DDR5 contract ▲ 80–90% QoQ · HBM3E sold out through next FY · Wafer mix → reallocating to HBM · DC power 485 TWh → ~950 TWh by 2030 · Accelerated servers +30% demand/yr · Relief forecast ≥ 2027, contested · DDR5 contract ▲ 80–90% QoQ · HBM3E sold out through next FY · Wafer mix → reallocating to HBM · DC power 485 TWh → ~950 TWh by 2030 · Accelerated servers +30% demand/yr · Relief forecast ≥ 2027, contested ·

Memory & Compute Desk

There is no RAM on the shelf, because the shelf is now an accelerator.

Memory is a finite wafer budget. Every stack of high-bandwidth memory stacked next to an AI accelerator is silicon that will not become a laptop module, a phone chip, or a router in a factory.

See the signals we track Methodology

Scarcity rarely arrives as an empty warehouse. It arrives as a lead time, a line-item surcharge, and a spec sheet that quietly got smaller.

The shortage in one paragraph

Why consumer memory prices moved before consumer demand did.

Memory fabs have roughly fixed capacity in the short run. When accelerator vendors began reserving a large share of that capacity for high-bandwidth memory (HBM) — a lower-bit-per-wafer, margin-rich product — the wafer supply left over for conventional DDR5 and DDR4 contracted. Fewer bits chasing the same laptop, handset and server demand means price rationing first and quantity rationing second. Buyers see it as surcharges and allocation; smaller buyers see it as quotes that expire in 48 hours.

+400%Estimated DRAM price rise, start of 2024 to end of 2026 (J.P. Morgan Global Research estimate)

~90%Quarter-over-quarter jump in conventional DRAM contract prices reported for early 2026 (Counterpoint, TrendForce-derived)

~70%Share of high-end DRAM output projected to be absorbed by AI data centres in 2026 (reported by WSJ)

2030+Earliest plausible normalisation, per supplier warnings; independent forecasts split between 2027 and the decade's end

Index, Q1 2024 = 100. Reconstructed illustration of the "~+400% by end of 2026" estimate — the intermediate points are interpolated, not measured. Replace with a real series before publishing.

Where the wafers went

HBM is a wafer-multiplier: the same die area yields fewer usable gigabytes than commodity DRAM, and it is bought first by the largest purchasers.

AI data centres ~70%

Everyone else ~30%

Structural, not rumoured

Long-term supply agreements lock output a year or more ahead.
Capacity additions take new fabs: 18–36 months, not one quarter.
Older nodes (DDR4) retire as capacity chases newer nodes, shrinking fallback supply.

The energy overlay

Data-centre electricity projected to roughly double, 485 TWh (2025) to ~950 TWh (2030) — about 3% of global demand (IEA).
Accelerated-server demand growing ~30% per year vs ~9% for conventional servers (IEA).
Grid interconnection and power capacity, not chips, now gate many new builds.

Second-order effects

Device makers trim base memory configs instead of raising sticker prices.
Dual-sourcing and lifetime-buy orders distort the visible market.
Refurbished and second-hand modules reprice faster than new retail.

Who absorbs the cost first

Qualitative exposure ranking; sensitivity depends on memory content per unit and contract cover.
Segment	Why exposed	Visible symptom
Consumer PCs & workstations	High memory content per unit, thin margins, spot purchasing	Config downgrades, bundled RAM cuts, price hikes
Game consoles & set-top	Fixed BOM, long product cycles locked to old pricing	Delayed revisions, storage trims, mid-cycle repricing
Smartphones	DRAM + NAND both compete for the same capacity	Base tiers shrink, Pro-tier gap widens
Automotive & industrial	Low priority for suppliers, qualification locks out substitutes	Long lead times, allocation queues
Cloud & hyperscale	Contract cover, but power and racks are the binding limit	Instance price rises, capacity waitlists

Signals worth watching

Six indicators that separate a real inflection from a headline.

Contract vs spot

Spread direction and width
Quote validity windows shrinking
Module makers' gross margins

Wafer allocation

HBM share of total DRAM wafer input
DDR4 end-of-life notices
Capex announcements and fab start dates

Inventory

Supplier and OEM weeks of inventory
Channel stock in distribution
Lifetime-buy volume

Power & grid

Interconnection queue length
Regional industrial tariff pressure
Reported PUE and utilisation

If you are buying in this market

Individuals

Buy the memory you'll actually use, not a round number.
Prefer upgradeable sockets over soldered base configs.
Second-hand enterprise modules are a legitimate supply.

Teams & procurement

Forecast 2–4 quarters out; spot buys price the panic.
Split volume across suppliers, accept mixed sourcing.
Recover idle RAM in existing fleets before ordering.

Engineers

Right-size instances; memory is now the expensive axis.
Quantise, offload, cache to cheaper tiers.
Measure cost per served request, not per GPU-hour.

Methodology & sources

Figures are third-party estimates gathered from public reporting; each should be re-verified against the primary source and dated before this page goes live.

Counterpoint Research — quarterly DRAM price movement reporting.
J.P. Morgan Global Research — DRAM price trajectory estimate through end of 2026.
TrendForce-derived contract price tracking, via secondary reporting.
The Wall Street Journal — share of high-end DRAM absorbed by AI data centres, 2026 projection.
IEA, "Energy and AI" / "Key Questions on Energy and AI" — data-centre electricity baselines and 2030 projections.
Supplier commentary (memory manufacturers) on shortage duration.

The chart's intermediate points are interpolated for illustration and are not measured data. Forecasts of when supply normalises conflict materially between sources; we show a range rather than a date.