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

Krishna Rao

The compute we procure is the lifeblood of our business. It is the most important thing in the company. It's the canvas on which everything else gets built. So the decisions we make in how much compute to buy are some of the most consequential and hardest decisions in the entire company. Think of it this way: 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. So we talk a lot about this cone of uncertainty. These purchases have real-world implications — you can't just go out and buy a gigawatt of compute and have it delivered next week. You have to think ahead and plan for it.

Krishna Rao

We take a very disciplined approach. We look bottoms-up: we model what we think demand will be — and sometimes we get that wrong — we think about the compute we need to stay at the frontier, and we try to estimate that well ahead. Then, as we go out and actually do these deals, flexibility is really important to us. We build flexibility into the deals themselves, and into how we use the compute, because the way we bridge from where we are today to where we want to go, when the business is growing exponentially, is to use that compute as efficiently as possible. I'd say I spend 30 or 40% of my time on compute even today.

Krishna Rao

A couple of different things. Number one, we use three different chip platforms — we're customers of Amazon's Trainium chip, Google's TPUs, and Nvidia's GPUs. We use these chips fungibly. If you think about the compute we buy, we use it for model development, we use it internally to speed up our own product and model development, and we use it to serve customers. Across those three platforms, we're using compute for all of those internal and external uses. That flexibility took us a long time to build. We've invested in it over multiple years to be, I believe, the most efficient users of compute among any of the frontier labs. That didn't happen overnight. When we started using TPUs — I think it was the third generation we first used at scale — people thought, oh, you're crazy, everyone's using GPUs, why aren't you? We've invested very heavily to be able to use that compute flexibly, and we look across the different generations of those platforms and use each generation for the best workload internally. So we built an orchestration layer that gives us that flexibility, and in doing so we get the most value out of it.

Krishna Rao

That's part of the journey for sure, but it's also been pretty collaborative. We work really closely with the Annapurna Labs team at Amazon to help influence the road map of these chips, because we're really stressing the limits of what they're capable of. That means a dollar of compute inside our organisation goes further than I think it does anywhere else. We want to utilise each chip to its best purpose within the company. So we're building our own compilers, building things from the chip level up, to have the customisation and flexibility to use it internally the way we think will generate the most ROI.

Krishna Rao

When you're growing a business exponentially, really small movements in monthly or weekly growth rates compound into very different outcomes. It's hard to predict this business. Humans mostly think linearly and incrementally — I've been at the company two years, and that's a paradigm I've had to break for myself, to stop thinking linearly and think on the exponential. When you're on that exponential, the range of outcomes starts to be really, really wide. We look at a range of scenarios, at different points in that cone over a one- to two-year period, and we work backwards. What we want is to still be at the frontier — that's the most important thing — to be able to serve customers, and to have enough internal compute to accelerate our employees. It's interesting: if we said to our employees, you can't use our models anymore, we could serve billions of dollars of revenue with the compute we allocate to them internally. But we want to take a long-term view, because we want to range towards the top end of these outcomes — and we have to plan for that. The most important question is what happens if you're at one point in the cone but you've only bought compute for a different point. That's where compute efficiency has really helped us out.

Krishna Rao

In addition to meeting about procurement, we meet a lot about allocation. What's important is that our culture is incredibly collaborative, and that informs how the conversation happens — there aren't fiefdoms, it's not zero-sum. There's a level of compute for development that we will not go below. Even if it makes it harder to serve customers, we want to keep making that long-term investment in the best models, because the returns to frontier intelligence are extremely high, especially in enterprise. So that puts a floor on the compute allocated to model development. Internal use helps us speed up that development and find the compute-efficiency multipliers that get more from each dollar. Each team represents what they'd do with the compute, we have an open and frank discussion about ROI, and because we can allocate so dynamically, we can make adjustments on a relatively short time horizon.

