← Home
Stratechery · July 14, 2025 · 52m

Google's Gemini and the TPU Advantage

Ben Thompson

Thompson analyzes Google's Gemini models and argues that Google's custom TPU chips provide a structural cost advantage that may matter more than model quality in the long run — AI is becoming an infrastructure competition.

This summary was generated from show notes and public descriptions, not from a full transcript review. Details may contain inaccuracies.

Canon

•
The false self of AI benchmarks — models that score similarly on benchmarks may have very different real-world performance, but the benchmark scores are what gets reported
Thompson warns that AI benchmarks present a false picture of model equivalence. Two models that score 90% on MMLU may have very different performance on the specific tasks users care about. Benchmarks are the false self of model capability.
ClaudeChatGPTGemini

Highlights

•
AI competition is shifting from model quality to inference economics — the company that can serve AI cheaply wins, regardless of who has the best model
Thompson argues the AI competition is evolving: early competition was about model quality (who has the best model?). The next phase is about inference economics (who can serve the model cheapest?). Google's custom TPUs may provide a decisive cost advantage.
ClaudeChatGPTGemini
Was this useful?