TrendForce: Memory to Account for 68% of Cloud Capex by 2027
Nashnova编辑部
TrendForce projects global cloud capex to surge 98% year-on-year in 2026, with memory's share rising from 47% to 68% by 2027 — driven by consecutive triple-digit contract-price jumps in DRAM and NAND.
Where is the cloud money going?
Major cloud providers' capex is set to rise 98% YoY in 2026, with growth still at 50% in 2027.
Memory (DRAM + NAND flash) will climb from 47% of total capex in 2026 to 68% in 2027.
This means → for every $100 in cloud capex, nearly $70 will flow directly to memory chips — making storage the single largest line item in this AI expansion cycle.
How much are memory prices actually rising?
Server DRAM contract prices: up a cumulative 64% in H2 2025, with another ~270% increase forecast for 2026.
Enterprise SSDs (NAND flash): up ~35% in H2 2025, with a cumulative ~235% rise projected for 2026.
HBM — high-bandwidth memory, a stacked chip designed for AI accelerators — may rise another 70%–140% in 2027.
In plain terms = this is not modest repricing. Memory prices are doubling year after year, and that is the fundamental reason its capex share is surging.
What is happening on the supply side?
TrendForce estimates HBM and RDIMM — a server-grade memory module — will together account for 51% of DRAM bit supply in 2026.
This means → memory makers are prioritizing limited capacity for servers, squeezing supply to consumer segments further.
By 2027, as process migration advances and new fabs ramp in H2, combined server-DRAM and HBM bit supply is expected to grow 27%.
What knock-on effects will the price surge create?
First effect: chip prices follow. Rising HBM costs give Nvidia and other AI-chip vendors stronger justification to raise prices. Cloud providers then need even more capex to maintain planned procurement volumes.
Second effect: architecture redesign. Cost pressure is pushing cloud providers to rethink AI-system memory architecture — options include adjusting RDIMM configurations, cutting HBM capacity in future AI chips, and exploring ASIC designs that bake models directly into silicon.
In plain terms = memory is too expensive, so cloud firms either spend more or engineer ways to do the same work with less memory. Both paths reshape AI hardware economics.
How long can this pricing cycle last?
TrendForce notes that some long-term agreements signed from Q2 2026 include price caps, which may limit further increases.
Overall, however, memory contract prices are expected to stay elevated through 2027.
This reflects an unresolved debate over this "AI memory super-cycle" — whether 2027 prices hold as forecast will be the key test of the cycle's durability.
Content is for reference only, not financial advice.