Mixture of Experts Anthropic’s TPU move and NVIDIA’s Starcloud
15 snips
Oct 31, 2025 
Gabe GoodHart, an AI architect specializing in model deployment, Chris Hay, a distinguished engineer known for his skeptical take on AI trends, and Kate Soule, director at Granite focusing on AI governance, dive into Anthropic's TPU deal with Google Cloud. They discuss the implications of NVIDIA's plan for data centers in outer space and ponder the feasibility of insuring superintelligent AI. They also evaluate OpenAI's approach to sensitive conversations, weighing ethics and effectiveness in AI's emotional engagements.
AI Snips
Chapters
Transcript
Episode notes
Hybrid Compute Strategy Wins
- Anthropic will expand Google TPU usage to scale inference while keeping NVIDIA for novel training work.
- Providers balance cheaper, broad inference hardware with CUDA-optimized GPUs for cutting-edge models.
Software, Not Just Chips, Locks In GPUs
- CUDA and tuned kernels keep NVIDIA dominant for novel architectures and rapid innovation.
- Alternate hardware catches up on common tensor ops, enabling mass inference portability.
Insurance Could Create A Safety Flywheel
- Underwriting Superintelligence frames AI safety as an insurance-driven incentive flywheel to raise standards.
- Panelists note the flywheel needs consumer money flow and verifiable standards to actually influence behaviour.