Rising HBM Prices Become a Major Cost for Hyperscale Data Centers
zephyr_z9 · x · 2026-07-07
Analysts point out that the "memory tax" in AI training and inference is being reframed as an "Nvidia tax." In reality, however, about 35%-40% of capital expenditure in hyperscale data centers goes toward memory (HBM).
Recently, multiple hyperscalers have revised their capital expenditures upward, largely driven by rising memory prices rather than just expanding compute capacity. This indicates that the supply of memory, such as HBM, has become a primary cost driver for AI infrastructure.
More from Infra
- AWS tutorial: fleet-wide edge CV model lifecycle with SageMaker SDK v3, ~$0.07 per run — AWS ML Blog · 2026-10-27
- Qwen3.8-27B hits 40 tok/s on Jetson Thor with INT4 + MTP, accuracy beats FP16 — No-Dragonfly6246 · 2026-10-27
- Celesto launches CI runners nearly 90% cheaper than GitHub Actions, with agent sandboxes — aniketmaurya · 2026-10-27
- Supabase founders on rebuilding the backend for a billion AI-built apps — Latent Space · 2026-10-27
- AW-0 runs at 230ms on a single RTX 5090, 2.4x faster than autoregressive code planning — trevordarrell · 2026-10-27
- Site tracks US city council topics: data centers rank only behind pickleball courts — zakkohane · 2026-10-27