Krishna Rao - Anthropic's CFO on Compute, Scaling to $30B ARR, and the Returns to Frontier Intelligence

Key insights

Media referenced

• Machines of Love and Grace - article - Dario Amodei's essay on AI's transformative potential, cited by Rao as the industry's best articulation of the upside case (drug discovery, developing-world living standards).

Companies

• Anthropic - Subject of the episode; CFO Krishna Rao describes compute strategy, pricing, culture, and fundraising.
• Amazon - Trainium chip partner; Anthropic co-develops the chip roadmap with the Annapurna Labs team and signed a deal for up to five gigawatts of Trainium capacity.
• Google - TPU partner; Anthropic signed a five-gigawatt TPU deal with Google and Broadcom starting in 2027, over $100 billion.
• Nvidia - GPU supplier, one of Anthropic's three fungible chip platforms alongside Trainium and TPUs.
• Broadcom - Co-party on the multi-gigawatt TPU deal with Google.
• Microsoft - Named as one of Anthropic's three cloud/hyperscaler partners and distribution engines.
• xAI - Anthropic announced a compute partnership tied to xAI's Colossus facility in Memphis, Tennessee, discussed as an example of opportunistic near-term compute sourcing.
• Meta - Cited as the source of large compensation packages used to poach AI talent industry-wide; Rao says Anthropic lost only two people to it versus dozens lost by other labs.
• OpenAI - Referenced as a peer frontier lab also pushing recursive self-improvement and model capability.
• DeepSeek - The DeepSeek news broke on the same day as the close of Anthropic's Series E, triggering investor re-underwriting of AI valuations broadly.
• FTX - FTX's liquidation of Anthropic shares was part of the backdrop when Rao joined during the Series D raise.
• Airbnb - Rao's prior role; he helped lead Airbnb's financing during the pandemic when revenue fell 70% in seven weeks, an analog he draws on for navigating extreme uncertainty.
• Blackstone - Rao's earlier career stop, in the private equity group; source of his grounding in granular financial analysis.

Techniques and frameworks

• Cone of uncertainty - Anthropic's framework for compute planning: model a range of exponential growth scenarios one to two years out rather than a single point forecast, and buy/build flexibility to cover the spread.
• Compute fungibility across chip platforms - Anthropic runs Trainium, TPUs, and GPUs interchangeably across training, internal use, and serving, via a custom orchestration/compiler layer, to maximize ROI per dollar of compute.
• Jevons paradox pricing - Cutting the price of Opus increased total consumption by more than the price cut, because it unlocked previously underused capability rather than just discounting existing usage.
• Recursive self-improvement - Using current models (Claude Code) to help build and accelerate the next generation of models; over 90% of Anthropic's internal code is written by Claude Code.
• Talent density over talent mass - Anthropic's stated hiring philosophy: a smaller, denser concentration of top AI research and inference engineering talent beats a larger headcount.
• Race to the top - Anthropic's framing for wanting competitors to emulate its safety and responsibility practices rather than treating them as proprietary advantage.

Summary

Krishna Rao, Anthropic's CFO, walks Patrick O'Shaughnessy through the mechanics of running finance inside a business where compute, not headcount or capital structure, is the central constraint. His framing device is the "cone of uncertainty": because Anthropic's growth is exponential and genuinely hard to forecast, the company plans compute purchases against a range of one-to-two-year scenarios rather than a point estimate, deliberately building flexibility into contracts and internal usage so it can adapt as reality lands somewhere in that range. Buying too much compute risks the business financially; buying too little means losing the frontier and failing customers. That tension, and the daily meetings it generates about allocating compute across model training, internal tooling, and customer serving, occupies 30 to 40 percent of Rao's time.

The technical core of Anthropic's compute advantage, as Rao describes it, is fungibility: the company runs Amazon Trainium, Google TPUs, and Nvidia GPUs interchangeably, deploying each chip generation to its best-fit workload through a custom-built compiler and orchestration layer developed over multiple years. He credits this flexibility, plus deep co-development relationships with chip partners (notably Amazon's Annapurna Labs team), for making Anthropic what he claims is the most compute-efficient of the frontier labs. That efficiency compounds with a pricing philosophy built around the Jevons paradox: when Anthropic cut Opus pricing at the 4/4.5 transition, customer consumption rose by more than the discount, because previously underused capability became accessible, letting the company hold prices stable through subsequent model releases rather than repricing each time.

Rao repeatedly returns to "returns to frontier intelligence" as the business's central thesis, illustrated by run-rate revenue moving from roughly $9 billion to over $30 billion within about a quarter, driven by model-led growth rather than sales-force expansion. He argues this is why Anthropic treats compute as a unified pool measured on overall ROI, not a per-customer variable cost, since the same hardware can serve inference in the morning and model training that evening. He also describes recursive self-improvement as already operational inside the company: over 90 percent of Anthropic's internal code is written by Claude Code, which is the practical justification for allocating scarce compute to internal use instead of selling it as customer-facing revenue.

On strategy, Rao positions Anthropic as platform-first, building vertical products like Claude Code or Claude for Financial Services only where the company has a capability lead the market hasn't caught up to, or wants to demonstrate how partners should build on the platform, with the expectation that more economic value accrues to customers than to Anthropic. He connects the company's safety and interpretability research directly to enterprise trust: nine of the Fortune 10 now run sensitive workflows on Claude, and he argues that trust depends partly on Anthropic's track record in alignment and interpretability work, which was pursued for mission reasons but became commercially load-bearing. He also traces how investor skepticism evolved across funding rounds, from doubting whether a frontier model was even necessary at the Series D, to doubting whether enterprise adoption could sustain the growth rate at the Series E, closing the day DeepSeek's release rattled AI valuations broadly.

Rao closes on culture and his own adaptation. He describes a hard cultural-fit hiring bar that rejects highly capable candidates who fail it, a company norm against internal politics and credit-taking, and biweekly unscripted all-hands with Dario Amodei, crediting this culture with retention through the industry's talent wars: Anthropic lost two people to Meta's high-profile compensation offers versus dozens lost by other labs. Personally, he traces his own shift from linear to exponential thinking to an early-2024 walk with chief compute officer Tom Brown, whose predictions sounded like science fiction at the time and have substantially come to pass, and draws on his prior experience financing Airbnb through a 70-percent pandemic revenue collapse as his closest analog for operating without a precedent to follow.

Notable Quotes

"The compute that we procure, it's the lifeblood of our business. It is the most important thing in the company. It's like the canvas on which everything else gets built." - Krishna Rao

"If you buy too much compute, you go out of business. If you buy too little compute, you can't serve your customers." - Krishna Rao

"We don't really think about models as closed or open. We think of them as frontier or not." - Krishna Rao

"Our competitors are incredibly capable and success is far from guaranteed." - Krishna Rao, quoting the sticker on Anthropic employees' laptops

"This is going to bend all paradigms of not just things I've seen, but what most people have seen." - Krishna Rao, recalling his reaction to an early-2024 conversation with chief compute officer Tom Brown