Memory for AI accelerators is sharply rising in price amid NVIDIA Vera Rubin chips

Memory for AI Accelerators Is Sharply Rising Amid NVIDIA Vera Rubin Chips

Memory costs for AI accelerators keep climbing — Bernstein links the uptick to preparations around the Vera Rubin supply. The HBM market, according to their note, has moved into a fresh phase of price pressure; not exactly a quiet shift.

Bernstein est. HBM4 at about $53/GB by 2027 (GB = gigabyte). Meanwhile, on system pricing: analysts expect NVIDIA to pass those memory hikes on to buyers — i.e., higher component costs become higher rack prices.

For context, Morgan Stanley had earlier put a single NVL72 rack at roughly $7.8M, with memory making up about $2M of that total. Bernstein calls that conservative and allows for the NVL72 rack price to rise to about $9.1M (USD). That number feels big — and for datacenter budgets, it will be noticeable.