Krishna Rao

A couple of ways. From a model perspective, the analogy people reach for when new models come out is cars — you had a sedan, then a higher-end sedan, and you're moving up the chain. That's true in terms of model intelligence. Where the analogy breaks down is people think, okay, I'm going from the sedan to the sports car, and I'll get much less fuel efficiency — I'm not buying the sports car for the gas mileage. In our case, we see both: huge improvements in capability, and improvements in efficiency. If you look at going from Opus 4 to 4.5, 4.6, now 4.7 — each of those leaps, not equal, but each has a multiplier in how much more efficiently it processes tokens. That doesn't just serve customers; it helps us internally, because if we're doing reinforcement learning on the model, that's basically inference within a sandbox with a reward function. If the model is better at efficient inference, that RL is more efficient too. So the customer gets more capability when we release a new model, and we can serve that model sometimes multiples more efficiently than the prior generation. Between generations, we're dynamically deploying efficiency improvements. It's always getting more efficient over time, and what fuels that is the research team.

Krishna Rao

Every time we have a new model, there's a set of capabilities that are different. People tend to think about model intelligence as IQ — a single number, 110 to 125. We think of it differently. Intelligence for us is multi-dimensional; it's not just a score. We find a lot of the published benchmarks are saturated. We publish them too, but our real measurement is what customers tell us: what is the real-world capability of this model? As we've released better models, it's not just outright intelligence — it's the ability to do long-horizon tasks, to use tools and computer use, to do agentic tasks faster. If you have two employees who are equally capable, but one takes a week and one takes a day, that second person can be seven times better. All of that factors into how customers experience it. And what we've found very consistently is that by releasing new models, the TAM gets unlocked in a unique way — more use cases become possible.

Krishna Rao

A good illustration is the last four months. We started the year with about $9 billion of run-rate revenue, and we ended the quarter north of $30 billion of run-rate revenue. That kind of change is enabled by these model-intelligence leaps and the products we build around them. I think that's unique to enterprise — in consumer you don't always see it as readily. Our enterprise customers are always pushing the limits. It started with coding, but it's expanded well beyond that. Each model generation gives you the chance to do more, better, more efficiently, and customers invest heavily in more tokens with the newer models. We've seen that cycle play out again and again. That's a core thesis of our business: especially in enterprise, the returns to frontier intelligence are not slowing down.

Krishna Rao

We do see progress accelerating. I can't speak for other companies, but for us the scaling laws are alive and well. Right now, 90-plus percent of our code is actually written by Claude Code, and a lot of Claude Code's own code is written by Claude Code. That's part of why we allocate compute internally, why we'd forgo revenue for it — the models themselves are helping us build the next generation of models. On top of the capability leap you get from scaling laws, talent is really important, and talent with the best models can accelerate development. We don't think about models as closed or open; we think of them as frontier or not. The ones at the frontier are capturing the economic value and driving real ROI, so we invest behind that thesis — both compute and the talent to use it. The other piece is the products built on top: we had 30 different product and feature releases in January. That pace is enabled by using the models with the talent we have.

Krishna Rao

The core of our company is still a research lab — maybe not as well understood from the outside. We're doing experiments that push the limits of what our models can do, and that engine is upstream of everything else. It's enabled by the models today, but it's not entirely done by the models. Over time they'll get better and more helpful in that process, but having the best talent to set the direction — not just priorities, but the new areas of discovery — actually makes that research talent even better. We talk a lot about how talent density beats talent mass. We want the densest collection of AI research and inference-engineering talent, and that, enabled with the best models, is a winning combination.

Krishna Rao

We think about the world as scenarios. It's very hard to have a point estimate in this business, so we keep a very low bar for updating our priors. Something that was true a month ago might just not be true today, and that breaks your model. The old approach — we'll forecast once a quarter and revisit at the next board meeting — doesn't work. It's so dynamic that we always have to ask: our models couldn't do this before and they can now, so what does that mean for the TAM? We saw this in coding first. Around Sonnet 3.5, 3.6, we started to see a remarkable jump in capability, followed by adoption, usage, and revenue. Now we can use coding as an analogue for a lot of what's happening elsewhere in the economy and in our business. We look at pattern recognition in our own business to predict what's going to happen.

