Invest Like The Best: Anthropic's Revenue Exploded From $9B to $30B Run Rate in a Single Quarter — and the CFO Says the Scaling Laws Show No Signs of Stopping
Krishna Rao sits down with Patrick O'Shaughnessy on May 13, 2026, to pull back the curtain on Anthropic's compute strategy, exponential growth, and why frontier intelligence commands a structural premium
From $9 Billion to $30 Billion in One Quarter
The number that will stop most investors cold came almost as an aside. Anthropic CFO Krishna Rao told Invest Like The Best that the company entered 2026 with roughly $9 billion in annualized revenue run rate and exited the first quarter at north of $30 billion. That is not a full-year figure. That is a single quarter. For context, Rao joined the company two years ago when run rate revenue stood at approximately $250 million. The trajectory defies every conventional framework for enterprise software adoption, and Rao was candid that even he spent his early months mentally anchored to linear thinking before being forced to rewire his assumptions.
The engine behind that growth, in Rao's telling, is not sales execution or pricing power in isolation. It is model intelligence compounding on itself. "Each model generation gives you the chance to do more with it, to do it better, to do it more efficiently," he said. "The returns to frontier intelligence are extremely high, and it's extremely high especially in enterprise." Net dollar retention sits above 500 percent on an annualized basis. Nine of the Fortune 10 are now customers. During the car ride to the studio, Rao signed two double-digit million-dollar commitments.
A $100 Billion Compute Commitment and the "Layer Cake" Strategy
Last month, Anthropic signed what amounts to one of the largest compute procurement commitments in corporate history. The company inked a deal with Google and Broadcom for up to five gigawatts of TPU capacity beginning in 2027, and a separate deal with Amazon for up to five gigawatts of Trainium. Combined, the two agreements represent a commitment exceeding $100 billion. Rao also disclosed a newly announced partnership with SpaceX to access its Colossus facility in Memphis for near-term compute, primarily targeting consumer and prosumer workloads.
Rao describes the resulting infrastructure as a "layer cake" — tranches of compute from different providers, different chip generations, and different start dates, all being dynamically orchestrated across three distinct use cases: model training and development, internal acceleration of employees, and serving external customer inference demand. "We use these chips fungibly," he said. Anthropic runs across Amazon's Trainium, Google's TPUs, and Nvidia's GPUs simultaneously, and has built its own compiler stack and orchestration layer to extract maximum utilization from each. Rao believes this makes Anthropic the most efficient user of compute among any frontier lab — a claim he attributes to years of deliberate investment rather than any single architectural decision.
The strategic logic for maintaining this multi-cloud, multi-chip posture is straightforward: flexibility is the hedge against uncertainty. "If you buy too much compute, you go out of business. If you buy too little compute, you can't serve your customers and you're not at the frontier — and that's the same thing," Rao said.
The "Cone of Uncertainty" — and Why It Drives Everything
The single most important conceptual framework Rao uses to run the finance function at Anthropic is what he calls the "cone of uncertainty." On an exponential growth curve, small differences in weekly or monthly growth rates produce radically divergent outcomes over a one-to-two-year horizon. The range of scenarios therefore becomes extremely wide, and point estimates become nearly useless. Anthropic's response is to build flexibility into every compute deal, maintain a hard floor on compute allocated to model development regardless of near-term demand pressure, and dynamically reallocate the remainder based on ROI discussions that happen continuously rather than quarterly.
Rao is explicit that the company has historically operated toward the top end of its cone. The reason capital raising continues at scale — $75 billion raised since he joined, with another $50 billion coming in from the Amazon and Google deals — is not to fund operating losses, he argued, but to pre-position for scenarios where growth continues to exceed expectations. "The reason we raise this capital is more because of that cone of uncertainty than it is to fund actual losses in the business today," he said.
Recursive Self-Improvement Is Already Happening
One of the most consequential disclosures in the conversation was Rao's confirmation that recursive self-improvement is no longer a theoretical construct at Anthropic — it is a present operational reality. More than 90 percent of the code written internally is now produced by Claude Code, and a significant portion of Claude Code itself is written by Claude Code. Anthropic dedicates meaningful internal compute to this loop not because it forgoes revenue to do so, but because the models are materially accelerating the development of the next generation of models.
"The models themselves are helping us to build that next generation of models," Rao said. "In addition to the capability leap you would have just from the scaling laws, talent is really important — and that talent with the best models can really accelerate the development of the capabilities." The company released 30 product and feature updates in January alone, a pace Rao said has only accelerated since. Anthropic's internal thesis is that this recursive loop, combined with the scaling laws it believes remain intact, is widening the gap between frontier labs and everyone else.
Scaling Laws: No Slowdown Observed
On the central question that has divided the AI investment community — whether scaling laws are hitting a wall — Rao was unequivocal. "For us, that's a fair characterization. The scaling laws are not slowing down." He acknowledged the founders include several of the original authors of the scaling laws papers and that the internal culture demands scientific skepticism of previously held assumptions. But from what Anthropic observes across pre-training loss curves, reinforcement learning runs, and ultimately customer feedback on real-world capability, there is no evidence of deceleration. Pre-training, post-training, and reasoning are all moving, and Rao's view is that a true ceiling would require all of them to collapse simultaneously.
