Summary

The episode’s core thesis, stated as fact by Anthropic’s CFO: “the returns to frontier intelligence are extremely high” — especially in enterprise. Anthropic started the year at ~$9B run-rate revenue and ended the quarter “north of $30 billion”, a jump Rao says was enabled by model-intelligence leaps unlocking new TAM. The use-the-6-month-old-cheaper-model camp keeps being wrong: each generation adds long-horizon capability, not IQ points, and customers invest heavily in more tokens on newer models.
Compute is the business’s lifeblood: “If you buy too much compute, you go out of business. If you buy too little compute, you can’t serve your customers.” Rao spends 30-40% of his time on compute and manages a “layer cake” of deals — 5GW of Google/Broadcom TPUs starting 2027, up to 5GW of Amazon Trainium in an over-$100B commitment, plus the just-announced Colossus Memphis capacity. A large additional compute drop would be “deployed very rapidly.”
The margin heresy investors struggle with most: it doesn’t fit a simple variable-cost-per-customer model. The metric is return on the whole compute envelope, fungible across uses — the same chip runs inference in the morning and model development at night — and those returns are “robust.” Q1’s revenue triple happened without onboarding new compute; capital raised ($75B since Rao joined, $50B more committed from Amazon and Google) reflects uncertainty more than “actual losses in the business today.”
Recursion is live: “90 plus percent of our code is actually written by Claude Code — a lot of Claude Code’s code is written by Claude Code.” Scaling laws are “alive and well” and “not slowing down”; the relevant divide is not open vs. closed but “frontier or not,” and the frontier captures the economic value.
Demand proof for skeptics: net dollar retention over 500% annualized, nine of the Fortune 10 as customers, and Rao signed “two double-digit million-dollar commits” in a 20-minute Uber to the studio. Patrick’s summary: “not pilots anymore.”
Pricing runs on Jevons, not margin capture: cutting Opus pricing at 4.5 sent consumption up “way, way more” than expected, and 4.6 slotted in at the same price. Efficiency compounds with capability — the sports car that also gets better mileage — because RL is “inference within a sandbox with a reward function,” so more efficient inference makes RL more efficient too.
Safety became a commercial advantage: interpretability (“an MRI for the model”) makes better models and supports enterprise trust for the most sensitive workloads. Mythos found 250 security vulnerabilities in an open-source codebase where a prior model found 22 — “kind of scary” — which drove the first phased, defense-focused release, “a template for the future.”
The stated pre-mortem for landing at the bottom of the cone: enterprise diffusion hitting a wall (“the use cases are playing catch-up to the model capability”), scaling laws leveling off (they don’t see it, but won’t claim 100% certainty), or losing the frontier in a competitive market — “that’s not guaranteed, either.”

Deep dive

1. Compute is the lifeblood — mis-buy in either direction and you die

Rao’s opening frame: compute is “the canvas on which everything else gets built” and the hardest decision class in the company — “if you buy too much compute, you go out of business. If you buy too little compute, you can’t serve your customers.” You can’t “buy a gigawatt and have it delivered next week,” so procurement is modeled bottoms-up against demand and frontier requirements. Even today he spends 30-40% of his time on compute.
The “cone of uncertainty”: on an exponential, tiny movements in weekly growth compound into wildly different outcomes, so Anthropic plans across scenarios over a 1-2 year horizon and works backwards — deliberately ranging “towards the top end.” The nightmare is landing at one point in the cone having bought for another; compute efficiency is the bridge that has “really helped us out.”
Allocation happens in standing meetings with no fiefdoms — but with a hard floor under model-development compute they will not breach even if serving customers requires “unnatural things.” The internal-use bucket is genuinely expensive: the compute allocated to employees “could serve billions of dollars of revenue” if redirected, and they spend it anyway because it accelerates model development.