Krishna Rao

We announced a partnership for the Colossus facility in Memphis. We're really excited — it'll let us keep expanding, especially on the consumer and prosumer side. But that's just one example of us looking for near-term compute wherever we can get it. As the compute base grows, that near-term compute becomes a smaller fraction of what's available. We look at whether we can deploy it productively — sometimes yes, sometimes no. If we can, we look at the economic return based on price, duration, location, what type of compute it is, and how efficiently we can run it. We use that same process to assess longer-term deals too. Last month we signed a five-gigawatt deal with Google and Broadcom for TPUs, starting in 2027. We also signed a deal with Amazon for Trainium for up to five gigawatts. It was an over-$100-billion commitment, and a lot of that compute is already landing and will land through the rest of this year and into next. It's a bit of a layer cake of compute starting at different times with different capabilities, and we're very dynamically comparing its price-performance over time.

Krishna Rao

Across three chip platforms we also have multiple generations — TPU v5e, v6, v7, Trainium 2, Trainium 3 — all at different places on the price-performance curve. Then we look at how we'll utilise each. Price-performance matters because of efficiency; speed matters for certain use cases. So we look at compute down to a very granular level: what it can deliver and when. Our compute team leads that, but we collaborate closely across the business to say where we need it and for what. We might need certain chips for RL, or more leading-edge compute for our best and fastest models or for training them. It's customer demand, but it's also very granular in terms of what each chip is best for and what we'll have when.

Krishna Rao

This goes back to fungibility. We're constrained across those use cases internally today. A year or two ago it would have been harder to consume a heterogeneous compute drop quickly, because these platforms are different — some are harder to operate, some have idiosyncrasies. Today, getting a lot more compute, I think it would be deployed very rapidly across those use cases, probably with the same allocation we do today. It's become a lot easier for us to spin up quickly and deploy almost any type of compute, and that's a real advantage.

Krishna Rao

Most of what we're building is platform. There are so many examples where a platform can accrue a lot of value, but the customers building on it create even more. It's maybe akin to the early days of AWS. If you think about the Claude platform and all the tools and services built into it — it's not just raw model access, it's prompt caching, the ability to use virtual machines, Claude Code being called within it, the Claude Agent SDK, managed agents. All of these are vectors to access that model intelligence for other companies to build into their own products. That's most of where we think the business is going. That said, we'll build our own applications on that same platform where a couple of things are true. One, if we have a vision into where the models are going and can demonstrate customer value there — that might be something like Claude Code. A lot of what was out there was developer-led; Claude Code is Claude-led. When it launched a little over a year ago the models couldn't quite do that, but we believed they'd get there, and they have. Two, demonstrating value for the ecosystem that others might emulate — Claude for financial services, Claude for life sciences, Claude security. We're building on the same platform as our customers, which creates a level playing field, and we've done these in a collaborative, partnership-oriented way. So I think of our strategy as mostly horizontal; we'll build vertical where we have something to add. A lot of the value is going to accrue to the customers building on top.

Krishna Rao

Part of what's hard in this business is that it changes so quickly. The model capabilities sometimes even surprise us. When we release models or products, there's an element of what happened over five, ten, twenty years in prior waves happening in months now. People are surprised by it in the same way we were surprised. But fundamentally we try to be very partner-oriented toward the ecosystem — early-access programmes, working closely with customers, listening to the capabilities they want. That doesn't mean releases aren't sometimes moments where you think, wow, that's more powerful than I thought. Part of that is the reality of where we are in this cycle, but our approach is to make those capabilities really accessible, which should accrue a lot of value to customers too.

Krishna Rao

The company is only a little over five years old. This past March was the third anniversary of the first dollar of revenue, and we only had a frontier model for real for the first time in March of 2024. Our pricing has been relatively stable across Haiku, Sonnet, and Opus. Mythos is newer. But we've made very few pricing changes. The biggest one was bringing down the price of the Opus family when we launched Opus 4.5. We found Opus-class models were underutilised relative to their capability — people were trying to fit an Opus problem into a Sonnet workload. Because of efficiency improvements, we could serve it very efficiently and still bring the price down, which made it more accessible. We're in the very early innings on these use cases, and the best way to proliferate this intelligence — from startups to the largest companies in the world — is to make it accessible enough that they get a lot of value. When we changed the Opus pricing, you saw a Jevons paradox: we lowered the price, but consumption went up way more than you'd expect. Then when we released Opus 4.6, they could slot it in — we didn't change the price. Pricing stability matters, and pricing to get that value and see that Jevons paradox matters too.

