Memory Eats the AI Chip: HBM Hits 63% of Component Costs

The News

High-bandwidth memory has quietly become the heart of the AI economy. Fresh data from Epoch AI researcher Venkat Somala shows HBM now accounts for 63% of the bill of materials for a typical AI accelerator in Q4 2025, up from 52% in Q1 2024. Logic dies, the silicon that actually performs the maths, take only 13%. Advanced packaging accounts for 15% and auxiliary components the remaining 10%.

The shift sits on top of huge growth in absolute spending. Total component outlay across the four chip designers Epoch tracks (Nvidia, AMD, Google and Amazon) climbed from roughly $22 billion in 2024 to about $52 billion in 2025. HBM alone went from $12 billion to $32 billion year on year. That $20 billion swing in memory accounts for most of the overall growth.

Hyperscalers are already baking the inflation into next year's plans. Microsoft has set aside $25 billion in its FY2026 capex specifically for component price increases, and Meta has lifted the top of its 2026 range by $10 billion.

Why It Matters

For two decades chip economics were dictated by logic. Intel and TSMC fought over nanometres and the wafer was the prize. The Epoch numbers mark a quiet inversion: the most expensive thing on an AI board is no longer the transistor, it is the SDRAM stacked next to it. Three vendors, SK Hynix, Samsung and Micron, sit on a near-monopoly over HBM3E and HBM4, giving them rare pricing power.

The last time a single component category swallowed this much of a system's cost was during the 2021 GPU shortage, when DRAM contract prices doubled in nine months. The current squeeze is structurally tighter because HBM is bespoke, low-yield and bonded into the GPU package, not a spot-market part. Hyperscaler operating margin will become more sensitive to a few thousand HBM wafers from a single Korean factory than to almost any other variable.

Indian Angle

For Indian policymakers the data lands well. The Tata-PSMC fab at Dholera and Micron's $2.75 billion assembly, test and packaging plant at Sanand have both been pitched as central to semiconductor sovereignty, but their commercial logic always looked softer than the press releases suggested. If memory is now where the gross margin lives, an OSAT site that can package HBM dies is suddenly more strategic than a logic fab three nodes behind the frontier.

For Indian AI startups the news is less welcome. Sarvam, Krutrim and the rest of the foundation-model wave rent compute in dollars on Nvidia capacity. As HBM cost share rises, the rental price per H100 or B200 hour will follow, and the rupee's recent slide widens the gap further. Expect founders to lean harder on smaller models, distillation and inference-time scaling rather than training larger checkpoints.

TCS, Infosys and Wipro will feel the second-order effect. A Microsoft capex line being eaten by Korean memory leaves less room for systems-integration spend, a ratio worth watching when quarterly guidance lands in July.

FAQ

How accurate is the 63% figure?

Epoch puts the range at 60% to 67% for Q4 2025. The precise share depends on product mix between Nvidia's Blackwell and AMD's MI series, reconstructed from a bill-of-materials analysis weighted by production volume.

Why has HBM become so expensive?

HBM stacks dies vertically using through-silicon vias and is bonded directly to the GPU package. The process is low-yield and only SK Hynix, Samsung and Micron operate at commercial scale. Demand from every hyperscaler chasing the same wafers lifted contract prices through 2025.

Could India produce HBM domestically?

Not in this decade. HBM4 needs leading-edge DRAM nodes that no Indian facility is currently licensed for. The realistic near-term role is packaging and testing at Micron Sanand or future OSAT sites, which capture a smaller but meaningful slice of the value chain.

Where can I read the original analysis?

Epoch AI's full data note is on its Data Insights page, authored by Venkat Somala and published on 21 May 2026.

This story was reported by Epoch AI. Read the full original coverage at Epoch AI.