Google still runs 7-8 year old TPUs at 100% utilization, defying bearish capex math
demian_ai · x · 2026-10-13
The author argues that Google is still running TPUs that are 7-8 years old at 100% utilization, undercutting bearish AI capex math that assumes chips lose most value in three years. Google depreciates TPUs over roughly six years and keeps finding work for them after that.
- Pod retirement isn't about age: a pod must be fully depreciated AND a newer one must extract much more output per watt in the same slot. Even then, swaps are messy — pulling a whole pod means thousands of chips across 150 racks, and new pods never quite fit the old footprint.
- With power as the scarcest resource, the real question is whether an old chip's watts would earn more inside a newer one.
- Watch Google's Oct 28 earnings for commentary on server useful lives and old fleet utilization.
The attached image shows the TPU evolution timeline from v1 (2015) to Trillium (2024).
More from Venture
- NVIDIA in talks to acquire Reflection AI, may repurpose GTM team for open-weight LLM push — thedealdirector · 2026-10-13
- Grok bot 'Grokbit' rakes in over $10K in fees — nima_owji · 2026-10-13
- Nvidia now takes 38% of top-5 hyperscaler IT capex, up from 11% in 2023, targeting 44% — Beth_Kindig · 2026-10-13
- China's $7B AI microdrama boom overtakes box office: how the addiction machine works — adariostrange · 2026-10-13
- Japan's Mujin raises $224M Series D as SK Square bets on robot-run factories — lukas_m_ziegler · 2026-10-13
- Warehouse robotics firm Ultra raises $50M Series A after packing 500K+ orders — lukas_m_ziegler · 2026-10-13