Research and drafting: Claude (Anthropic) — “AI-in-Chief”. Editing and publication: Geoff Leask.
Figures checked 27 August 2026. All claims are linked to sources; statements that are the authors’ judgment rather than sourced fact are marked [Analysis].
TL;DR
- DRAM contract prices are still rising, but the pace of increase has collapsed: roughly +90–95% quarter-over-quarter in Q1 2026, ~+60% in Q2, and a projected +13–18% in Q3.
- The slowdown is caused by buyers refusing to pay more, not by improved supply. The shortage itself is intact.
- The viral “AI bubble” stories contain real events — US$130B of data-centre projects blocked or delayed in Q1 2026, and OpenAI’s famous 40%-of-global-DRAM deal turning out to be a non-binding letter of intent — but none of them has returned wafer capacity to consumer memory.
- Every major supplier and analyst puts meaningful supply relief at late 2027 to 2028, when new fabs come online. SK Hynix’s CEO calls 2027 the industry’s worst supply year on record.
- The one scenario that lowers prices sooner is a genuine AI capital-expenditure pullback. Memory supercycles historically end in crashes, and forecasters have been badly wrong before — in both directions.
1. Where prices actually are
Street pricing tells the story faster than contract indices. Figures below are point-in-time observations from the linked reporting, not continuous tracking.
| Item | ~Early/Mid 2025 | Late 2025 – Mid 2026 | Change | Source |
|---|---|---|---|---|
| DDR5-6000 32GB kit | ~US$120 | US$400+ | ~3.3× | igor’sLAB (PCPartPicker data) |
| DDR5 64GB kits | — | US$700+ (many configs) | — | igor’sLAB |
| DDR4 modules (various) | baseline | roughly doubled to tripled | 2–3× | igor’sLAB |
| NVIDIA RTX PRO 6000 Blackwell 96GB (GDDR7) | US$8,565 at launch (Mar 2025) | US$13,250 (Jun 2026) → US$16,000 (Aug 2026) | +87% | Tom’s Hardware |
| Memory as share of a PC’s bill of materials | ~20% | ~35% | +15 pts | StorageSwiss |
Two second-order data points show how far upstream the squeeze reaches: Apple removed the 256GB and 512GB memory options from the Mac Studio in March and May 2026 amid rising RAM prices (MacRumors), and Micron discontinued its Crucial consumer memory brand to prioritise data-centre customers (igor’sLAB).
2. The trajectory: deceleration, not decline
TrendForce’s contract-price surveys across 2026:
| Quarter | Conventional DRAM (QoQ) | NAND Flash (QoQ) | Source |
|---|---|---|---|
| Q1 2026 | +90–95% | +55–60% | The Register |
| Q2 2026 | +58–63% (projected at the time) | — | StorageSwiss |
| Q3 2026 | +13–18% (projected) | +10–15% | Tom’s Hardware |
That looks like a market cooling off. The mechanism matters, though: TrendForce attributes the Q3 moderation to consumer-electronics manufacturers being unwilling and unable to absorb further cost increases — demand destruction at the buyer’s end — not to any improvement in supply (Tom’s Hardware). Server DRAM remains undersupplied, with a portion of purchases cushioned only by long-term contracts.
[Analysis] In other words: the second derivative turned down while the first derivative stayed positive. Prices are still compounding on a base that already tripled. A slowdown in increases is not a path to lower prices — it may simply mark the level at which the consumer market stops clearing.
3. How the shortage happened
The proximate trigger was demand-side and sudden. In October 2025, OpenAI signed agreements with Samsung and SK Hynix — as part of its Stargate initiative — for up to 900,000 DRAM wafers per month, roughly 40% of global DRAM output (Conrad Gray / Medium). The announcement hit a market whose safety stock was already thin, and prices moved within weeks.
The structural driver underneath is the reallocation of wafer capacity from conventional DRAM to High Bandwidth Memory (HBM) for AI accelerators. StorageSwiss characterises this as a structural reallocation rather than a cyclical shortage, with DRAM lead times stretching beyond 40 weeks and TrendForce projecting the total memory market to grow from US$551.6B in 2026 to US$842.7B in 2027 (StorageSwiss).
4. The “bubble is bursting” stories — what’s true and what isn’t
The viral claims each trace back to real events. None of them, so far, has translated into cheaper memory.
| Claim circulating online | The underlying facts | Why memory prices didn’t fall |
|---|---|---|
| “A whole bunch of data-centre projects have been cancelled” | True: 75 US AI data-centre projects worth ~US$130B were delayed or rejected in Q1 2026 alone (TechRadar), vs US$156B for all of 2025. Of 12–16GW planned for 2026, only ~5GW was under construction (Yahoo Tech). | Causes are grid capacity, transformer/switchgear shortages and community opposition — not collapsing demand. UBS’s August read: 25GW under construction globally, no cooling, 2026 growth forecast raised to 20–25% (UBS via Deep News). |
| “The big RAM agreement fell through” | Substantially true: the OpenAI–Samsung–SK Hynix arrangement was a non-binding letter of intent, with minimal actual 2026 delivery volumes (The Buildout). Oracle and OpenAI also scrapped the Abilene Stargate expansion in March 2026 (The Deep Dive). | The demand signal had already reshaped supplier behaviour — capacity reallocations and consumer-line exits (e.g. Crucial) — before any wafers shipped. Meanwhile, harder commitments replaced the soft ones: ~US$950B in supply deals signed in July 2026 lock HBM to Nvidia and Broadcom through 2030 (TechTimes). |
| “Spot prices crashed — OpenAI cancelled orders” | Spot volatility in early 2026 was attributed by reporting to Chinese stockists offloading hoarded inventory into a weakening secondary market, with long-term contract prices unaffected (The Buildout). | One-off inventory dumps move spot prices for weeks, not contract prices for quarters. |
[Analysis] The uncomfortable takeaway is that the shortage no longer needs its original trigger. Even if the OpenAI demand never fully materialises, the capacity that left the consumer market has not returned, and the July 2026 mega-deals re-anchored demand with counterparties far less likely to wobble.
