If you run UK infrastructure, the instinct you half-remember - "hold off, memory always gets cheaper" - is exactly wrong for this cycle. Server DRAM contract prices have risen more than fivefold since Q3 2025, TrendForce still had them climbing through Q3 2026, and the shortage now stretches into 2027. This is a deliberately hard-nosed market-timing take: why "wait for prices to fall" backfires in 2026, what the delay actually costs in pounds, and how to phase purchases so a memory spike does not detonate your refresh budget.
The 2026 memory spike, in one line
Memory, not the CPU or the chassis, is what has repriced the whole server. TrendForce's quarterly contract-price data tells the story: server DRAM rose 43-48% quarter-on-quarter in Q4 2025, conventional DRAM jumped a further 93-98% in Q1 2026, then 58-63% in Q2 2026, before the increase moderated to 13-18% for Q3 2026. Compound those steps and a unit of server memory that indexed to 100 in Q3 2025 sits above 520 a year later - an indicative 5.3x.
The macro numbers are just as stark. TrendForce put total DRAM industry revenue at roughly US$97bn in Q1 2026, an 81% jump in a single quarter, with Samsung, SK hynix and Micron all posting 60-95% revenue gains. On the supply side, reporting through the period suggested memory makers were filling only around 70% of server DRAM orders. This is not a soft, cyclical wobble; it is an allocation market, and buyers are competing for parts that are being deliberately rationed.
The cause is structural rather than seasonal. Fab capacity is being diverted to high-bandwidth memory (HBM) for AI accelerators, which crowds out conventional DDR5 and DDR4 server DRAM. Our own AI server data study shows why that pull is so relentless: a single training node can carry more memory than a rack of general-purpose servers did a generation ago.
Why "wait for prices to fall" backfires this cycle
In a normal DRAM cycle, patience is rewarded. Prices overshoot, suppliers over-build, a glut arrives, and the buyer who waited two or three quarters pockets the difference. That reflex is baked into most procurement teams. It is also the single most expensive assumption you can make in 2026.
Three things have broken the old playbook. First, the demand shock is AI capex, which is structural and price-insensitive - hyperscalers are signing long-term agreements (LTAs) that lock supply away from the merchant market. Second, the big three suppliers are prioritising HBM and high-end server DRAM margins over volume, so there is little incentive to flood the market. Third, TrendForce's own forecasts had prices still rising, not falling, through Q3 2026, with meaningful relief unlikely before 2027. Micron's leadership publicly guided tightness into 2027.
So the buyer waiting for the dip is not waiting out a three-month spike; they are betting against a multi-year, supply-constrained trend that every major analyst house expected to persist. When the reference price is climbing every quarter, delay is not a saving strategy - it is a compounding cost. If you must move workloads or extend runway rather than buy, a consolidation route such as our VMware alternatives sizing tends to beat holding out for a memory price that is not coming.
What waiting actually costs: the phased-buy vs wait model
Put pounds on it. Take a modest four-node refresh where each node needs about £10,000 of memory at Q3 2025 prices, and price each option off the DRAM index above. Buying all four upfront at Q3 2025 would have cost roughly £40,000. Phasing one node per quarter across the year - paying the going rate each time - lands near £98,700. Doing what the "wait for the dip" instinct suggests, and buying all four at Q3 2026, costs about £211,000. That is 5.3x the upfront figure and more than double the phased approach.
The point is not that hindsight is cheap - of course Q3 2025 was the moment to buy. The point is directional and forward-looking. In a rising market, the buyer who committed early beat the buyer who waited by a factor of five, and even the buyer who dripped purchases in monthly still beat the waiter by more than 2x. Every quarter of delay in 2026 has behaved like a tax, not a discount.
For a buyer standing in mid-2026, that reframes the decision. You cannot rewind to last year's prices, but you can refuse to repeat the mistake by holding out for 2027. Model the cash impact against your own build with our IT finance calculator before you commit a phasing plan - financing the memory-heavy portion now can be cheaper than paying a higher cash price later.
Where memory sits in your bill of materials
Memory has become dear enough to move the whole quote, because it is no longer a rounding error on the invoice. On a memory-rich general-purpose server - SemiAnalysis models a 1TB, 512GB-per-socket CPU box - DRAM is close to 40% of the bill of materials, with the CPU, mainboard and everything else around 45% and storage roughly 15%. Lighter, mainstream configurations typically run 20-30% memory, but the direction of travel is one-way.
That share is exactly why the 2026 spike is so painful. If memory is 40% of the box and its unit price has risen roughly fivefold while the rest of the bill barely moves, the finished-server price lands near 2.6x its former level - an uplift of about 150-170%, not the modest bump buyers expect. Even a lighter, mainstream config where memory is 20-30% of the box still faces an 80-120% rise. Buyers who benchmark against last year's per-server price and assume a small uplift are consistently under-budgeting.
It also changes where the smart savings are. Trimming the storage tier or shaving a CPU SKU no longer moves the needle much; getting the memory decision right does. That is a strong argument for speccing carefully rather than over-provisioning DIMMs "to be safe" - use a disciplined build tool such as our server configuration workflow, or the Lenovo configurator, to right-size capacity to the workload instead of the habit.
