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

TAKEAWAYS

My biggest takeaways from Krishna Rao, CFO of Anthropic

01

Buy too much compute and you go out of business; buy too little and you fall off the frontier.

Krishna calls compute the most important thing in the company, the canvas on which everything else gets built, which makes the procurement decision existential in both directions. His answer is the 'cone of uncertainty': model a range of one-to-two-year scenarios, work backwards from them, and build flexibility into every deal so a miss on either side is survivable.

“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.”
Krishna Rao · 3:53
02

The returns to frontier intelligence are not slowing down.

This is the core thesis of the business for Krishna, and enterprise is where it shows up hardest. Run-rate revenue went from about $9 billion at the start of the year to $30 billion by the end of the quarter, Anthropic now sells to nine of the Fortune 10, and net dollar retention runs over 500% annualized. Each new model unlocks TAM, customers pour in more tokens, and the cycle repeats.

“that's a core thesis of our business set, especially in enterprise, the returns to frontier intelligence are not slowing down”
Krishna Rao · 13:02
03

Compute is not a variable cost; it is a fungible envelope, and the right measure is the return on the whole thing.

Krishna says the hardest thing for investors to grasp is that treating compute as a per-customer marginal cost forces Anthropic into a software paradigm that doesn't fit. The same chips run inference in the morning and model development in the evening, across Trainium, TPUs and GPUs, on an orchestration layer built over years. So Anthropic measures the return on the full compute envelope, not the cost to serve one customer.

“This idea of a variable cost that's on the incremental to serve a customer, it doesn't really fit our business. It tries to maybe fit our business into a software paradigm. That's not the case.”
Krishna Rao · 35:44
04

The frontier ahead is a virtual collaborator for $40 trillion of knowledge work.

Krishna sizes the target as roughly $40 trillion of knowledge work done annually and describes the goal as something with context inside your organization, use of your specific tools, memory, and long-horizon ability, mirroring what already happened in coding. Early signal: Cowork is growing faster than Claude Code was when indexed to the same point in time.

“I think it is towards this vision or this goal of like a virtual collaborator. And so think of this as something that has context within your organization, that can use all of the tools that are specific to you”
Krishna Rao · 1:04:39
05

The models are now building the next models.

Krishna reports that over 90% of Anthropic's code is written by Claude Code, much of it Claude Code writing its own code, and the company shipped 30 product and feature releases in January. That is why Anthropic forgoes near-term revenue to keep compute on model development: each generation accelerates the one after it, and the leap from Opus 4 to 4.5 to 4.6 to 4.7 brought a multiplier in token-processing efficiency each time.

“right now, within the company, 90 plus percent of our code is actually written by Claude Code”
Krishna Rao · 17:11
06

There is a floor on model-development compute that Anthropic will not go below, even if customers suffer for it.

Krishna is explicit that serving customers does not get first claim on the chips. When compute is tight, Anthropic will do unnatural things on the serving side before it cuts research below the floor, because the long-term return on being at the frontier outweighs any quarter's revenue.

“There's a level of compute for model development that we will not go below. Even if it means it's harder to serve customers or we have to do unnatural things when it comes to that, we want to continue to make that long-term investment in developing the best models because we think the returns to frontier intelligence are extremely high.”
Krishna Rao · 9:21
07

Cutting Opus's price didn't shrink revenue; it triggered a Jevons paradox.

Krishna says pricing has stayed stable across Haiku, Sonnet and Opus, and the one big move was bringing down the price of the Opus family with Opus 4.5. Consumption rose far more than expected, which he reads as validation for pricing to value rather than raising prices as the models get better.

“Changing the pricing for Opus, you see this Jevons Paradox. We lowered the price of it, but the consumption went up way, way more than what you would have expected.”
Krishna Rao · 33:18
08

Safety work done for the mission turns out to be an enterprise sales asset.

Investors keep asking Krishna whether AI safety and the business are in tension. His answer is that the investment in safety, interpretability and alignment was made for mission reasons, but enterprises handing over sensitive workloads want a company they can trust, so it pays back commercially downstream. He puts it alongside the other recurring investor doubts: that you need a frontier model at all, and that a small salesforce can sell this.

“When you have this investment that we made and will continue to make in safety, interpretability, alignment, that actually endures to the benefit of the enterprise customers as well”
Krishna Rao · 45:00
09

Culture, not compensation, is why the talent stays.

Krishna's evidence is retention: all seven co-founders and most of the first 20 to 30 employees are still there, and when Meta and others made mega-offers, Anthropic lost two people while other labs lost dozens. He describes the culture as rigorous debate followed by real alignment, biweekly transparent all-hands from Dario, no fiefdoms, and a humility captured in the line 'success is far from guaranteed.'

“The culture underpins the reason why we've been able to attract and retain some of the best talent in the industry, because we don't always pay people the most.”
Krishna Rao · 59:55
10

Quarterly point-estimate forecasting doesn't work at Anthropic; run scenarios and keep a low bar for updating priors.

Krishna runs finance the way the company runs research: pre-training, RL and customer pain points all feed one loop, and from what Anthropic sees the scaling laws are not slowing down, so a plan revisited at the next board meeting is stale by then. He recalls joining at about $250 million of run-rate with a plan to reach a billion and asking, 'in what year?'

“this old like, well, we'll forecast once a quarter and we'll revisit this in three months at the next board meeting. That doesn't work for our business.”
Krishna Rao · 22:12
11

Mythos's phased release is the template for shipping a model that spikes on a dangerous capability.

Krishna describes Mythos as broadly capable but with a spike on cyber: on an open-source codebase where a prior model found 22 security vulnerabilities, Mythos found 250. Rather than a general launch, Anthropic released it in expanding phases focused on defensive use, and he thinks that pattern will carry to future capable models.

