Inside Anthropic's $100 Billion Al Compute Commitment | CFO Krishna Rao
- Source
- youtube.com
- Author
- Invest Like the Best
- Date
Context
Anthropic's chief financial officer, Krishna Rao, describes buying computing capacity as planning for a wide range of outcomes. In an Invest Like the Best interview, Rao says too much compute can sink a company while too little leaves it unable to serve customers or stay at the frontier, so Anthropic models demand scenarios one to two years ahead and works backward. This is Rao's account, not an outside audit.
To make a wrong guess cheaper, Rao says Anthropic uses chips from Amazon (Trainium), Google (TPUs) and NVIDIA (GPUs) interchangeably for developing models, speeding up its own staff and serving customers, with a floor on model-development compute. Rao describes a Trainium deal for up to 5 gigawatts, a commitment over $100 billion, and a 5 gigawatt TPU deal with Google and Broadcom starting in 2027.
Rao also says each Opus release from 4 through 4.7 processes more efficiently, which helps both serving and training because reinforcement learning is largely in a with a reward function. No efficiency multiple is given.
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
- inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
- sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.
Checking sign-in…
Loading comments…




