AI, greed cause massive spike in memory prices

News
Nov 5, 20253 mins

DRAM prices could increase over 23% in coming year, report says.

Broken piggy bank with coins money spilled against a white background, close-up
Credit: Gecko Studio / Shutterstock

DRAM prices have skyrocketed thanks to the explosion in construction of AI data centers, which can require massive amounts of memory. Prices are high, getting higher, and won’t come down until well into next year, according to predictions.

TrendForce, a market research firm that specializes in the memory market, says that server DRAM contract prices are strengthening in the fourth quarter of this year, driven by ongoing data center expansion among global cloud service providers. This momentum is lifting overall DRAM pricing and affecting both DDR4 and DDR5 memory. (See also: Storage constraints add to AI data center bottleneck)

Looking forward to 2026, global server shipments are projected to increase by approximately 4% each year. At the same time, cloud service providers are swiftly transitioning to high-performance computing platforms to accommodate large-scale AI models, which is significantly boosting the memory requirements per server. As a result, overall DRAM bit demand is expected to surpass previous forecasts, prolonging the ongoing supply shortage.

Although final contract pricing for the quarter is still being negotiated, suppliers are showing a greater willingness to raise quotes as cloud service providers increase order volumes. TrendForce has revised its 4Q25 outlook for conventional DRAM pricing upward, from an earlier forecast of 8-13% growth to 18-23%, with a strong likelihood of further upward revision.

And that’s a modest estimation. Jim Handy, president of Objective Analysis, follows the memory market and says the rate of increase in prices “depends on how good of a customer you are. Solid volume purchasers have seen more moderate price increases, while spot market prices for DRAM chips have increased by anywhere from 3 to 11 times what they sold for in March.”

The problem is not just in purchasing of DRAM. The few memory makers that make up the market have shifted their limited manufacturing from regular DRAM to HBM memory used in GPU accelerators for AI machines, because HBM is much more profitable than standard DDR memory.

TrendForce says that as of 2Q25, HBM3e still commanded a price premium more than four times that of DDR5, so it’s hard to fault the memory manufacturers for wanting to make a buck. However, as DDR5 prices continue to rise, the gap between the two is projected to narrow significantly in 2026, and DDR5 profitability could surpass that of HBM3e starting in the first quarter of next year, TrendForce stated.

The seven hyperscalers Objective Analysis tracks used to spend an average of 12-15% of their revenues on “PPE,” which stands for “property, plant, and equipment.” Over the past six quarters, that share has zoomed to more than 25%, which means that they are spending on equipment without getting any return, basically betting that the market for AI will grow to justify their investment. 

“A company can only do that for so long before it exhausts its bank account, but it’s more rational to anticipate that investors will pressure management into changes before that point is reached. It comes down to a question of the investors’ appetite for risk,” said Handy.