects a fourfold full-year rise. Samsung, SK Hynix and Micron are steering wafer capacity from conventional DRAM (the chips in servers and PCs) and NAND toward HBM (high-bandwidth memory) and server DDR5.
HBM is harder to make and yields fewer bits per wafer, so each HBM unit displaces two or more conventional DRAM wafers. Every wafer steered to AI memory is capacity denied elsewhere, and AI wants a lot: NVIDIA’s GB200 carries 192GB of HBM per GPU, and Rubin will carry 288GB.
The reallocation shows in decisions. Micron is exiting its Crucial consumer brand to prioritise AI customers, and TrendForce logged conventional DRAM contract prices up 90-95%, with 32GB DDR5 kits moving from $270 to $359.99 at retail.
The pressure persists because most AI workloads are built around memory; dial the footprint down and performance breaks. The Stanford DAM tracker charts cost per gigabyte for DRAM, NAND and HBM, though retail lags contract. That chip-market squeeze is already climbing the next rung: devices.
Why does AI’s roughly 70% share of the world’s memory chips push up prices for MacBooks, consoles and phones?
AI data centres are expected to consume up to 70% of all high-end memory in 2026, so AI-set prices become the price everyone else pays. Memory’s share of device bills of materials has grown, and Apple, Microsoft, Sony and Nintendo are all passing that through.
That 70% covers high-end memory: HBM, server DRAM and enterprise SSDs, not every chip sold. Memory has gone from a tenth of a smartphone’s materials cost to two or three times that, roughly 30% of a budget phone and 23% of an entry-level laptop; Gartner expects the sub-$500 entry-level PC segment to disappear by 2028.
Apple’s MacBook Air starts £100 higher, Sony lifted the PS5 by up to £90, and Microsoft added £20-50 to Xbox prices. Nintendo pushed the Switch 2 from $450 to $500, citing the RAM shortage, and Sony may delay its next console to 2028 or 2029.
For your business, the knock-ons are refresh cycles, fleet costs, 40-week-plus DRAM lead times and fewer low-end models. On 2026-27 procurement, weigh timing, volume commitments and flexibility: vendors are letting quotations expire and shifting to allocation-only ordering.
Memory supercycle or boom-bust: will AI memory prices stay elevated or crash?
Plan for two scenarios; credible sources disagree. The supercycle case says AI demand is structural and supply discipline holds, with SK Hynix expecting shortage into 2030. The boom-bust case says 2027 capacity and any AI capex pullback could flip the market. Goldman Sachs Research models undersupply easing from about 4.9% in 2026 to 2.5% in 2027.
The bull case rests on structure: AI demand spans training and inference, suppliers remember past busts, and multi-year deals lock it in. Broadcom has supply locked through 2028, and Deloitte expects the Big Three’s capex up nearly 340% by 2027.
The bear case is history. Every past memory boom has ended in oversupply and price collapses, new capacity lands from 2027, and CXMT is expanding on IPO money. AI capex is discretionary, and a moderation could flip a tight market.
One wrinkle: HBM prices fell while conventional DRAM surged, on HBM3E discounts and the HBM4 launch delay, though the decline should narrow as HBM4 ramps with NVIDIA Vera Rubin and AMD MI455X racks shipping from Q3 2026.
What to watch: contract versus spot versus retail prices, supplier inventory and capacity announcements, and long-term agreement (LTA) signings; the Stanford DAM tracker and TrendForce’s contract-price outlooks track the price side, though the tracker’s HBM series is modelled rather than market data.
Is the AI boom going to make my cloud bill go up, and how do I assess my exposure?
Yes, directionally, for memory- and GPU-heavy workloads in power-constrained regions. The transmission channel: memory is about 30% of AI data centre investment, so DRAM, HBM and SSD prices flow into infrastructure economics before they ever show on an invoice. How much any bill moves depends on your contracts, regions and renewal timing.
No public data traces memory costs into AWS, Azure or GCP price lists, so assess exposure with three variables. One provider has signalled a number: OVHcloud’s CEO expects some cloud prices to rise 5-10% by September 2026, a single vendor’s read.
One: workload intensity, where memory-hungry and GPU-heavy services sit closest; storage-heavy estates face the SSD leg: enterprise SSD prices rose 53-58% in Q1. Even Google’s TPUs carry 192GB of HBM, so the memory exposure carries over to alternative silicon.
Two: region, where grid approvals in congested markets run 24-36 months, and Gartner projects 40% of AI data centres power-constrained by 2027.
Three: renewal timing, whether contracts renew into a rising market with prices committed or floating.
Then estimate pass-through into your pricing: which products carry real hardware cost, and where competitors’ prices have moved? Tiered SLAs and flexible scheduling absorb some of the rest. See why AI data centre capacity keeps slipping, the equipment and labour squeeze behind rising costs, and how the crunch reaches tech buyers.
What should I check when choosing a cloud region or provider, and before signing long-term capacity contracts while builds are delayed?
Choose on power runway, interconnection status, water and cooling fit, and moratorium risk, alongside hardware specs. Before signing, stress-test term length against build slippage, escalation guarantees and exit terms. Australia and Southeast Asia are the main diversification options to weigh.
Australia’s numbers: data centres use 2% of grid electricity today, 6% by 2030, with NSW going from 4% to 11% and modelling warning of 26% higher wholesale prices by 2035. Facilities take 18-24 months to build while transmission takes five to ten years.
