Most companies allocate their most expensive resources toward the thing that makes money right now. Anthropic is doing the opposite.
In an appearance on the “Invest Like the Best” podcast, Anthropic CFO Krishna Rao laid out the company’s compute allocation strategy in unusual detail. The headline: Anthropic’s internal research team gets the biggest slice of the company’s computing power, even ahead of customer-facing workloads. The logic is counterintuitive but deliberate. Better models today mean more efficient tokens tomorrow, which means more capacity for paying customers down the road.
The math behind giving researchers first dibs
Rao described a system where Anthropic holds daily internal meetings to assess what he calls “return on compute,” essentially a real-time audit of whether each unit of processing power is being deployed where it generates the most value. Those meetings allow rapid reallocation of resources across research projects, customer training jobs, and inference workloads.
Anthropic maintains a strict minimum allocation for research that doesn’t bend to commercial pressure. Rao suggested the company is knowingly leaving billions of dollars in immediate revenue on the table to protect that floor.
