Soaring memory prices are set to drive AI infrastructure capex even higher, with industry analyst firm TrendForce projecting a 98% year-on-year increase in 2026 and another 50% in 2027.

Memory lead times from some of the biggest providers are already stretched as far out as 2028. But spending on such hardware in the coming years will continue to soar.

TrendForce suggests combined dynamic random-access memory (DRAM) and non-volatile (NAND) flash memory spending will go from 47% of cloud service provider total capex in 2026, to 68% in 2027.

Component shortages are already having a knock-on effect, raising prices on everything from server lines to cloud service offerings. Nvidia has raised some product price lines in recent weeks, and Everpure admitted earlier this year that some of its services had risen around 70%. Meanwhile, Extreme Networks’ CEO said price increases showed the “price inelasticity of networking.”

TrendForce noted that contract prices for server DRAM in the second half of 2025 alone jumped 64%, a number that could rise as high as 270% by the end of 2026. The same is noted for flash, with enterprise solid state drives (SSDs) rising by around 35% during the second half of 2025, and projected to surge by a cumulative 235% in 2026.

While some long-term agreements signed beginning with the second quarter of 2026 having included price ceilings aimed at curbing further increases, TrendForce warns prices, specifically for high-bandwidth memory products, “could still rise by 70–140% in 2027.”

The report suggests operators may need to raise capex budgets further to meet price increases or opt for more “aggressive” optimizations of their existing AI memory stacks while still meeting chip and server shipment targets.

“Such adjustments could take several forms, including, but not limited to, adjusting RDIMM (registered dual in-line memory module) configurations and the amount of HBM (high-bandwidth memory) integrated into future AI chips, as well as exploring AI ASICs (application-specific integrated circuits) with the model architecture hardwired into the silicon,” the report reads.

Vendors are already getting innovative to get the memory crunch, such as through Nvidia’s “scale-in” networking update, which adds acceleration for north-south traffic, offloading heavier security and storage tasks from the main server processor to boost compute performance.