Krishna Rao serves as CFO of Anthropic, one of the leading AI companies developing frontier language models. He joined the company two years ago when it had $250 million in run rate revenue and has overseen its explosive growth to over $30 billion in annual run rate.

The conversation centers on Anthropic's approach to compute procurement and allocation, which Krishna describes as 'the canvas on which everything else gets built.' The company uses compute fungibly across three different chip platforms for model development, internal acceleration, and customer serving.

Krishna explains Anthropic's 'cone of uncertainty' planning framework for navigating exponential growth scenarios, the returns to staying at the frontier of AI capabilities, and how the company balances platform versus application development. He also discusses the cultural tenets that have enabled Anthropic to retain top talent and maintain collaborative decision-making as it scales.

The Cone of Uncertainty: Planning for Exponential Growth

Anthropic uses a 'cone of uncertainty' framework because small changes in weekly growth rates compound into vastly different outcomes when growing exponentially.

"If you buy too much compute, you go out of business. If you buy too little compute, you can't serve your customer and you're not at the frontier" - Krishna on the high-stakes nature of compute procurement decisions.

The company models scenarios across 1-2 year periods, working backwards from different points in the cone to ensure they can stay at the frontier while serving customers and accelerating internal development.

Fungible Compute Across Three Chip Platforms

Anthropic uses Amazon's Tranium chips, Google's TPUs, and NVIDIA's GPUs fungibly, switching workloads between inference, model development, and internal use throughout the day.

"We run workloads on one day in the morning on a chip for inference, and in the afternoon or evening, we use it for model development" - Krishna on dynamic compute allocation.

The company invested years building orchestration layers and compilers to achieve this flexibility, working closely with chip teams like Amazon's Annapurna Labs to influence roadmaps.

Krishna spends 30-40% of his time on compute decisions, with daily meetings to allocate resources between model development, internal use, and customer serving.

Revenue Explosion: $9B to $30B Run Rate in One Quarter

Anthropic's revenue grew from $9 billion to $30 billion annualized run rate in Q1 2025, driven by frontier model capabilities unlocking new enterprise use cases.

"The returns to frontier intelligence are extremely high, especially in enterprise" - Krishna on why staying at the cutting edge drives exponential growth.

Each new model generation creates step-function improvements in capability and efficiency, with customers immediately switching to the latest models for better performance.

Net dollar retention exceeds 500% annually, with nine of the Fortune 10 companies as customers making "double-digit million-dollar commits."

Recursive Self-Improvement and Scaling Laws

Over 90% of Anthropic's code is now written by Claude, with Claude Code writing much of its own code in a recursive improvement loop.

"The models themselves are helping us to build that next generation of models" - Krishna on how AI accelerates AI development beyond just human researchers.

Scaling laws continue to hold across pre-training, post-training, and reasoning, with no signs of slowdown in model capability improvements.

The company released 30 different product and feature updates in January 2025 alone, with pace accelerating due to model-assisted development.

Internal AI Transformation: Finance Team as Power Users

Anthropic's finance team produces statutory financial statements using Claude, with over 70 finance-specific skills in their internal library.

"What used to take hours to produce a weekly report for what's driving revenue or compute utilization now comes down to 30 minutes" - Krishna on productivity gains.

The head of tax is the team's top token user, focusing on automating tax policy engines and large workload portions rather than just junior employees adopting AI.

Monthly financial reviews are now 90-95% automated, allowing teams to focus on strategic implications rather than data reconciliation.

Massive Capital Commitments: $125B in Compute Deals

Anthropic has raised $75 billion since Krishna joined, with another $50 billion committed through Amazon and Google deals extending to 2027 and beyond.

The company signed deals for up to 5 gigawatts each with Google/Broadcom for TPUs and Amazon for Tranium chips, representing over $100 billion in commitments.

"We announced a partnership with SpaceX for their Colossus facility in Memphis" - Krishna on creative approaches to securing near-term compute capacity.

Compute procurement follows a "layer cake" approach with different capabilities starting at different times, optimized for price-performance across multiple generations.

Platform Strategy and Application Development Balance

"Most of what we're building is a platform" with applications like Claude Code built to demonstrate capabilities and create customer value ahead of model readiness.

Anthropic builds vertically where they have unique perspective or can demonstrate platform value, while partnering collaboratively in areas like security and financial services.

The company is "the only model that's on all three clouds" and "the only Large Language Lab using all three chip platforms," enabling broad distribution.

"We think there's so much value that's going to accrue in these areas that our customers can win and we can win as well" - Krishna on the collaborative approach to market development.

Culture and Talent Retention in Hypergrowth

All seven co-founders remain at the company along with the vast majority of the first 20-30 employees, despite aggressive external recruiting.

"When Meta and others were out with these huge packages for technical talent, we lost two people and other labs lost dozens" - Krishna on culture-driven retention.

The company conducts rigorous culture interviews that can disqualify otherwise excellent candidates, prioritizing collaboration over individual achievement.

Dario Amodei holds transparent all-hands meetings every two weeks with open Q&A, fostering intellectual honesty and alignment across the organization.

Future Vision: Virtual Collaborators and Healthcare Breakthroughs

The frontier goal is creating "virtual collaborators" with organizational context, tool access, memory, and ability to work on long-horizon tasks across the $40 trillion knowledge work economy.

"Cowork is growing faster than Cloud Code was if you index them to the same point in time" - Krishna on expanding beyond coding to general knowledge work acceleration.

Krishna is most excited about biotechnology applications, envisioning a world where diseases become curable within patients' lifetimes through AI-accelerated drug discovery.

As described in Machines of Love and Grace, the potential includes transforming drug development, healthcare delivery, and ensuring economic gains accrue globally rather than to a small elite.