2. Three chip stacks, one orchestration layer — the quiet edge

Anthropic runs Amazon’s Trainium, Google’s TPUs, and Nvidia’s GPUs fungibly across training, internal use, and serving — the only lab on all three chip platforms and all three clouds. The bet took shape when they adopted possibly third-generation TPUs at scale: “people thought, ‘oh, well, you’re crazy — everyone’s using GPUs.’” They build their own compilers “from the chip level up” and match each chip generation to its best workload.
The payoff claim: “a dollar of compute inside our organization goes further than I think it does anywhere else.” The relationships run deeper than procurement — teams work closely with Amazon’s Annapurna Labs to help influence the chip roadmap itself, and the clouds double as distribution engines.
The deal book is a “layer cake of compute”: the Colossus facility partnership in Memphis announced 15 minutes before taping (the host attributed it to xAI in Tennessee; Rao said SpaceX in Memphis), a 5GW Google/Broadcom TPU deal starting 2027, and an Amazon Trainium deal up to 5GW — “an over 100 billion-dollar commitment,” much of it landing this year and next. Every tranche gets scored on price-performance over time, landing date, and workload fit.
On metabolism, Patrick’s thought experiment — air-drop 2x or 10x compute tomorrow — gets a direct answer: Anthropic is constrained across all use cases today, and while a heterogeneous drop would have been hard to absorb a year or two ago, now a large additional compute drop would be “deployed very rapidly” at roughly the current allocation ratios.

3. Returns to frontier intelligence: the $9B→$30B quarter

Rao rejects the IQ frame: “Intelligence for us is multi-dimensional. It’s not just a score” — and “a lot of those benchmarks are saturated”; the real measurement is what customers can do. New generations add long-horizon task ability, tool and computer use, and speed: two equally capable employees, one finishing a week’s assignment in a day, means the second “can be seven times better.”
Each release unlocks TAM, and the proof is the last four months: ~$9B of run-rate revenue at the start of the year to “north of $30 billion” at quarter-end — “that kind of a change is really enabled by these model intelligence leaps,” plus 30 product and feature releases in January alone.
The thesis is enterprise-specific: consumers rarely push model limits, but enterprise customers hit the edge, report it, and then “invest really heavily in more tokens with the newer models” — a cycle Anthropic has “seen play out again and again.” It answers Patrick’s setup: the everyone-catches-up-on-old-models camp “just hasn’t been the case” — even he flips to Opus 4.7 or GPT 5.5 the day it ships.
The demand evidence Rao volunteers from the investor’s chair: net dollar retention over 500% on an annualized basis, nine of the Fortune 10 as customers, and “I signed two double-digit million-dollar commits like in the car ride” — a 20-minute Uber. Patrick’s gloss: “not pilots anymore.”

4. Sports cars that get better mileage — efficiency compounds with capability

The car analogy, as told: people assume moving from sedan to sports car costs you gas mileage. “In our case, we actually see both improvements” — going from Opus 4 to 4.5, 4.6, and now 4.7, “each one of those leaps, they’re not equal, but each one has a multiplier” in token-processing efficiency.
The flywheel: reinforcement learning “is basically inference within a sandbox with a reward function, right? So if the model’s better at more efficient inference, that RL is more efficient as well” — a win-win where customers get more capability while Anthropic serves it at a multiple more efficiently, with efficiency improvements deployed dynamically between model generations.
On scaling laws: “alive and well… the scaling laws are not slowing down.” He notes founders authored the scaling-laws papers, but insists “we can be a bit of a skeptical bunch” — a scientific-method culture constantly re-testing priors against pre-training loss curves, RL, and customer feedback, where pain points become training targets (no training on enterprise data; prosumer only opt-in).

