The AI boom is no longer hitting only GPUs, power supply and data centres. It is now reshaping ordinary memory markets as well. TrendForce expects DRAM and NAND to account for 68% of major cloud providers' capital expenditure in 2027. That is a remarkable share for components that used to be treated as relatively interchangeable parts of a server.
The effect reaches mass-market devices. PC makers, smartphone vendors and SSD buyers are not directly competing for the exact same chips used in an AI cluster. They are, however, competing for manufacturing capacity, investment and priority from the same memory suppliers.
The bottleneck goes deeper than HBM
AI memory is usually associated with HBM sitting next to Nvidia or AMD accelerators. The current pricing pressure reaches much further. Large cloud providers also consume huge quantities of server DRAM and enterprise SSDs, giving suppliers a strong incentive to prioritise products with higher margins and predictable long-term demand.
TrendForce already described this shift in March, with capacity moving toward HBM, server DRAM and enterprise SSDs. It forecast conventional DRAM contract prices to rise 58–63% quarter over quarter in the second quarter, while NAND was expected to increase 70–75%. Growth moderated in the third quarter but remained clearly positive.
A 270% server-DRAM increase shows the scale of the shift
The change is especially visible among cloud providers. TrendForce expects server DRAM contract prices to rise by about 270% in 2026. At the same time, total capital expenditure by major cloud service providers is projected to jump 98% in 2026 and another 50% in 2027.
That does not mean every retail RAM module becomes 270% more expensive. Contracts, memory types, capacities and customer segments differ substantially. This is where the broad “RAMageddon” headline becomes too blunt: the pricing shock is real, but not every memory product follows the same curve.
Why consumers still end up paying
For PC and smartphone vendors, direct demand for their specific memory chips is only part of the equation. What matters is which products Samsung, SK hynix, Micron and other suppliers choose to prioritise on constrained production lines. If server products carry higher margins and cloud customers sign long-term supply agreements, consumer products lose allocation power.
For the third quarter, TrendForce still forecasts another 13–18% increase for conventional DRAM and 10–15% for NAND despite weaker consumer demand. That creates an unusual market: softer PC and smartphone demand does not automatically push prices down because supply is simultaneously being redirected toward data centres.
What the fivefold headline does not tell you
The Financial Times reports that some DRAM pricing measures have risen roughly fivefold within a year. As an indicator of the scale of the disruption, that matters. As a blanket claim about all “RAM prices”, it would be too imprecise.
Spot markets, OEM contracts, server RDIMMs, LPDDR, GDDR and HBM are different markets. Using one price series to claim that every notebook vendor now pays five times as much for memory would overstate the case. The defensible conclusion is narrower: several memory segments are seeing exceptional increases, while AI infrastructure is structurally changing capacity allocation.
Capacity planning matters again for administrators
For enterprises, the trend has a practical consequence. RAM and SSD pricing should again be treated as a separate procurement risk in server, virtualisation and private-cloud projects. Organisations planning major 2027 refreshes should not assume the memory prices of earlier upgrade cycles.
Wasteful VM reservations also become more expensive. When memory was cheap, generous overprovisioning was convenient. At materially higher prices, measuring real utilisation, storage tiers and retention becomes worthwhile again — without cutting resilience on critical systems.
Pandorex View
“AI is eating all the RAM” is catchy but technically too simple. The stronger mechanism is economic: cloud providers accept high prices, secure supply through long-term agreements and pull manufacturers toward higher-margin server products. That makes memory scarcer and more expensive even in systems that never run an AI model.
That matters more than any single spectacular pricing figure. If DRAM and NAND really consume such a large share of cloud CapEx, memory is moving from supporting component to strategic bottleneck in AI infrastructure.
Relevance: 9/10 · Hardware impact: 9/10 · Administrator relevance: 8/10 · Confidence in trend direction: 9/10