OpenAI started using Google's TPUs through Google Cloud to reduce costs and diversify from reliance on NVIDIA
- OpenAI compute bill is projected to be $14 billion in 2025
- Cost is 50% training / 50% inference
- NVIDIA still dominates training, but competition for inference is hitting up
- Google, AMD, and Amazon have solutions that can infer more tokens per dollar than NVIDIA
- The era of NVIDIA's high margins might be coming to an end
News: Google convinced OpenAI to pay for tensor processing units and shift away from Nvidia.
OpenAI has ~billions and billions~ of reasons to seek cheaper AI chip alternatives