For decades, memory has been the most reliably cyclical commodity in tech: demand rises, prices rise, buyers pull back, supply catches up, prices fall. Most buyers have internalised that rhythm, and 2026 looked like the next turn of the wheel. Then the data stopped cooperating. Conventional DRAM contract prices rose 90 to 95% quarter-on-quarter in Q1 2026, with a further 58 to 63% expected in Q2, even though consumer demand is soft. That is the one combination the old model says cannot happen. This is the memory crisis rewriting software economics. The answer turns on two mechanisms, and neither behaves like a normal cycle.
Why is the memory crisis structural rather than a normal supply-and-demand cycle?
The memory crisis is structural because the normal boom-bust correction has stopped working. In that cycle, prices rise, demand cools, supply catches up and prices fall again. Here, capacity is committed and reallocated before conventional demand is ever met, so higher prices no longer trigger the familiar correction.
That boom-bust pattern held through 2010, 2016 and 2021. The difference is that the price signal and the demand signal have come apart: DRAM contract prices are rising sharply while consumer demand is weak. TechInsights puts it plainly: this is no longer a normal cycle.
Why does demand destruction not kick in? Because memory is now a minimum-viable input to AI infrastructure, not an optional upgrade. Cutting memory capacity cuts compute performance, so buyers keep paying. IDC frames memory as a strategic infrastructure input whose demand compounds between quarters rather than normalising. There is no hidden buffer waiting to be released: DRAM inventory fell from 31 weeks in early 2023 to roughly eight weeks by late 2025, and lead times on larger orders stretch beyond 40 weeks.
The cyclical view would need demand destruction, spare inventory and a fast supply response. None are present. TrendForce’s DRAM and NAND contract price releases are the reference point; track the trend through supplier earnings calls and your distributors’ allocation notices. Turning those forward prices into a cost model starts here.
What is high-bandwidth memory (HBM), and why does producing it starve conventional DRAM supply?
The first mechanism is physical.
HBM is stacked DRAM with the dies bonded by through-silicon vias (TSVs), placed beside a GPU or AI accelerator to deliver bandwidth conventional DDR5 cannot. Each HBM3E bit consumes about three times the wafer supply of a DDR5 bit (and HBM4 widens that gap), so every HBM ramp is a zero-sum conversion that displaces conventional DRAM output.
The bandwidth gap explains why AI cannot swap back to DDR5. HBM3E delivers around 8 terabytes per second per package; a DDR5 module manages roughly 100 gigabytes per second, leaving expensive compute idle when memory cannot feed it. When Samsung, SK Hynix and Micron reallocate wafer lines toward HBM for NVIDIA’s accelerators (where H200 ships with 141GB of HBM3E and Blackwell Ultra 288GB), conventional DRAM output shrinks.
This is where the HBM and DDR5 price stories meet. The 90 to 95% then 58 to 63% surges are conventional DDR5 and RDIMM repricing, but HBM demand is the engine pulling capacity away. Server RDIMMs out-earned HBM per wafer in Q1 2026, a sign of how tight the conventional pool has become.
The packaging layer makes it harder still. HBM also consumes CoWoS advanced packaging capacity, and TSMC supplies roughly 90% of it from Taiwan, concentrating the bottleneck in one place. That displacement through to endpoint and server pricing is covered here.
How do hyperscaler long-term agreements (LTAs) lock away 2027 memory supply?
Even the conventional DRAM that survives wafer crowding is already spoken for. That is where long-term agreements come in.
Hyperscalers lock away 2027 supply by signing multi-year, often pre-paid long-term agreements (LTAs) with Samsung, SK Hynix and Micron. These agreements reportedly book nearly all 2027 DRAM and HBM output before it is produced, leaving conventional buyers to compete for the residual.
An LTA converts a spot shortage into a structural one by selling capacity before it exists. Google, Microsoft and Amazon have reportedly signed three-to-five-year contracts with prepayments, and hyperscalers have reportedly locked in almost all 2027 DRAM production with advance deposits. The result is a two-tier market. Committed hyperscaler volume sits on one side, and you’re left fighting over a thin allocation on the other.
New supply cannot rescue 2027. A leading-edge fab takes three to five years from groundbreaking to meaningful output, and a die shrink takes 12 to 18 months to qualify. The crunch also extends beyond DRAM: NAND flash contract prices are expected to rise 70 to 75% in Q2 2026 on enterprise SSD demand, so the datacentre cost shock compounds.
If you don’t have hyperscaler-scale LTAs, your response is longer planning horizons and qualifying a second source early. And the software margin exposure from the residual allocation is where the story reaches your own economics.
The old cycle isn’t coming back
The two mechanisms resolve the puzzle the price data posed. HBM wafer crowding physically removes conventional DRAM capacity, and pre-paid hyperscaler LTAs pre-sell the 2027 output that remains. With capacity committed before demand is met, no price signal can trigger the old self-correction.
