Summary of Krishna Rao - Anthropic's CFO on Compute, Scaling to $30B ARR, and the Returns to Frontier Intelligence - [Invest Like the Best, EP.471]

Overview of Invest Like the Best with Krishna Rao

In this episode, Patrick O’Shaughnessy speaks with Krishna Rao, CFO of Anthropic, about how the company thinks about compute as the core asset powering model training, product delivery, and internal productivity. The conversation focuses on Anthropic’s capital allocation discipline, its use of multiple chip platforms, the returns to staying at the frontier of AI capability, and how the company balances being a platform with building its own applications. Rao also discusses pricing, margins, investor perceptions, Anthropic’s culture, regulatory/safety considerations, and where he sees the biggest long-term opportunity: AI-driven advances in biotechnology and healthcare.

The central role of compute at Anthropic

• Rao describes compute as the “lifeblood” of the business and “the canvas on which everything else gets built.”
• Anthropic’s biggest challenge is deciding:
  • how much compute to buy,
  • when to buy it,
  • and how to allocate it across training, internal R&D, and serving customer demand.
• The company treats compute planning as a high-stakes, probabilistic exercise because:
  • buying too much could be financially dangerous,
  • buying too little could prevent them from serving customers or staying at the frontier.
• Anthropic uses a “cone of uncertainty” framework to plan for a wide range of possible growth outcomes, since the business is growing exponentially rather than linearly.

How Anthropic uses compute flexibly

Three chip platforms

Anthropic runs compute across three major platforms:

• Amazon’s Trainium
• Google’s TPUs
• NVIDIA GPUs

Fungibility and orchestration

• The company has built systems that allow it to use these chips interchangeably across different workloads.
• Compute is used for:
  • model development,
  • internal acceleration,
  • customer inference and serving.
• Rao says this flexibility took years to develop and is a major source of efficiency.
• Anthropic has also worked closely with chip and cloud partners to influence hardware roadmaps and build custom compilers and infrastructure.

Frontier intelligence and the “returns” to being ahead

• Rao argues that the returns to frontier AI capability are still extremely high, especially in enterprise.
• He rejects the idea that users can simply stick with older, cheaper models and “catch up” later.
• In practice, better models unlock:
  • more capabilities,
  • longer-horizon tasks,
  • better tool use and agentic workflows,
  • more real-world ROI for customers.
• Anthropic sees this play out in revenue growth and customer adoption:
  • the company reportedly grew from about $9B to $30B in run-rate revenue over a recent four-month period.
• The core thesis: better frontier models unlock new TAM, not just cheaper delivery of the same tasks.

Scaling laws, model improvement, and recursive self-improvement

• Rao says Anthropic believes scaling laws remain alive and well.
• The company evaluates progress through:
  • pre-training loss curves,
  • reinforcement learning progress,
  • customer feedback on real-world performance.
• Customer pain points are treated as training targets, though Anthropic does not train on enterprise customer data unless explicitly opted in.
• He also highlighted how the company increasingly uses its own models to build better models:
  • much of Anthropic’s code is now written by Claude Code,
  • internal tools accelerate research and product development.
• The company sees this as an early form of recursive self-improvement, though still heavily guided by human talent.

Platform vs. applications: where Anthropic wants to play

• Anthropic views most of its business as a platform business:
  • APIs,
  • model access,
  • prompt caching,
  • agents SDKs,
  • managed tooling.
• But it also builds select applications when it believes it can:
  • demonstrate the future of the platform,
  • lead the market,
  • or create a useful reference implementation.
• Claude Code is the clearest example of this “build ahead of the models” strategy.
• The company generally prefers a horizontal strategy, while selectively going vertical in areas like:
  • security,
  • financial services,
  • life sciences.

Pricing, margins, and capital intensity

• Anthropic has kept pricing relatively stable across its model families.
• Rao emphasized that the company often prefers to lower prices when it believes a model is underused relative to its capability.
• The goal is to drive usage, unlock more value for customers, and create a Jevons-paradox effect:
  • lower prices can lead to much higher usage.
• On margins, Anthropic focuses less on traditional software-style unit economics and more on the return on its full “compute envelope” across:
  • training,
  • inference,
  • internal acceleration.
• The company has raised massive amounts of capital, but Rao frames this as necessary to support growth and optionality, not primarily to fund losses.