5. Recursion is underway — and what would bend the cone down

Why forego revenue for internal compute: “90 plus percent of our code is actually written by Claude Code. A lot of Claude Code’s code is written by Claude Code.” The models are building the next generation of models; best talent plus best models accelerates the capability curve beyond what scaling alone gives.
On the open-source gap question, Rao redraws the axis: “We don’t really think about models as closed or open. We think of them as frontier or not” — the frontier captures the economic value. The complement is “talent density beats talent mass.”
The forecasting discipline that exponentials force: scenarios, not point estimates, with “a very low bar for updating your current priors — it could be the case that something a month ago was true that’s just not true today and that breaks your model.” Quarterly board-meeting forecasting “doesn’t work.” The Sonnet 3.5/3.6 coding jump — capability, then adoption, then revenue — is now the pattern-matching template for the rest of the economy.
His pre-mortem for the bottom of the cone: first, diffusion — “the use cases are playing catch-up to the model capability,” and change inside large human organizations is hard; second, scaling laws leveling off, which “we don’t see” but can’t rule out with 100% certainty; third, losing the frontier itself — “it’s a competitive market… that’s not guaranteed, either.”

6. Pricing for Jevons; margins as return on the compute envelope

Context Rao insists on: the company is “only a little over 5 years old,” this past March was the third anniversary of the first dollar of revenue, and it only had “a frontier model for real” from March 2024. Against that, pricing across Haiku, Sonnet, and Opus has been remarkably stable — very few changes.
The biggest change was a cut: Opus-class models were “underutilized relative to their capability” — customers “trying to fit an Opus problem into a Sonnet workload” — so efficiency gains funded a price reduction at Opus 4.5. Result: “this Jevons paradox — we lowered the price of it, but the consumption went up way, way more than what you would have expected,” and Opus 4.6 then slotted into existing workloads at the same price.
On why not just set a healthy margin: the business doesn’t reduce to a simple unit-cost model. Everything the compute does — inference today, model development that “unlocks TAM that drives revenue six months from now,” internal acceleration — supports revenue on different time scales, and “our returns on that compute expense today are robust.” The tell: the Q1 revenue explosion happened without onboarding new compute; capacity was set by ramps agreed ~12 months earlier.
This is also what investors find hardest: the fungibility paradigm. “You can’t repurpose R&D people into COGS” at a software company or a factory — here one chip does inference in the morning and model development in the evening, and that flexibility “is actually what helps to drive revenue short-term and long-term.”

7. Mostly platform; vertical only to build ahead and demonstrate

The strategy is “mostly horizontal,” explicitly analogized to early-days AWS: platform accrues a lot of value, but “the customers who are building on that platform actually create even more value.” The Claude platform — prompt caching, virtual machines, Claude Code dispatch, the agent SDK, managed agents — is a set of “vectors to access that model intelligence.”
They go vertical in two cases: building ahead of model capability (Claude Code was “Claude-led” where the market was developer-led — “we think the models can’t quite do that today… but we think they’ll get there, and they have”), and demonstrating patterns for the ecosystem to emulate — Claude for financial services, life sciences, security — launched in partnership, on the same platform customers use: “a level playing field.”
On customers being scared of Anthropic as a competitor, Rao doesn’t fully deny the dynamic: model capabilities “sometimes even surprise us,” and what took prior platform waves 5-20 years “is happening in months now.” The answer offered is partner orientation — early access programs, listening on capabilities — with the claim that Anthropic can actually accelerate front-footed adopters.

8. Anthropic runs on Claude — the finance team as the proof of concept

The dogfooding is concrete: statutory financial statements for all legal entities are produced with Claude (a human checks); a real-time platform called Ant Stats plus a library of ~70 finance-specific Claude skills; and an MFR skill that produces the monthly financial review “90 to 95% ready.” The point: “Claude is not just reporting the weather” — it explains drivers, and weekly revenue and compute-utilization reports went from hours to 30 minutes.
Rao keeps a token-usage dashboard (“we don’t compensate people on it — no one’s trying to token-max”), and the most senior people are the biggest users — the head of tax is number one, automating tax-policy workloads. “If we’re not super users of this, how can you expect customers to do that?”
Patrick’s pushback — worth keeping: it feels “ever so slightly dystopian” that “I’m just doing what it tells me versus me telling it what to do.” Rao’s reframe is “Jevons paradox, but for labor”: people became much more productive and they hired more people because of it, with time shifting from tying out numbers to deciding where to reinvest.
The frontier from inside is the “virtual collaborator” — organizational context, your tools, memory that learns from mistakes, and work over long horizons “on not just a task, but an actual idea.” Co-work (which grew out of using Claude Code as a digital co-worker) is “growing faster than Claude Code was if you index them to the same point in time,” and product development is now “shipping daily” with fleets of agents — “everyone kind of becomes a manager.”