Anthropic tracks model efficiency improvements alongside raw capability gains. Using the Opus family as an illustration — from Opus 4 through 4.5, 4.6, and now 4.7 — each generational step delivers not just greater intelligence but a multiplier on token processing efficiency. The implication is that the sports car analogy for AI models is wrong: better models are also cheaper to run. "We actually see both — huge improvements in capability, but also in model efficiency," Rao said. That efficiency gain cascades internally because reinforcement learning is effectively inference within a sandbox, so a more efficient model makes RL runs cheaper too.
Pricing: Jevons Paradox in Action
Anthropic's most significant pricing move in recent history was to cut the price of the Opus family when it launched Opus 4.5. The rationale was straightforward: Opus-class capability was being underutilized because customers were fitting Opus problems into Sonnet-tier pricing. By reducing the price, Anthropic triggered what Rao explicitly described as a Jevons paradox — consumption increased far more than the price cut would have implied at a constant utilization rate. Pricing on Sonnet and Haiku has remained broadly stable. Mythos, the newest and most powerful model, carries a higher price point commensurate with its capability tier.
Rao pushed back on the framing that compute constraints should logically drive price increases across the board. His preferred measuring stick is return on the total compute envelope, treating training, internal use, and inference as a unified pool of capital generating returns across different time horizons. "Our returns on that compute expense today are robust," he said, without providing specific margin figures. The key insight for investors is that as revenue grows rapidly without a corresponding near-term ramp in compute — because capacity is pre-committed 12 months in advance — the incremental economics on serving new demand can be highly favorable within a given compute vintage.
Platform vs. Application: Mostly Horizontal, Selectively Vertical
Rao positioned Anthropic's strategic orientation as primarily a horizontal platform business, drawing an explicit analogy to the early days of AWS. The goal is to build the best models and then offer the tools, SDKs, managed agents, prompt caching, and application-specific overlays that allow enterprise customers to build on top. Claude Code, co-work, Claude for Financial Services, Claude for Life Sciences, and Claude Security are characterized as vertical demonstrations designed to show the ecosystem what the platform can do — not land grabs into application territory.
Rao acknowledged the tension directly. Customers who build businesses on Claude are right to watch how far Anthropic extends down the stack, and he conceded that model releases sometimes surprise even Anthropic's own team. His answer to partner anxiety was essentially structural: Anthropic builds on the same platform it offers customers, vertical launches have been done collaboratively with ecosystem partners, and the company's commercial interest lies in proliferating intelligence broadly rather than capturing every application layer. Whether that posture holds as the platform matures is a question Rao left open.
How Claude Is Running Anthropic's Finance Team
Among the more granular disclosures was Rao's description of how his own finance organization has been transformed. Anthropic now produces statutory financial statements for all legal entities using Claude, with human review as a check rather than the primary production mechanism. It has built an internal platform called Ant Stats, with a library of over 70 finance-specific skills accessible across the team, including a monthly financial review skill that produces the MFR at 90 to 95 percent completion. A weekly revenue and compute utilization report that previously took hours now takes 30 minutes. Critically, Rao said the model does not merely report what happened — it analyzes drivers and causality. "Claude is not just reporting the weather, it's also helping to think about drivers — why did the number change in the way it did."
Rao tracks token usage across his team on a leaderboard and noted that the heaviest users tend to be the most senior people. His head of tax is the top consumer, building policy engines and automating large portions of tax workload. "If we're not super users of this, if we're not pushing the limits of it, how can you expect customers to do that?" he said.
Mythos, Cybersecurity, and the Phased Release Framework
Rao addressed the release of Mythos, which he described as a broadly capable model that happened to exhibit a particularly sharp spike in cybersecurity capability. Where a prior Claude model identified 22 security vulnerabilities in an open-source code base, Mythos found 250. The decision to release it in phases, to a controlled and expanding set of users, was framed as a template for how Anthropic intends to handle future models with asymmetric dual-use risk profiles. The company worked closely with the current U.S. administration in the lead-up to the release. Rao characterized the regulatory dynamic as one where Anthropic wants innovation to proceed quickly but within a responsibility framework — and where Anthropic's investment in interpretability and alignment research creates commercial benefits with enterprise customers that were not originally the motivation for that research.
What Could Derail the Thesis
Rao offered a candid pre-mortem on what could push Anthropic toward the low end of its cone of uncertainty. The first risk is diffusion rate: use cases in large enterprises are playing catch-up to model capability, and human organizational inertia could slow adoption. The second is a slowdown in the scaling laws themselves, which Rao noted cannot be ruled out with certainty even if Anthropic sees no evidence of it today. The third is competitive position: "Today we're at the frontier. I think we're defining the frontier of agentic AI. We need to stay there — and that's not guaranteed either."
The Long Game: Healthcare and Drug Discovery
Asked where he is most personally excited about AI's impact, Rao's answer was unambiguous: biotechnology and drug discovery. His vision is a world where a patient diagnosed with a currently incurable disease could see that disease cured within their lifetime, because AI has compressed the timeline for molecular research, protein analysis, and clinical trial processing by orders of magnitude. "If you think about these molecules and proteins — they're so complex, and such small changes have such big implications for the outcomes — AI is perfect for that," he said. He views a 10x to 100x increase in lab throughput as achievable and transformative in a way that would reach populations globally, including in areas where healthcare resources are scarce.