Krishna Rao - Anthropic's CFO on Compute, Scaling to $30B ARR, and the Returns to Frontier Intelligence - [Invest Like the Best, EP.472]
Krishna Rao, CFO of Anthropic, discusses the company's compute strategy, explosive revenue growth from $9B to $30B ARR in a single quarter, and the thesis that returns to frontier AI intelligence are exceptionally high in enterprise. He explains how Anthropic uses three chip platforms fungibly, navigates a 'cone of uncertainty' in forecasting, and why the company's culture of intellectual humility and collaboration has been a key competitive advantage.
Summary
Krishna Rao, CFO of Anthropic, sits down with Patrick O'Shaughnessy to discuss running the financial operations of one of the fastest-growing companies in history. The conversation centers on compute as the foundational resource of the business — Rao describes it as 'the canvas on which everything else gets built.' Anthropic uses three chip platforms (Amazon Trainium, Google TPUs, and NVIDIA GPUs) fungibly across three workloads: model training/development, internal use by employees, and serving customer inference demand. This flexibility, built over multiple years, is considered one of Anthropic's key competitive advantages and required building custom compilers and orchestration layers down to the chip level.
Rao introduces the concept of the 'cone of uncertainty' to describe how exponential business growth makes forecasting extremely difficult. Small differences in weekly or monthly growth rates compound into wildly different outcomes, requiring Anthropic to think in scenario ranges rather than point estimates and to build flexibility into compute procurement contracts. He notes that the company has signed deals totaling over $150 billion in compute commitments with Google/Broadcom (5 gigawatts of TPUs starting 2027) and Amazon (5 gigawatts of Trainium), while also opportunistically sourcing near-term compute such as the newly announced partnership with SpaceX's Colossus facility in Memphis.
A core thesis of the conversation is that 'returns to frontier intelligence are extremely high, especially in enterprise.' Rao argues that each new model generation doesn't just improve raw intelligence scores but unlocks new use cases and longer-horizon agentic tasks, driving a Jevons paradox effect where lower prices or better capabilities lead to dramatically higher consumption. This is evidenced by Anthropic growing from approximately $9B to $30B in annualized run rate revenue in a single quarter. The company serves nine of the Fortune 10, has a net dollar retention rate over 500%, and recently lowered Opus pricing, which caused consumption to increase far more than proportionally.
Rao discusses how Anthropic internally uses its own models extensively — over 90% of code is written by Claude/Claude Code, and the finance team has built over 70 Claude-powered skills including automated monthly financial reviews and statutory financial statement generation. He frames this internal use as essential: foregoing potential revenue from that compute is justified because it accelerates model development and compound productivity gains.
On the platform vs. application question, Rao says Anthropic is primarily a horizontal platform company (analogous to early AWS), but will build vertical applications like Claude Code where it has unique insight into model trajectories or can demonstrate value for the ecosystem. He describes the vision as enabling a 'virtual collaborator' that can use company-specific tools, maintain memory, and work on long-horizon tasks — targeting the estimated $40 trillion annual knowledge work market.
Regarding the competitive and regulatory environment, Rao notes that Mythos was released in a phased way due to its exceptional cybersecurity capabilities, which Anthropic views as a template for responsible release of highly capable models. He discusses Anthropic's collaborative relationship with hyperscalers as multifaceted partnerships involving chip roadmap influence, capacity planning, and distribution — not just procurement. On talent retention, despite large competing offers from companies like Meta, Anthropic reportedly lost only two technical employees, which Rao attributes to the company's distinctive culture of intellectual humility, radical transparency, collaborative decision-making, and mission alignment around safe AI development.
The interview closes with Rao identifying biotechnology and drug discovery as the area he is most excited about, envisioning AI dramatically accelerating the lab throughput for molecular research and potentially enabling cures for diseases within patients' lifetimes.
Key Insights
- Rao argues that compute is so central to Anthropic that buying too much causes bankruptcy while buying too little prevents staying at the frontier — there is no safe middle ground, making procurement the most consequential decision in the company.
- Anthropic uses three different chip platforms (Trainium, TPUs, GPUs) fungibly across training, internal use, and customer inference — a capability Rao claims took years to build and makes them the most efficient compute users among frontier labs.
- Rao claims that each new model generation produces simultaneous improvements in both capability and inference efficiency, unlike the car analogy where better performance means worse fuel economy — meaning customers get more intelligence at lower cost with each release.
- Anthropic grew from approximately $9B to $30B in annualized run rate revenue in a single quarter (Q1), which Rao attributes primarily to model intelligence leaps unlocking new enterprise use cases rather than sales team expansion.
- Rao states that Anthropic's net dollar retention rate exceeds 500% annualized, indicating that existing enterprise customers dramatically increase their spend as they find more use cases — driven by the Jevons paradox effect of better, cheaper models.
- Rao reveals that over 90% of Anthropic's internal code is written by Claude Code, and that a significant portion of Claude Code's own code is written by Claude Code — describing this as the practical reality of recursive model-assisted development.
- Anthropic has committed to over $150 billion in future compute purchases (5 gigawatts each from Google/Broadcom and Amazon), with Rao framing these as long-horizon investments managed alongside near-term opportunistic compute like the SpaceX Colossus partnership.
- Rao argues that Anthropic's investment in AI safety research (interpretability and alignment science) was motivated by mission but produced an unexpected business benefit: enterprise customers entrusting sensitive workloads prefer a company with a demonstrable safety track record.
- When Anthropic lowered prices on the Opus model family, consumption increased far more than expected — a Jevons paradox effect Rao says validates their thesis that accessible pricing to unlock enterprise TAM is more valuable than higher per-token margins.
- Rao describes Anthropic's culture as one where a candidate can excel on every technical dimension but be rejected for failing the culture interview — citing that when Meta made large competing offers to AI talent, Anthropic lost only two people while other labs lost dozens.
- Rao frames Anthropic's strategy as primarily horizontal platform (building the intelligence layer and tools/SDKs for others to build on), with selective vertical applications like Claude Code built only where Anthropic has unique foresight into model trajectory or ecosystem demonstration value.
- Rao identifies the biggest investor misconception about Anthropic as treating compute as a separable variable cost analogous to traditional software R&D, when in reality compute is a single fungible resource simultaneously driving near-term revenue, long-term model development, and internal productivity — making ROI on the full compute envelope the correct measuring stick.
Topics
Full transcript available for MurmurCast members
Sign Up to Access