Krishna Rao on Anthropic’s Compute Economics, the $100 Billion Commitment, and Financing Frontier AI

Anthropic’s chief financial officer explains why compute — bought a year ahead, used fungibly across three chip platforms — is the single lever that governs how fast the company can grow, and why he measures the whole business by return on that compute rather than by software-style margins.

Key ideas

1. Compute is the business’s binding constraint and its riskiest decision. Buy too much and you go out of business; buy too little and you fall off the frontier and cannot serve customers. Rao spends 30–40% of his time on it, planning purchases a year or more ahead against a ‘cone of uncertainty’ of exponential-growth scenarios.
2. Fungibility is the edge. Anthropic runs on Amazon Trainium, Google TPUs, and Nvidia GPUs, and moves a single chip between customer inference, model development, and internal use through the day. That orchestration layer, built over years, is what he calls being the most efficient user of compute among the frontier labs — and the hardest thing to explain to investors.
3. Growth is genuinely exponential. Run-rate revenue ran from about $9bn at the start of the year to north of $30bn a quarter later — roughly $250m when Rao joined two years earlier — driven by leaps in frontier model capability rather than by adding sales headcount.
4. The measuring stick is return on compute, not variable cost. Because the same compute supports revenue today (inference), in six months (model development), and internal acceleration, treating it as a per-customer variable cost mis-frames the business. Anthropic reports robust returns on the full compute envelope.
5. Safety spend turns out to pay commercially. Interpretability and alignment research — pursued for the mission — make models easier to build and give the nine-of-the-Fortune-10 enterprise customers a company they can trust with their most sensitive workloads.

Summary

Compute as the canvas — and the hardest call in the company

Rao frames procured compute as ‘the lifeblood of our business’ and ‘the canvas on which everything else gets built’. The decision is brutal because it is asymmetric and slow: you cannot buy a gigawatt and have it delivered next week, so purchases are modelled bottom-up against demand a year or two out. He works from a ‘cone of uncertainty’ — a fan of exponential scenarios — and plans towards the top end while building flexibility into both the deals and the usage. Flexibility means running three chip platforms (Amazon’s Trainium, Google’s TPUs, Nvidia’s GPUs) fungibly across model development, internal acceleration, and customer serving, with an orchestration layer and in-house compilers built over several years. A floor of compute is ring-fenced for model development and never breached, because the returns to frontier intelligence — especially in enterprise — are the core thesis.

The exponential, and breaking linear habits of mind

The recurring theme is that humans think linearly and this business does not. Rao joined at roughly $250m of run-rate revenue and instinctively asked which year the plan reached a billion — the wrong question. Revenue moved from about $9bn to north of $30bn within a quarter, unlocked by model-capability leaps rather than headcount: each new Opus generation both raises capability and multiplies token-processing efficiency, so a price cut on Opus 4.5 triggered a Jevons-paradox jump in consumption. Internally, 90%-plus of Anthropic’s own code is now written by Claude Code, and the research lab remains upstream of everything. The company forecasts as a range of scenarios with a low bar for updating priors, using coding — where capability, adoption, and revenue arrived in sequence from Sonnet 3.5 onward — as the pattern for the rest of the economy.

The $100 billion commitment and financing the frontier

The capital story is large but, Rao argues, not speculative. He signed a five-gigawatt TPU deal with Google and Broadcom starting 2027 and an Amazon Trainium deal for up to five gigawatts — together an over-$100bn commitment, much of it landing across this year and next as a ‘layer cake’ of compute arriving at different times with different price-performance. Anthropic has raised roughly $75bn since he joined, with a further $50bn to come from the Amazon and Google deals; near-term capacity, such as the SpaceX Colossus facility in Memphis, is bolted on opportunistically wherever it can be deployed productively. The through-line: the raise funds growth against a wide cone of uncertainty, not losses, because the business runs efficiently. The fundraising history reads as a sequence of external shocks — a Series D closing as FTX liquidated Anthropic shares, a Series E first-close on the day of the DeepSeek news — met each time by the thesis proving out.

Return on compute, not variable cost

Rao’s sharpest correction to investor mental models is to reject the software-margin frame. Compute is not a per-customer variable cost; it is a single fungible resource that runs inference in the morning and model development in the evening, supporting revenue over different time horizons. So the right question is the return on the whole compute envelope, which he reports as robust. Pricing has been deliberately stable — few changes, the biggest being the Opus price cut that widened access — because the goal is to proliferate the intelligence and let customers capture most of the value, AWS-style, on a platform that is mostly horizontal with selective vertical products (Claude Code, Claude for financial services, Claude security). Net dollar retention runs over 500% annualised; customers are signing eight-figure commitments, not pilots.

Safety as a commercial asset, and the culture that holds talent

Investment in interpretability (‘an MRI for the model’) and alignment science was made for the mission, but Rao traces two downstream commercial effects: seeing inside a model makes you better at building it, and enterprises entrusting their most sensitive workflows to Claude want a company they can trust. The phased, government-engaged release of the Mythos model — capable across many dimensions but spiking on cyber, where it found 250 vulnerabilities in a codebase a prior model surfaced 22 in — is offered as a template for responsible deployment. Underpinning it is a culture Rao credits for retention: seven co-founders still present, a genuine culture-interview gate, rigorous debate resolving into real alignment, and Dario Amodei fronting an unscripted all-hands every two weeks. When rivals dangled outsized packages, he says Anthropic lost two people where other labs lost dozens.

Speakers

• Krishna Rao — Chief Financial Officer of Anthropic; joined in early 2024 during the Series D, having previously helped lead Airbnb’s pandemic financing and worked in Blackstone’s private equity group.

See also

• Gavin Baker on AI Infrastructure, Power Constraints, and Semiconductor Investing — a prior Invest Like the Best episode reading the same AI build-out through its hardware and capital constraints.
• Scaling Laws — the compute-scaling trajectory Rao says is ‘alive and well’ and underpins the frontier thesis.
• Big Blob of Compute Hypothesis — the treat-compute-as-one-fungible-resource idea that Rao’s cross-platform orchestration puts into practice.
• Token Economics — pricing stability, the Opus price cut, and the Jevons-paradox jump in token consumption.
• Five-Layer AI Cake — the platform-versus-application question and where value accrues on Anthropic’s stack.
• Private Credit — the financing backdrop to a $100bn-plus, capital-intensive compute commitment.
• Responsible Scaling Policy — the phased, government-engaged Mythos release as a deployment template.
• Mechanistic Interpretability — the ‘MRI for the model’ research Rao ties to both safety and enterprise trust.
• Dario Amodei — Anthropic co-founder and CEO, named as the sharper revenue forecaster and author of Machines of Loving Grace.
• Patrick O'Shaughnessy — host of Invest Like the Best.