Provider diligence means evidence over marketing: signed grid capacity, permitted interconnection, funded milestones. Check the pipeline arithmetic: developers file for five to ten times the capacity they will build, and six in seven early-stage megawatts never land. Then check moratorium risk in the markets you are buying into.
Before signing, ask whether the term length outruns the provider’s build schedule and what exit and portability terms look like if construction slips.
Which brings it back to your budget: chip contract prices became device prices, and device prices are becoming cloud bills and capacity contracts. The surge is structural while the buildout runs, but it is navigable: plan for two futures and stress-test what sits downstream of them. Scenarios, an exposure map and contract diligence beat forecast confidence. See the complete overview of the boom’s collateral damage.
Frequently Asked Questions
Is RAMageddon caused by AI, or by tariffs and geopolitics?
The surge is an AI demand story, not a geopolitics story. Hyperscale buying collided with constrained manufacturing, and Samsung, SK Hynix and Micron steered wafer capacity from conventional DRAM and NAND to HBM and server DDR5. Tariffs can add cost elsewhere in the chain, but they do not explain why conventional DRAM contract prices jumped 90% to 95% in a single quarter.
Why can’t memory makers simply build more factories to end the shortage?
Building memory capacity takes years and billions of dollars before a single wafer ships, so supply cannot respond quickly. Samsung, SK Hynix and Micron are adding capacity, but most of it will not be running until 2027 at the earliest, and Deloitte expects the Big Three’s combined capex to rise nearly 340% between 2024 and 2027. Suppliers are also cautious about overbuilding after past busts.
Why does AI need so much memory in the first place?
AI models need their parameters and working state held close to the processor, and inference workloads keep adding to that footprint as context grows. AI accelerators show the scale: a single NVIDIA GB200 GPU carries 192GB of HBM, and the Vera Rubin platform carries up to 288GB of HBM4 per GPU. Demand scales with every rack deployed.
What is HBM, and why can’t AI just use ordinary RAM?
HBM, or high-bandwidth memory, stacks DRAM dies and connects them for far more bandwidth than standard DDR5, which is why AI accelerators depend on it. That complexity is also the problem: HBM is harder to manufacture and yields fewer bits per wafer, so one HBM unit displaces two or more conventional DRAM wafers and starves commodity supply.
Is it true that AI data centres will use 70% of the world’s memory?
Not exactly. The figure applies to high-end memory: HBM, server DRAM and enterprise SSDs, where AI data centres are expected to consume up to 70% of that supply in 2026, not 70% of all memory. The direction is what matters. Consumer and commodity segments get the leftovers, which is why AI-set prices become the price everyone else pays.
Will SSDs and flash storage get more expensive too?
Yes. The squeeze is a memory story, not a RAM-only story. Manufacturers have shifted capacity towards higher-margin AI parts, and TrendForce logged NAND contract prices up 55% to 60% quarter-on-quarter, with enterprise SSD prices up 53% to 58%. If your estate is storage-heavy, that leg of the bill is exposed alongside DRAM and HBM.
Will graphics cards get more expensive too?
Yes. Graphics cards carry their own dedicated memory, so they face the same shortage and its knock-on effects. Console pricing shows how makers respond when memory costs jump: Sony raised PS5 prices by up to £90, Microsoft’s Xbox rose £20 to £50, and Nintendo lifted the Switch 2 from $450 to $500 while citing the RAM shortage.
Why did Micron stop selling Crucial memory to consumers?
Because the industry is prioritising AI buyers. Micron exited its Crucial consumer business to improve supply and support for larger strategic customers, and the same logic runs across the Big Three: every wafer steered to AI-grade memory is capacity denied to consumer channels. Expect fewer consumer-focused products and less discounting while AI margins dominate.
Should I buy a new laptop or PC now, or wait for prices to fall?
Treat it as a scenario test rather than a forecast. If you need the device or upgrade now, buying into a rising market can beat waiting: retail 32GB DDR5 kits moved from about $270 to $359.99, and prices are still climbing. If you can defer, a boom-bust turn could bring relief once new capacity lands from 2027, but plan for high prices in the meantime.
Are refurbished devices a good way to avoid the price rises?
Often, yes. Refurbished electronics are cheaper, and their prices tend to rise with residual values rather than full component costs, so increases are likely smaller than on new devices. Two cautions: older models still on sale can be marked up by opportunistic retailers, and repairs may beat replacement. If you must replace something, moving sooner rather than later is the safer bet while prices climb.
What is a long-term agreement (LTA), and why does it matter for buyers?
An LTA is a multi-year supply contract that locks in volumes and often price floors or ceilings. They matter because scarce supply is being committed years ahead: hyperscalers have moved from one-year deals to long-term contracts, Broadcom is locked in through 2028, and buyers who sign early secure allocation while everyone else competes for what remains.
How can I track memory prices and know when the shortage is easing?
The Stanford DAM Memory Price Tracker charts $/GB for DRAM, NAND and HBM, while TrendForce and Counterpoint Research publish contract-price outlooks and shortage forecasts. Keep two caveats in mind: retail prices lag contract prices, and HBM has no public spot market, so its figures are analyst estimates. Track contract, spot and retail prices together, since each moves at a different speed.