9. Investors kept fitting the wrong mold — and safety became a commercial advantage

The Series D, closing as Rao joined two years ago, “was not a straightforward fundraising”: the company only got a frontier model mid-raise, and the FTX transaction was liquidating Anthropic shares at the tail end. Investor objections were mold-fitting — why do you need a frontier model, “aren’t AI safety and building a really big business at odds?”, your sales force is too small to be an enterprise software company.
The Series E closed at the end of 2024 near $1B run rate — and the first close landed on the day the DeepSeek news broke. Even supportive investors said “there’s no way you’re going to be able to keep it up… look at how long it took with cloud.” Rao owns his own linear thinking: he joined at $250M run rate, saw a plan to $1B and asked “Great, in what year?” — and “Dario has been a much better predictor of the revenue than I have.”
The non-obvious interlink: safety research done for mission reasons turned out commercial. Interpretability is “like an MRI for the model,” and “it turns out if you can look inside the model, you’re better at building them” — while alignment and trust helped support enterprise adoption: nine of the Fortune 10 put their most sensitive workloads on Claude.
Capital math: $75B raised since Rao joined, another $50B coming from the Amazon and Google deals closed last month — raised “more because of that uncertainty than to fund actual losses in the business today.” His questions if he sat in the investor’s seat grilling labs: all-up ROI on compute and its shape over time; whether customers are testing or deploying at meaningful scale; and where future compute comes from when your suppliers also compete for it.

10. Mythos, the government, and a culture that lost two people to Meta

Mythos was “maybe misconstrued as just a cyber model” — it’s broadly capable but spiked in cyber: an open-source codebase where a prior model found 22 security vulnerabilities, “Mythos then found 250. So that is kind of scary, right?” Hence the first phased release — an expanding access group, defensive framing — “a template that could be used for the future.” On government pre-approval proposals and the Department of War episode: “regulation has a role to play,” the posture is “very America first,” and the Mythos process with the administration is his example of balancing innovation speed with responsibility.
On AI polling below Congress: the industry should paint the picture (Dario’s Machines of Loving Grace — drug development, rare disease, healthcare in the developing world) while honestly articulating risks: “if somebody’s just telling me all the good news and none of the bad news… do I really trust this perspective?” Rao’s own top hope is drug discovery when lab throughput rises 10-100x: “you’re diagnosed with a disease that is not curable, but in your lifetime that cure can be found… you actually might not die of that disease.”
The culture, described for his parents: seven co-founders, all still at the company; a culture interview that is a genuine bar (“somebody could be flying colors on everything else… we won’t hire them”); no fiefdoms, no confetti at milestones — “what’s next?”; Dario fronting the whole company every two weeks with unplanted questions. The empirical test: when Meta came with huge packages for LLM talent, “I think we lost two people. And other labs lost dozens.”
The personal arc: a 2.5-hour walk around the Mission with chief compute officer Tom Brown in early 2024 — “if even 10% of that is true, this is going to bend all paradigms,” and much of it has come true. His prior unprecedented crisis was leading Airbnb’s pandemic financing after it lost 70% of revenue in seven weeks; the internal phrase is “holding light and shade.” His closer, on kindness: his older brother quietly chose an in-state college 25-30 years ago, wanting to give Krishna the opportunity to go wherever he wanted.

The Founders Who Left Tesla to Rebuild America | a16z How We Grew Koch Inc. to $150 Billion Without Going Public: Charles & Chase Koch