Nvidia Tells Customers AI Servers Getting ~17% Pricier as Memory Costs Surge
rohanpaul_ai · x · 2026-08-23
Nvidia has informed major customers that servers powered by its AI chips are getting 17% pricier in many cases, driven by memory inflation in Grace Blackwell and Vera Rubin systems.
- Memory is the pressure point: Vera Rubin NVL72 packs 20.7TB of HBM4 and 54TB of LPDDR5X per rack; TrendForce expects DRAM supply to stay tight through 2027 as AI servers pull production toward HBM and server memory.
- The increase could add at least $5B to 1GW builds, before power, cooling, networking, buildings, and financing costs.
- Cloud providers can absorb some of it, but pass-through would raise the cost of renting Nvidia compute for training and inference.
- Nvidia's land-and-power investments (Cloverleaf, SB Energy) look more consequential: with more capital tied up per rack, delays idle more money, giving Nvidia a financial reason to secure site and power readiness itself.
Related event: Nvidia Tells Top Customers AI Server Prices to Rise Over 15%(15 posts)→
More from Venture
- YC-backed Overshoot launches inference optimization, claims $5M saved for AI teams — ycombinator · 2026-10-16
- Vercel launches Token Pass: $30/mo for $80 of GLM 5.3 and Kimi K3 credits — evilrabbit_ · 2026-10-16
- Polymarket gives Meta just 3% to lead AI by year-end; Anthropic at 44% — Polymarket · 2026-10-16
- Ferryman: solo-built cross-posting SaaS charges $30-$100/mo, syncs via Claude and Codex — KevinNaughtonJr · 2026-10-16
- 6 open models, 80 M&A memos, double-blind AI judges: Step 5 Preview claims a new Pareto frontier — Aizkmusic · 2026-10-16
- Kevin Kwok: Vercel's agent play is 'manufacturing your own customers,' TSMC-style — kevinakwok · 2026-10-16