Krishna Rao

We think about the return on our compute spend writ large — all the different workloads, whether serving customers or model development, all in support of revenue over different time scales. If I serve inference, it's in support of revenue today; if I do model development, it might unlock TAM that drives revenue six months from now, and everything in between. Our returns on that compute expense today are robust, and we think of it as the return on the full envelope of compute. When revenue grew in Q1, it's not like we onboarded a bunch of new compute in that period — compute comes on a ramp that might have been determined twelve months ago. So the idea of a variable cost on the incremental to serve a customer doesn't really fit our business; it tries to force us into a software paradigm. In actuality, compute supports all of these activities, we're generating a robust return on it, and that's our measuring stick.

Krishna Rao

We're fortunate to have great partners in Amazon, Google, and Microsoft, and also Broadcom and Nvidia. We're the only model on all three clouds today, and the only language lab using all three of these chip platforms. These collaborations are much deeper than just procurement — that's often lost. Our teams are deeply embedded with the Annapurna Labs team at Amazon; we're really good users of Trainium, we plan capacity together, and the three clouds are also great distribution engines for us — though we have a robust first-party business too. These are multi-faceted partnerships: developing the chips, landing the capacity, serving it, and ultimately distributing to customers.

Krishna Rao

This is really interesting, because we were using Claude Code about a year ago — I started asking, is everyone just live-coding? We started using it as an assistant, a digital co-worker, not just for coding tasks. That was early in what eventually became Cowork — an extension of Claude Code to say that what it did for agentic software development, it should do for all of knowledge work. Today, for all of our legal entities, we can produce the statutory financial statements using Claude. A human checks them, but all of those financial statements are produced with Claude. We also have a more real-time platform called Ant Stats. It used to take a lot of time to sift through the data, get to conclusions, write a memo. We now have a library of skills for Claude specific to finance — last I checked, 70 of them — that everyone can access through a common repository. On top of that we built a monthly-financial-review skill. It produces our MFR 90 to 95% ready, and then all the discussion becomes about what we do and the implications — not what happened, because Claude isn't just reporting the weather; it helps think about drivers, why a number changed. What used to take hours to produce a weekly report now comes down to 30 minutes, and we spend our time on the strategic implications and get it in the hands of business leaders much faster.

Krishna Rao

I have a dashboard I look at for token usage across the team. We don't compensate people on it — no one's trying to token-max — but it's really interesting, because some of the most senior people on the finance team are the biggest users of tokens. It's not just the 22-year-old who joined with a coding background. Our number-one user is our head of tax; he's focused on tax-policy engines and automating large parts of the team's workloads. I tell people, if we're not super-users of this, if we're not pushing its limits, how can we expect customers to?

Krishna Rao

I have a slightly different view. It's made our already talented people so much more productive. There's a bit of a Jevons paradox but for labour: people become more productive, and we've actually hired a lot more people because of it, because there's no shortage of work to do. Now, with Claude, people spend less time reconciling some number in the MFR and more time thinking about how to reinvest in the business, how to dynamically allocate resources. Before, I'm working to tie out a number, or taking a long time to close the books. So I think of it more optimistically — as an accelerant to our productivity that lets us get a lot more done even as we grow the team, and that's starting to be true across many companies.