Customer demand and enterprise traction

• Anthropic is seeing strong enterprise adoption:
  • it now sells to nine of the Fortune 10.
• Rao argues customers are using Claude in real production workflows, not just pilots.
• He points to strong net dollar retention and meaningful repeat usage as evidence that enterprise ROI is real.
• He also notes that demand could absorb substantially more compute if it were available, though the company still needs to allocate it across competing needs.

Anthropic’s internal use of Claude

• Anthropic uses Claude extensively inside the company, especially in finance and operations.
• Examples include:
  • producing statutory financial statements,
  • generating monthly financial reviews,
  • automating tax workflows,
  • speeding up reporting and memo writing.
• Rao says a large portion of the finance team’s work now takes far less time, allowing the team to focus on interpretation and strategy instead of manual reconciliation.
• He views this as a productivity accelerant rather than a threat to human work.

Culture: collaboration, humility, transparency

• Rao describes Anthropic’s culture as unusually collaborative and low-ego.
• Key cultural traits:
  • no fiefdoms,
  • strong intellectual honesty,
  • rigorous debate,
  • high transparency,
  • alignment after decisions are made.
• Dario Amodei’s open company updates and regular written communication were highlighted as important cultural mechanisms.
• Anthropic is highly selective on culture fit, and Rao says that helps the company retain talent even when competing firms offer massive compensation packages.

Safety, regulation, and the “Mythos” release

• Rao emphasized that Anthropic takes safety seriously and sees it as part of its mission.
• The discussion around the model release he described as “Mythos” illustrated Anthropic’s cautious approach:
  • the model was especially strong in cyber-related capabilities,
  • Anthropic chose a phased release rather than a broad launch,
  • the goal was to enable defensive use while reducing misuse risk.
• He argued this is a template for future releases: powerful models may require staged access and close oversight.
• On regulation, Anthropic is working closely with government and believes innovation and responsibility must coexist.

The biggest misunderstandings investors still have

• Earlier investor questions centered on:
  • whether Anthropic needed a frontier model,
  • whether safety and scale were compatible,
  • whether enterprise adoption would be too slow.
• Today, a major misunderstanding is still how compute works:
  • many still think of it like a simple variable cost,
  • Anthropic views it as a shared, fungible resource that supports multiple workstreams.
• Rao says investors increasingly understand the business, but the dynamic compute model remains hard for people to fully internalize.

The frontier Anthropic is aiming for next

• Rao sees the next major frontier as the “virtual collaborator”:
  • an AI system with memory,
  • organization-specific context,
  • tool use,
  • long-horizon task execution,
  • and the ability to learn from mistakes.
• He believes this could transform knowledge work broadly, across a roughly $40T global market.
• The company is already seeing early versions of this in coding and in its internal products like Claude Code and Cowork.

What could slow Anthropic down

Rao identifies a few risks to the optimistic path:

• enterprise diffusion of AI could slow if organizations adopt more slowly than expected,
• model capability improvements could level off,
• Anthropic could lose its frontier position in a competitive market.

What excites Rao most

• His strongest long-term excitement is in biotechnology and healthcare.
• He believes AI can:
  • accelerate drug discovery,
  • speed up clinical and regulatory workflows,
  • help discover treatments faster for both common and rare diseases.
• He sees the potential for 10x or 100x more experimental throughput in scientific research.

Personal reflection and closing note

• Rao says the biggest personal challenge has been learning to think exponentially instead of linearly.
• He credits strong hiring, first-principles thinking, and staying intellectually open to new realities.
• The most memorable kindness he shared was from his older brother, who chose a college path partly to preserve Rao’s future options and opportunities.
• That story underscored a theme that ran through the interview: long-term thinking, sacrifice, and creating room for future possibility.