Anthropic CFO: compute strategy supports $30B run rate

Anthropic’s chief financial officer, Krishna Rao, described the company’s compute strategy-strategic procurement, use of three chip platforms and close supplier cooperation-as supporting its growth toward a $30 billion annualized revenue run rate while improving model development speed and cost efficiency.

Rao joined Anthropic in May 2024. He previously served as chief financial officer at Fanatics Commerce and helped secure a $1 billion lifeline for Airbnb during the 2020 pandemic.

Rao warned that mismanaging compute can harm a technology company: “If you buy too much compute you go out of business; if you buy too little compute you can’t serve your customers.” He said Anthropic treats compute procurement as a strategic financial decision and keeps a minimum level of compute for model development while shifting capacity between training and serving as needs change.

Anthropic runs workloads on three platforms: Amazon’s Trainium chips, Google’s CPUs and Nvidia’s GPUs. The company built the ability to move workloads across different hardware and now uses that flexibility to optimize costs and performance. Rao added that partnerships with chip manufacturers let Anthropic extract more value from each dollar spent on compute: “A dollar of compute inside our organization goes further than I think it does anywhere else.”

Engineers can reallocate capacity on short notice, moving compute between development and inference to speed experiments and reduce training costs. Rao noted the firm will not let development compute fall below a set threshold even if that makes customer serving more difficult in the short term.

Rao described model intelligence as multidimensional and said new model releases have increased capability and efficiency, which has created new enterprise use cases. He said the company is seeing faster progress as models scale and training techniques improve.

On financials, Rao said Anthropic began the year with about a $9 billion run rate and is scaling toward a much higher annualized figure as models and products reach more customers. He described the company’s operational approach as balancing capability gains with cost control through dynamic compute allocation.