What happens when data centers must replace hundreds of billions in GPUs?
No-Papaya-9289 · reddit · 2026-09-26
A Reddit discussion on the long-term financial logic of AI infrastructure: companies have spent hundreds of billions building data centers around the latest GPUs, but how long will those chips last — two, three, five years — before requiring equally massive replacement? Unlike cloud storage drives, these are the most expensive components in the data center. Aside from Nvidia profiting, the poster worries AI companies may be burning money forever under the current approach.
More from Infra
- Running Qwen3.8 Locally on 6x3090s with exllamav3 Hits 80-120 tok/s — takoulseum · 2026-09-26
- Gemma 4 31B UD-Q8_K_XL Suddenly Hits "Exceeds Shared Memory" at 68k Context — Few_Professional6859 · 2026-09-26
- Meta Open-Sources Muse Glimmer: 30B Agentic Model That Fits a 24GB Consumer GPU — bibryam · 2026-09-26
- SGLang team launches $ARKIE memecoin that routes 2% trade fees into GPUs for open AI infra — BanghuaZ · 2026-09-26
- Google confirms Gemini 4 is in early post-training as it readies to send TPUs to space — 量子位 · 2026-09-26
- TSMC approves $60.7bn in capital appropriations in 91 days, betting big on advanced nodes — julsimon · 2026-09-26