The UK buyer angle
UK buyers carry two extra risks on top of the global price curve. The first is currency: DRAM is priced and contracted in US dollars, so any sterling weakness stacks on top of the dollar price increase. A 5x dollar move can land as an even larger pound move if the exchange rate turns against you mid-cycle, which makes locking pricing and lead times more valuable than usual.
The second is queue position. With suppliers filling only part of the order book, distribution allocation matters. UK resellers with standing supplier relationships and forward-committed stock can quote availability that the spot market cannot, and being early in the queue is worth real money when parts are rationed. Lead times on high-capacity RDIMMs have stretched, so a "we'll order it when we need it" posture risks both a higher price and a slipped project date.
There is also a legitimate demand-reduction lever unique to this moment. Not every workload needs new DDR5. Extending the life of kit you would otherwise refresh - via third-party maintenance once it passes vendor end-of-support - keeps capacity in service without buying into a peak memory market. Check where your fleet sits with our server end-of-life checker before you assume a refresh is unavoidable.
A practical phasing playbook
"Phase your purchases" is easy to say and easy to fumble. The aim is to commit enough now to hold your project timelines and your queue position, while spreading exposure so you are not betting the whole budget on one quarter's price. In a market that is rationed and rising, being in the queue beats being clever about the exact bottom.
Concretely, that looks like a short set of moves you can apply this quarter:
- •Buy the memory you can firmly justify now - do not wait for a dip that analysts do not forecast until 2027.
- •Lock pricing and lead times in writing; in an allocation market a firm quote is worth more than a lower spot number you cannot actually fill.
- •Right-size DIMM capacity to the workload with a configurator rather than over-provisioning "to be safe" - every extra DIMM is now a 5x-inflated line item.
- •Use certified refurbished servers and reclaimed DDR4/DDR5 for non-critical, dev/test and edge tiers, where you do not need to buy into peak pricing.
- •Extend, don't replace, where you can - pair third-party maintenance with an end-of-support plan to keep serviceable kit running past its refresh date.
- •Finance the memory-heavy tranches rather than paying peak cash prices upfront, so cash flow, not the DRAM spot market, sets your cadence.
When might it ease - and the honest caveats
The optimistic reading is that Q3 2026's moderation to a 13-18% quarterly rise is the first sign the vertical part of the curve is flattening as consumer demand hits affordability limits and high base effects bite. The realistic reading is that flattening is not the same as falling: TrendForce and Micron both pointed to tightness persisting into 2027, and no credible forecast in the period had prices returning to early-2025 levels within the planning horizon. Slower increases still mean higher prices.
Treat every forward number here as directional, not gospel. Contract prices are volatile, our compounded index blends server and conventional DRAM figures, and the pound-denominated modelling is illustrative and ex-VAT - your actual quote depends on capacity, form factor, timing and exchange rate on the day. The takeaway does not hinge on the exact multiple, though. Whether memory has risen 4x or 5.5x, the strategic conclusion is identical: this is a supply-constrained market where waiting has cost money every quarter.
So the answer to "buy now or wait?" is: buy what you genuinely need now, phase the rest, and do not price your 2026 plan on a 2027 rescue that may not arrive. If the spike is squeezing budgets, the highest-leverage responses are consolidating workloads onto fewer, denser hosts, extending serviceable kit, and financing the memory-heavy portion - not sitting on your hands and hoping the old cycle repeats.
Sources
The market data, price steps and cost-share figures in this article draw on the following primary and secondary sources.
- •TrendForce — AI Server Demand Continues to Support Memory Prices in 3Q26 (conventional DRAM +13-18% QoQ): https://www.trendforce.com/presscenter/news/20260703-13134.html
- •TrendForce — Rapid Contract Price Surge Drives 1Q26 DRAM Industry Up 81% QoQ (1Q26 +93-98%, 2Q26 +58-63%, US$97bn revenue): https://www.trendforce.com/presscenter/news/20260601-13070.html
- •TrendForce — DRAM Prices to Continue Rising in 4Q25, Server Demand Surges (server DRAM +43-48% QoQ): https://www.trendforce.com/presscenter/news/20250924-12733.html
- •SemiAnalysis — AI Server Cost Analysis: Memory Is The Biggest Loser (server BOM memory ~40%): https://newsletter.semianalysis.com/p/ai-server-cost-analysis-memory-is
- •TechPowerUp — Server DRAM Pricing Jumps 50%, Only 70% of Orders Getting Filled: https://www.techpowerup.com/342331/server-dram-pricing-jumps-50-only-70-of-orders-getting-filled
- •Network World — Server memory prices could double by 2026 as AI demand strains supply: https://www.networkworld.com/article/4093752/server-memory-prices-could-double-by-2026-as-ai-demand-strains-supply.html
- •Wikipedia — 2024-present global memory supply shortage (structural HBM crowd-out, timeline into 2027): https://en.wikipedia.org/wiki/2024%E2%80%93present_global_memory_supply_shortage