5. When supply relief actually arrives
| Event / forecast | Timing | Source |
|---|---|---|
| Samsung: “significant shortages” across memory products persist | Through at least 2027 (cleanroom space constrains 2026–27 expansion) | Tom’s Hardware, DCD |
| SK Hynix CEO: 2027 the industry’s “worst year” for supply; demand may exceed capacity beyond 2030 | Stated 10 July 2026 | ArticSledge (citing Reuters) |
| TrendForce: 2027 server RDIMM bit supply grows only 15–20% YoY, lagging demand | 2027 | ArticSledge |
| Gartner: ~3.6% market undersupply persists | Through 2027 | WION (citing Gartner) |
| Goldman Sachs: Samsung DRAM average selling prices rise a further ~27% | 2027 | WION (citing Goldman Sachs) |
| SK Hynix M15X fab in volume | Mid-2027 | DCD |
| Samsung P5 (Pyeongtaek) operational | 2028 | DCD |
| Earliest plausible window for sustained retail price declines | Late 2027 – 2028 | ArticSledge |
For context, one scenario analysis published in January 2026 assigned 60% probability to normalisation by Q1–Q2 2027 (SoftwareSeni); events since — the supplier warnings above — have shifted the consensus toward that analysis’s worst case.
6. The scenario where prices fall early
There is one credible path to cheaper RAM before new fabs arrive: a genuine AI capex pullback, in which cloud providers cut GPU orders, HBM demand softens, and wafer capacity flows back to consumer DDR5 (RAM Price History). Two reasons not to dismiss it:
- History. Memory supercycles end abruptly: the 2016–17 shortage ran ~16 months and the 2020–21 shortage ~14 months before reversing, and past booms (2017–18) were followed by steep crashes once oversupply hit (SoftwareSeni, BaCloud).
- Forecast humility. As late as November 2024, TrendForce and most industry watchers predicted DRAM prices would decline through 2025 (BaCloud). They were spectacularly wrong. The error bars on today’s confident shortage forecasts deserve the same scepticism.
[Analysis] Signals we would treat as a genuine turn rather than a wobble: a hyperscaler publicly cutting AI capital expenditure; spot prices declining for two or more consecutive months; or supplier inventory-weeks visibly climbing. Absent those, dips are likely inventory noise.
7. What this means if you’re buying hardware
[Analysis — this section is the authors’ judgment, not sourced forecasting.]
- Don’t budget on reversion. Base case for the next 12 months is continued increases at a decelerating rate, with a possible plateau in 2027. Plan purchases assuming today’s prices or worse.
- Buy memory-heavy items early, compute-heavy items late. RAM and VRAM capacity is the component class with the clearest upward pressure; raw compute is comparatively better supplied.
- Configuration RAM is the trap. Vendors that solder or fix memory at order time (Apple being the obvious example) pass the shortage through with no aftermarket escape hatch — and have already begun withdrawing high-capacity options entirely.
- Stay trigger-ready. If a correction comes, history says it comes fast. Keep funds liquid and watch the turn signals in §6 rather than waiting for a scheduled review date.
Sources
- Tom’s Hardware — Memory price surge begins to cool as consumers hit affordability limit
- The Register — DRAM prices expected to double in Q1
- StorageSwiss — Memory and Flash Prices Are Not Coming Down (Through 2027)
- igor’sLAB — Memory bottlenecks until at least Q4 2027
- ArticSledge — When Will RAM Prices Go Down? (prices checked 24–25 Aug 2026)
- Tom’s Hardware — Samsung and SK Hynix warn AI-driven memory shortages could last until 2027 and beyond
- Data Center Dynamics — Samsung and SK Hynix post record profits but warn shortages persist into 2027
- Conrad Gray (Medium) — OpenAI crashed the global memory supply
- The Buildout — OpenAI memory supply agreement limited to letter of intent
- The Deep Dive — OpenAI locked up 40% of global RAM with no obligation to buy any of it
- TechRadar — $130 billion worth of AI data center projects cancelled or delayed in Q1 2026
- Yahoo Tech — Half of planned US data center builds face delays or cancellations
- Tech Insider — US AI data center delays: capacity crisis (Omdia cost data)
- UBS report via Deep News — UBS dismisses AI bubble concerns
- TechTimes — Samsung and SK Hynix lock in AI chip supply through 2030 with $950B in US deals
- Tom’s Hardware — Nvidia doubles RTX PRO 6000 Blackwell’s MSRP to $16,000
- Thunder Compute — NVIDIA RTX PRO 6000 Blackwell pricing (GDDR7 shortage as driver)
- MacRumors — Mac Studio roundup (RAM options removed March/May 2026)
- SoftwareSeni — When Will DRAM Prices Normalise?
- BaCloud — When Will RAM Prices Drop? Global memory market outlook
- WION — Will RAM prices go down in 2027? (citing Gartner and Goldman Sachs)
Disclosure: this article was researched and drafted by Claude Fable 5 (Anthropic’s AI assistant) from web sources retrieved 27 August 2026, and reviewed and edited by a human before publication. Figures marked as projections are third-party forecasts and may not age well — see §6 on forecast humility.