“we think that's a template that could be used for the future, but because of this one particular area, we wanted to be cognizant of that in how we released it”
Krishna Rao · 56:50
12

AI is a Jevons paradox for labor: more productive people mean more hires, not fewer.

Krishna's own finance team is the proof. Statutory financial statements for every legal entity are produced with Claude and checked by a human, the team keeps a library of over 70 finance-specific skills, the monthly financial review comes out 90-95% ready, and weekly revenue reports dropped from hours to about 30 minutes. The result has been hiring more people, because there is no shortage of work to do.

“I think of it again like Jevin's paradox, but for labor, which is that we have people become incredibly more productive. We've hired a lot more people because of that, because there's no shortage of work to do.”
Krishna Rao · 42:07

post Copy as a post — the same 12 takeaways as plain text, timestamps included

My biggest takeaways from Krishna Rao, CFO of Anthropic:

1. Buy too much compute and you go out of business; buy too little and you fall off the frontier. Krishna calls compute the most important thing in the company, the canvas on which everything else gets built, which makes the procurement decision existential in both directions. His answer is the 'cone of uncertainty': model a range of one-to-two-year scenarios, work backwards from them, and build flexibility into every deal so a miss on either side is survivable. [3:53]

2. The returns to frontier intelligence are not slowing down. This is the core thesis of the business for Krishna, and enterprise is where it shows up hardest. Run-rate revenue went from about $9 billion at the start of the year to $30 billion by the end of the quarter, Anthropic now sells to nine of the Fortune 10, and net dollar retention runs over 500% annualized. Each new model unlocks TAM, customers pour in more tokens, and the cycle repeats. [13:02]

3. Compute is not a variable cost; it is a fungible envelope, and the right measure is the return on the whole thing. Krishna says the hardest thing for investors to grasp is that treating compute as a per-customer marginal cost forces Anthropic into a software paradigm that doesn't fit. The same chips run inference in the morning and model development in the evening, across Trainium, TPUs and GPUs, on an orchestration layer built over years. So Anthropic measures the return on the full compute envelope, not the cost to serve one customer. [35:44]

4. The frontier ahead is a virtual collaborator for $40 trillion of knowledge work. Krishna sizes the target as roughly $40 trillion of knowledge work done annually and describes the goal as something with context inside your organization, use of your specific tools, memory, and long-horizon ability, mirroring what already happened in coding. Early signal: Cowork is growing faster than Claude Code was when indexed to the same point in time. [1:04:39]

5. The models are now building the next models. Krishna reports that over 90% of Anthropic's code is written by Claude Code, much of it Claude Code writing its own code, and the company shipped 30 product and feature releases in January. That is why Anthropic forgoes near-term revenue to keep compute on model development: each generation accelerates the one after it, and the leap from Opus 4 to 4.5 to 4.6 to 4.7 brought a multiplier in token-processing efficiency each time. [17:11]

6. There is a floor on model-development compute that Anthropic will not go below, even if customers suffer for it. Krishna is explicit that serving customers does not get first claim on the chips. When compute is tight, Anthropic will do unnatural things on the serving side before it cuts research below the floor, because the long-term return on being at the frontier outweighs any quarter's revenue. [9:21]

7. Cutting Opus's price didn't shrink revenue; it triggered a Jevons paradox. Krishna says pricing has stayed stable across Haiku, Sonnet and Opus, and the one big move was bringing down the price of the Opus family with Opus 4.5. Consumption rose far more than expected, which he reads as validation for pricing to value rather than raising prices as the models get better. [33:18]

8. Safety work done for the mission turns out to be an enterprise sales asset. Investors keep asking Krishna whether AI safety and the business are in tension. His answer is that the investment in safety, interpretability and alignment was made for mission reasons, but enterprises handing over sensitive workloads want a company they can trust, so it pays back commercially downstream. He puts it alongside the other recurring investor doubts: that you need a frontier model at all, and that a small salesforce can sell this. [45:00]

9. Culture, not compensation, is why the talent stays. Krishna's evidence is retention: all seven co-founders and most of the first 20 to 30 employees are still there, and when Meta and others made mega-offers, Anthropic lost two people while other labs lost dozens. He describes the culture as rigorous debate followed by real alignment, biweekly transparent all-hands from Dario, no fiefdoms, and a humility captured in the line 'success is far from guaranteed.' [59:55]

10. Quarterly point-estimate forecasting doesn't work at Anthropic; run scenarios and keep a low bar for updating priors. Krishna runs finance the way the company runs research: pre-training, RL and customer pain points all feed one loop, and from what Anthropic sees the scaling laws are not slowing down, so a plan revisited at the next board meeting is stale by then. He recalls joining at about $250 million of run-rate with a plan to reach a billion and asking, 'in what year?' [22:12]

11. Mythos's phased release is the template for shipping a model that spikes on a dangerous capability. Krishna describes Mythos as broadly capable but with a spike on cyber: on an open-source codebase where a prior model found 22 security vulnerabilities, Mythos found 250. Rather than a general launch, Anthropic released it in expanding phases focused on defensive use, and he thinks that pattern will carry to future capable models. [56:50]

12. AI is a Jevons paradox for labor: more productive people mean more hires, not fewer. Krishna's own finance team is the proof. Statutory financial statements for every legal entity are produced with Claude and checked by a human, the team keeps a library of over 70 finance-specific skills, the monthly financial review comes out 90-95% ready, and weekly revenue reports dropped from hours to about 30 minutes. The result has been hiring more people, because there is no shortage of work to do. [42:07]

— Invest Like the Best, "Krishna Rao - Anthropic's CFO on Managing Compute, Scaling to $30B ARR, and the Returns to Frontier Intelligence" (2026-05-13)

.txt .json