That means memory is no longer a commodity that oscillates; it is strategic infrastructure being repriced. Higher prices now signal structural scarcity, not a cyclical peak, so waiting for the correction is a losing plan. The full structural repricing and what the crunch does to your cloud bill are the practical next steps. The old cycle is not coming back to rescue you if you wait for it.
Frequently Asked Questions
Why have RAM and SSD prices gone up so much?
RAM and SSD prices are being repriced because AI infrastructure is consuming the supply that would otherwise reach ordinary buyers. Conventional DRAM contract prices rose 90 to 95% quarter-on-quarter in 1Q26, and TrendForce expects a further 58 to 63% rise in 2Q26. NAND flash and enterprise SSDs face the same pressure, so the increases flow through to almost every system you buy.
Is this the same as the 2018 or 2021 memory price cycles?
No. Earlier upswings were cyclical peaks: prices rose, demand cooled, suppliers added capacity and prices fell again within a few quarters. This time capacity is reallocated to HBM and pre-sold through multi-year agreements before demand is met, so the self-correction that ended those cycles cannot engage. The structural differences, not the size of the increase, are what set 2026 apart.
Could an AI spending slowdown or a recession bring prices back down?
Unlikely in the near term. Even a sharp pullback in AI spending would not release the capacity already committed to HBM and locked into hyperscaler agreements, and memory is now a minimum-viable AI input rather than an optional upgrade. New fabs take three to five years and die shrinks take 12 to 18 months to qualify, so supply cannot respond quickly enough to rescue prices.
Can I dodge the price rises by switching memory type or supplier?
Only partly. Switching between DDR5 suppliers rarely helps when all three, Samsung, SK Hynix and Micron, are reallocating the same wafer capacity toward HBM. Older DDR4 is also rising as lines convert. A genuine hedge is qualifying a second source early and committing volume on longer contracts, because availability, not headline price, is now the binding constraint.
Does the memory premium affect consumer laptops and phones, or only datacentre buyers?
It reaches everyone eventually. Datacentre buyers absorb capacity first and at the highest prices, but the same wafer shortage feeds through to notebooks, desktops and phones as suppliers prioritise higher-margin server and HBM products. Consumer demand is soft, yet conventional DRAM still rose 90 to 95% quarter-on-quarter in 1Q26, which shows the increases are supply-driven rather than demand-driven.
How long is the structural repricing likely to last?
At least through 2027, and probably longer. Because nearly all 2027 DRAM and HBM output is reportedly booked before production begins, and new fabs take three to five years from groundbreaking, there is no fast supply response that could reset prices. Plan against elevated memory costs for the medium term rather than waiting for a correction that the market’s structure has removed.
Why do spot market prices differ so much from contract pricing?
Spot and contract prices measure different things. Contract pricing reflects negotiated, often multi-quarter agreements and is the benchmark TrendForce tracks; spot pricing moves faster as brokers react to short-term scarcity. When hyperscaler long-term agreements absorb most forward output, the residual volume left for spot and smaller buyers is thin, so spot premiums swing far more sharply than the contract increases do.
DDR5 price increases versus HBM price increases: which one actually drives the repricing?
Both matter, but for different reasons. HBM demand is the engine that pulls wafer capacity away from conventional production, while the DDR5 and RDIMM increases, 90 to 95% then 58 to 63% quarter-on-quarter, are the visible repricing ordinary buyers actually pay. Server RDIMMs briefly out-earned HBM per wafer in 1Q26, which shows how tight the conventional pool has become.
Where can I find TrendForce’s DRAM and NAND contract price data?
TrendForce publishes quarterly DRAM and NAND flash contract price releases, which are the figures most media and analysts cite for the 90 to 95% and 58 to 63% quarter-on-quarter increases. They are available through TrendForce’s own research releases and are widely summarised in technology press. Treat them as the primary reference when you need a defensible number for budget or board papers.
Where can engineering leaders track memory price trends and forward forecasts?
Beyond TrendForce’s contract price releases, watch supplier earnings calls from Samsung, SK Hynix and Micron, CoWoS packaging updates from TSMC, and lead-time or allocation notices from your distributors. Contract price trackers and hyperscaler capital expenditure disclosures add context. Tracking two or three of these consistently tells you more about forward availability than any single price headline.
What can buyers without hyperscaler-scale LTAs do to protect themselves?
Conventional buyers face the residual allocation, so the practical defences are longer planning horizons, early dual-sourcing and qualifying a second supplier before you need it. Committing volume on multi-quarter contracts helps secure supply even at high prices. Accept that lead times and allocation, not negotiation, now determine what you can actually buy, and build that assumption into your budgets.
