HBM memory has already accounted for 63% of AI chip costs, with SK Hynix, Samsung, and Micron holding the pricing power over computing power.

Epoch AI’s latest research shows that the share of high-bandwidth memory (HBM) in AI chip component costs has risen from 52% in Q1 2024 to 63% in Q4 2025. Over the same period, absolute HBM spending surged from $12 billion to $32 billion, a 167% increase.
(Background: Korean stocks KOSPI surged 8% and triggered a trading halt! Samsung rose 7.4% intraday despite not striking; SK hynix jumped 11.2%)
(Additional context: Gavin Baker’s three contrarian bets: TSMC in a bid to stabilize the market, Trainium undervalued, and space computing power to be revealed within two years)

Epoch AI’s latest analysis tracks the component cost structures of four major chip designers—Nvidia, AMD, Google, and Amazon—covering the span from Q1 2024 to Q4 2025. The findings are clear: the AI compute cost bottleneck is shifting from chip design to memory supply.

What is HBM, and why is it so expensive?

High Bandwidth Memory (High Bandwidth Memory, HBM). The core concept is to vertically stack memory chips and directly package them next to the GPU.

AI model training and inference require GPUs to read and write massive amounts of data in extremely short timeframes. Traditional memory bandwidth is not enough—HBM is designed to solve this bottleneck. After multiple layers of DRAM chips are stacked, transfer speeds can reach several times those of traditional solutions, and latency is significantly reduced.

Simply put: HBM’s job is to stop the GPU from having to wait for memory, rather than making the GPU itself run faster.

The issue is that manufacturing this type of memory is extremely complex, and only three companies worldwide—SK hynix, Samsung, and Micron—have mass-production capability. After 2022, AI compute demand rose exponentially, and the three manufacturers’ bargaining power also increased substantially.

Epoch AI’s data clearly shows a shift in the cost focus. In the chip procurement mix of Nvidia, AMD, Google, and Amazon, the cost share of HBM increased from 52% in Q1 2024 to 63% in Q4 2025—an increase of 11 percentage points.

At the same time, the share of logic chips (the GPUs themselves) edged down from 14% to 13%, advanced packaging fell from 19% to 15%, and auxiliary components declined from 15% to 9%. When every other segment’s share is shrinking, only memory continues to expand.

How does cost pressure flow through to capital expenditure?

HBM price hikes are not just a supplier problem—they are cascading down the supply chain.

HBM makes up 63% of AI chip component costs, and AI chips are a major portion of tech giants’ capital expenditure. Therefore, any memory price increase directly shows up in financial reports. In its FY2026 capital expenditure plan, Microsoft separately lists a $2.5 billion component price increase factor—this is the portion that can clearly be attributed to the $25 billion increase. Meta likewise raised its 2026 capital expenditure range by $10 billion, citing rising component costs.

In absolute terms, the transmission magnitude is equally staggering. HBM spending jumped from $12 billion in 2024 to $32 billion in 2025—a 167% increase in a single year. Total AI chip component spending rose from $22 billion to $52 billion, up 136%. The entire supply-chain scale doubled within two years, making the structural problem of memory shortages even harder to resolve in the short term.

Epoch AI expects that in 2026, HBM’s share of AI chip component costs will continue to rise. In other words, higher capital expenditure does not necessarily mean computing power increases proportionally. A significant portion of it is used to absorb the costs of higher memory prices.