Morgan Stanley warns AI’s biggest bottleneck is now memory, not compute

The AI industry's next bottleneck is not about processing speed — it is about memory. Morgan Stanley ↘ 0.46% published a sweeping report this week arguing that the central constraint in artificial intelligence is migrating from a "compute wall" to a "memory wall," a shift the firm says will reshape the entire AI infrastructure investment landscape over the next half-decade.

The Widening Gap

The bank's analysis, led by Shawn Kim, Head of Morgan Stanley's Europe and Asia Technology Team, lays out the mismatch in stark terms. DDR5 single-channel bandwidth is expected to grow from 44.8 GB/s in 2024 to just 51.2 GB/s in 2026 — roughly a 14% increase over two years. In that same window, global AI inference token generation is projected to leap from approximately 10 trillion tokens per month to 3,200 trillion, a more than 320-fold surge. The result is a rapidly widening chasm between what processors can handle and what memory systems can deliver.moomoo

That imbalance is already showing up in costs. Storage-related components now account for as much as 73% of CPU server bill-of-materials expenses, and DRAM prices per gigabyte have climbed to their highest levels in nearly 30 years. According to TrendForce, conventional DRAM contract prices surged by more than 90% quarter-over-quarter in the first quarter of 2026, while Gartner estimates annual DRAM prices this year will rise by 125%.sokatec+3

A New Investment Frontier

Morgan Stanley forecasts that cloud storage spending will reach $418 billion by 2030, with memory's share of cloud providers' capital expenditures rising from 12% in 2023 to 40% by 2027. The total addressable market for novel memory technologies, including high-bandwidth memory (HBM), could reach $276 billion by 2030.macrostream+1

The firm identified six areas of innovation it expects to drive the next wave of AI infrastructure spending: advanced process nodes, memory architecture redesign, advanced packaging, peripheral interconnect chips such as CXL, processing-in-memory, and new materials. In a June podcast, Kim noted that memory prices have risen more than six-fold over the past year, describing the situation as "chipflation" — when memory chips stop getting cheaper and become harder to find.moomoo+1

Ripple Effects Beyond Data Centers

The supply crunch is already spilling into consumer electronics. Morgan Stanley projects that PC memory demand could face a 15% shortfall in 2027, equivalent to about 58 million units, while smartphones could see a 12% shortfall affecting roughly 134 million devices. Reuters reported in June that makers of devices from smartphones to PCs are being forced to choose between raising prices and accepting thinner margins.reuters+1

The firm's preferred investment plays include Micron ↘ 1.89%, Samsung, SK Hynix, and SanDisk Western Digital Corporation ↗ 0.93%, alongside semiconductor equipment makers such as ASML ↘ 0.43%. As Kim put it: "GPUs determine how fast AI runs, while memory determines how far AI can go."finance.yahoo+2