Claim: Gemini 4 Trained Entirely on Google's Own TPUs, No Nvidia Involved
gaganghotra_ · x · 2026-11-15
A widely shared analysis claims Google is the only tech giant training top-tier frontier models entirely without Nvidia: from Gemini 1 to Gemini 4, core model training has run on in-house TPU clusters — primarily TPU v6e/Trillium plus newly deployed TPU 8t/8i — using the JAX/XLA stack and ICI interconnects to build Pods with tens of thousands of chips. The Gemini 4 Argon model card reportedly states distributed training is built entirely on TPU hardware. Details are unverified beyond the model card; check the official card for specifics.
More from Infra
- JP Morgan sees AI accelerator shipments doubling to 31.2M units by 2028, custom silicon leading — Beth_Kindig · 2026-11-15
- Home Lab: 8x RTX 3090 Rig Runs 10 Parallel Agents at 370 tok/s — dazzou5ouh · 2026-11-15
- Alchemy adds Rivet, Effect Cluster and 8 sandbox platforms with swap-one-layer portability — samgoodwin89 · 2026-11-15
- Chamath explains why CoreWeave credit default swap spreads are jumping — rohanpaul_ai · 2026-11-15
- Solar-powered local AI: old MacBook + Meshtastic mesh keeps inference running off-grid — natesiggard · 2026-11-15
- Do developers want a simpler middle ground between AWS AgentCore and Cloudflare Agents? — line9779 · 2026-11-15