Krishna Rao

I joined about two years ago; we were closing our Series D at the time. That was not a straightforward fundraising — the company really only had a frontier model in the middle of it. Towards the tail end, the FTX transaction was happening, liquidating a bunch of Anthropic shares. The questions were: why do you need a frontier model, what are the returns? And around our mission — people said, aren't AI safety and building a really big business at odds? There were misconceptions: your sales force is really small, don't you need to scale it like all these enterprise software companies? People were trying to fit us into a mould that existed before. At the end of 2024 we raised the Series E — the business had scaled to close to a billion dollars of run-rate revenue — but the day of our first close was the day the DeepSeek news came out. We got the close done, but there was a ton of volatility as people asked, should I totally re-underwrite how I think about AI? They looked at our forecast and thought, okay, you've grown to a billion of run-rate revenue quickly, but there's no way you keep it up — laws of physics — and the enterprise adoption will take so much longer; look how long cloud took, how many people are still on-prem.

Krishna Rao

The business kept proving out the thesis that the return to frontier intelligence is really high — model-led growth, enabled by products and our go-to-market and distribution. And they saw this interesting interlink most people didn't understand or believe: we invest in research not just in model development but in AI safety research. We pioneered interpretability — think of it as an MRI for the model, to see inside the neural network. We pioneered alignment science — you want the model to do what you tell it, and you measure how often it strays. We did those things for our mission, but they had downstream effects: if you can look inside the model, you're better at building them. And the last linkage — if you're selling to enterprises, and we now sell to nine of the Fortune 10, those enterprises are entrusting us with customer information and data, and their most sensitive workflows. Our investment in safety, interpretability, and alignment inures to the benefit of those customers, because if they're going to entrust us with all that access, they want a company they can trust. That's not why we invested in it, but it's proven out again and again. We've raised about $75 billion since I joined, with another $50 billion to come from the Amazon and Google deals we closed last month. It's a tremendous amount of capital, but it's a capital-intensive business — and we raise it to support growth, not to fund losses in the business today.

Krishna Rao

When I joined, the business had about $250 million of run-rate revenue, and the plan was to get to a billion. I said, great — in what year? That was linear thinking. Consistently, Dario has been a much better predictor of the revenue than I have. The first time I saw it, you have all these arguments about the laws of physics and the law of large numbers — where is the revenue coming from, how can customers move this quickly, is this even possible in enterprise? All of those break down over time as you see how the business works internally, how the adoption curves and the underlying exponentials support the revenue exponential. That doesn't mean we're not disciplined about the forecast, but my thinking has shifted a lot from linear and incremental toward leaning into the exponential.

Krishna Rao

It's this paradigm of how compute is used. Not as a variable cost over some period, but as a resource so fungibly utilised. We run a chip for inference in the morning and use it for model development in the afternoon and evening. That paradigm doesn't exist in a software company or a factory. If you have a bunch of people doing R&D, they can't become cost of goods sold, and vice versa. Here you really have that fungibility, which is why the return on compute is so important. People are beginning to understand it, but there's still a tendency to treat it as two separate costs when in actuality they're self-reinforcing, and that flexibility is what helps drive revenue short-term and long-term.

Krishna Rao

First: what is the ROI on compute, all up? How are you utilising it, what return are you seeing today, and what's the shape of it over time? These are massive, unprecedented investments. Second: how do your customers see ROI? Are they just testing, or actually deploying at meaningful scale? For our business we see that in spades — our net-dollar retention rate is over 500% on an annualised basis, and with nine of the Fortune 10, these are real customers making significant buying decisions.

Krishna Rao

Exactly. On the way here I was in an Uber, and I signed two double-digit-million-dollar commits in a 20-minute car ride. We're being judged by some of the biggest companies in the world, the most sophisticated buyers, and startups that have choice and are choosing us. A third question I'd ask from the skeptical investor seat is: how do you think about compute in the future and where it comes from? Some of the places we buy from also sell to others and use it internally, so what's the balance over time?

Krishna Rao

That's right. That's right.

Krishna Rao

There have been other transformative waves — the Industrial Revolution, the internet, cloud. One thing that's different about AI is that it's all happening so quickly; years or decades of progress compressed into months. Going back to humans thinking linearly versus exponentially, that can be jarring. We're very optimistic about the potential — Dario wrote this essay, Machines of Loving Grace, about how this technology can transform how we live: drug development, curing mainstream and rare diseases, how healthcare is delivered, raising the standard of living in the developing world. We could do a better job painting that picture, and show more tangible results over time. But we also want to articulate the risks — I don't think we should tell everyone everything's going to be great, because there are likely to be bumps. People gravitate toward honest and balanced assessments; if someone's telling me only the good news, do I trust that perspective? So it's about clearly articulating the opportunities, thinking about the solutions — which no single company has the blueprint for — and being transparent about both. Over the long term the opportunity is far greater than the risks, but that doesn't mean it'll be perfectly smooth.

Krishna Rao

People maybe misconstrued Mythos as just a cyber model. It's an incredibly capable model across many dimensions, but we found cyber in particular was a place where it spiked. This was the first model we decided to release in a different way, consistent with our mission and principles. We took a phased approach, because when a model is this capable — cyber is what people focused on, but there are other things too — it can be used positively, to patch code bases. We had an open-source code base where a prior model found 22 security vulnerabilities and Mythos found 250. That's kind of scary, but it informed how we released it. We didn't say we'd never release it; we said, let's do it in a phased way, to a group that expands over time, focused on how that one cyber capability can be used defensively rather than offensively. We think that's a template for the future.

Krishna Rao

First and foremost, we prioritise a strong relationship, because regulation has a role to play in how these models are developed over time. We're very America-first in our approach — we want the technology to support the US and democratic countries around the world, which is one reason we've worked closely with the administration on something like Mythos. There's a balance: you want innovation to happen quickly and not be slowed down, but you also want a responsibility framework for how these things are deployed. We've long said this technology has implications and we should have an honest conversation about them, including with the government. The Mythos process is a good example of that.

Krishna Rao

The culture is a genuinely unique aspect of Anthropic, and it's different when you're living it. A few observations. First, we have seven co-founders — that shouldn't work on paper, but it does in practice, and they've set the example for what matters. We do a culture interview, and it's not pro forma — it's a real part of the evaluation. Someone could be the smartest person you've met for a role, and we won't hire them if they don't pass the culture bar. How would I describe it? One, incredibly collaborative — we don't tolerate fiefdoms, sharp elbows, or needing to take credit. It's humble; our competitors are incredibly capable and success is far from guaranteed. If we reach a milestone, there's no confetti on the floor — it's, what's next? Two, rigorous debate — intellectual openness and honesty, where people question things and express a point of view, but the dialogue is productive, and after we decide, there's real alignment. In something like compute allocation, people might have different perspectives, but once we come to a decision there's no second-guessing, no politics. Three, remarkably transparent: Dario gets up in front of the company every two weeks, usually writes a short document, talks about three or four topics, and takes open questions — not softballs, not planted. It's not a decision-making forum, but a window into how leadership is thinking. All seven co-founders are still at the company, and the vast majority of the first 20 to 30 employees are too. The culture underpins why we've attracted and retained some of the best talent — we don't always pay the most. When Meta and others were out with huge packages for technical talent across the labs, I think we lost two people. Other labs lost dozens.

Krishna Rao

It really is underpinned by the culture — and that's empirical, not just a feeling. When you talk to people, it's: I want to have the most impact possible, I want to work where talent density matters more than talent mass, in a place that's actually collaborative rather than fighting for one thing that wasn't debated in the right way. Most of our team just wants to do really good work, and they're attracted to the mission — developing this transformative technology in a responsible way. We have this concept of a race to the top: we don't always have the right answers, but we want others to look at what we do and emulate some of it, so the technology is developed in a better way across the industry.

Krishna Rao

Because we're focused on enterprise and on changing the productivity of knowledge work, it's this idea of a virtual collaborator. Think of something that has context within your organisation, can use all the tools specific to you — homegrown or purchased — has memory and the ability to learn from your mistakes and its own over time, and can work over a very long time horizon on not just a task but an idea. That means model capability has to keep growing, and the products we build on top can unlock it. But you have to get it in the right form factor — intelligence is multiple things, and the virtual collaborator combines them: not just generically smart, but smart for your use cases. What we're seeing in coding is what we expect to see elsewhere. Claude Code has led the way, but then something like Cowork comes along — Cowork is growing faster than Claude Code was if you index them to the same point. That's remarkable, because developers are fast adopters. Even our own product development today isn't one product manager with two engineers shipping over three months; it's shipping daily with a fleet of agents working across the company. Everyone becomes a manager. We're very early in that, but the potential is incredible.

Krishna Rao

It's really hard. The important thing is to think in first principles — everyone has priors when they come to something new. I spent a lot of time with Tom Brown, our chief compute officer, who was one of the first people to interview me. Before I started, we went on a walk around the Mission in San Francisco for two and a half hours, in early 2024, and he told me his vision for the future of the company. Honestly, it sounded crazy. I came home and told my wife, this is going to be wild — if even 10% of that is true, it's going to bend all paradigms. A lot of what Tom said has come to fruition. The other piece is hiring great people. I tell people in interviews, I'm not hiring you as a direct report, I'm hiring you as a partner — there might be things we disagree on, and I want to hear that. We've hired people from the best companies in the world, from hyperscalers, large software, financial services. In another lifetime I worked at Blackstone in private equity; that training in thinking at a granular level is really valuable. I'm not comfortable at 50,000 feet — that's not me — but you can't be at 500 feet on everything in this business; there's too much surface area. So having partners is critical. And I look for analogs to things that have happened before — I helped lead the financing Airbnb did in the middle of the pandemic, when the business lost 70% of its revenue in seven weeks. A harrowing time, without precedent, where you had to think clearly when things were changing rapidly and there was no good template. On a personal level, it's hard to balance family, friends, and this job. But maybe once a week, in a quiet moment, I just think: wow, this is really cool. It's an incredible opportunity to work with this group of people on this problem at this moment in time.

Krishna Rao

We talked a lot about the scale of the compute infrastructure and what models could do in a short amount of time. He described a world I'd have called sci-fi, but a lot of what we're experiencing here and now has roots in that conversation — and there were even more things beyond where we are today. The commonality was that everything is going to happen much quicker than we think, and both the implications and the capabilities can change. He also had an incredible optimism about the future — we talk internally about holding light and shade. I came away with a bunch of questions, but also a real sense of positivity.

Krishna Rao

The first thing would be the diffusion rate within our customers. The use cases are playing catch-up to the model capability, and we're talking about humans in large organisations with practices they've had for a long time — change is hard. If that diffusion hits a wall or slows down, that could affect the rate of revenue growth. Second, the scaling laws slowing down or not holding — we don't see that, and we can't say it with 100% certainty, but the model capabilities levelling off would be another thing. Third is how we think about being at the frontier. Today we're at the frontier — I think we're defining the frontier of agentic AI — and we need to stay there. It's a competitive market, and we'll keep investing in the technology, the compute, and the go-to-market required, but that's not guaranteed either.

Krishna Rao

The biotechnology and healthcare outcomes are what I'm most optimistic about. We may live in a world where you're diagnosed with a disease that isn't curable, but within your lifetime that cure is found much more rapidly, and you might not die of it. A lot of what we do today helps speed up the drug-development process — the paperwork, clinical study reports. I'm most excited about when it goes further back into drug discovery, because molecules and proteins are so complex that small changes have big implications — AI is perfect for that. If a lab's throughput goes up 10x or 100x and we can run that many more experiments, you get better results faster, and it doesn't have to be limited to a small set of diseases. That has the potential to greatly alter the way we live.

Krishna Rao

I have a brother five and a half years older than me. We lived in California when he went to college. He got into everywhere he applied and was going to go to medical school after. I didn't know any of this at the time — I had to pull it out of him years later — but we were solidly middle class, this was 25, 30 years ago, and the financial aid packages weren't as robust as today. A big factor in his decision to go to an in-state college was wanting to give me the opportunity to go wherever I wanted, even though that was six years out and who knew how I'd turn out. Twelve- or thirteen-year-old me would never have understood it, but many years later, that's something incredibly kind that I still hold with me.

Krishna Rao

Thanks for having me, Patrick. Really enjoyed it.