The really, really big increase in compute, and by extension, effectively agentic workflows, is going to be when I deploy agents—maybe in swarms, maybe just in background kind of operations at some kind of frequency—to go and do task execution for me that is, again, 100% decoupled from my ability to go and kick them off. That's just going to be an incredible amount of compute that is probably 100 or 1,000x more than what we're currently consuming in AI at the moment.
Hello everyone and welcome back to MTS. Today we are joined live with Aaron Levie. He is the co-founder and CEO of Box. And last week Aaron posted just about how much compute the next era of AI agents are going to require and what that means for the infrastructure we're building. So Aaron, welcome back to MTS.
Hey, good to be here. I want to be thoughtful on the credit. I think I quote-tweeted the CEO of Daytona about some of the trends that we're seeing, which was the math on every personal agent that Muse creates. Basically, you have—each user of Muse essentially is running a computer in the cloud. And so you can start to see, imagine billions of users of these personal agents, the amount of compute infrastructure that we're going to need. And that's obviously just one of the many, many use cases we're starting to see. And I think the really, really big use case is actually going to be on the agent swarms and all of the work that those agents are going to do. That's just going to be an incredible amount of compute that is probably 100 or 1,000x more than what we're currently consuming in AI at the moment.
Yeah, exactly. I wanted to talk about it. It was the quote that Ivan posted. He said he's building Daytona and he specifically said he calculated that Muse needs around 65,000 CPUs, 775 PB of DRAM to serve 100 million users, and that was the specific thing that you were quote-tweeting. You're talking about the magnitude, and this is just Muse. And what you shared is there's going to be a ton of other use cases. Even when it comes to just the personalized agent side of things, this is relatively new. This all happened within the last few months with like Instinct and Grokbot. Now we have Muse, we have Dots, we have all these things. What are the other trends you're seeing and why this compute infrastructure is just going to be insane and grow crazily?
Yeah. So personal agents—you can just kind of imagine that everybody will have between one and a few agents just kind of doing useful things for them in the background at almost all times. That will be a lot of computers, to be clear. There's going to be a lot of volume on that. But that will probably cap out at some point on a per-person basis, just because there's only so many kind of tasks simultaneously that the average consumer needs. I mean, there's going to be stuff around healthcare. There'll be stuff around their personal lives. But that somewhat caps out. Where the massive, you know, probably 100x increase in token and compute consumption will ultimately come from is all of the enterprise workflows and enterprise workloads. So if you think about it, most work that we do today, even with AI, but certainly pre-AI, is basically highly dependent on the amount of energy I personally can contribute to a task. My energy in terms of how much I'm thinking about something or researching something or typing something or communicating something to somebody else. That's effectively—all work in the economy is basically rate-limited in knowledge work, rate-limited by our ability to generate information or move information or produce information.
Agents are the first step into kind of being able to decouple my ability to execute with energy and the ability for just now computers to do work on top of that. But even today, most agents that we're all kind of looking at or thinking about are me sending a task to an agent, that agent basically doing some work for me to speed up three or five hours' worth of work in 10 minutes. But it's usually still rate-limited by my ability to give that agent a task. So when you're working with Claude or ChatGPT, these are obviously still—I'm sort of causing the task to go on and have the agent go work on it. The really, really big increase in compute, and by extension, effectively agentic workflows, is going to be when I deploy agents—maybe in swarms, maybe just in background kind of operations at some kind of frequency—to go and do task execution for me that is, again, 100% decoupled from my ability to go and kick them off.
So for instance, if you say, 'Hey, I want to go deploy 10,000 agents to basically go upgrade a major piece of software or go do life sciences research and do a bunch of tests on new medical experiments.' Or I want to have agents be able to get kicked off every time that some code changes in our codebase and then basically fan out to review every single change in the code and every permutation of risk that gets generated by those changes from a cybersecurity standpoint. These are actually where you're going to see just a complete nonlinear growth of AI use. Because what will happen is we will start to look at every problem through the lens of how much compute can we throw at this problem to solve it either faster or with higher efficacy, and it's going to completely change the way that we think about a lot of workflows in the economy. The average hedge fund will just basically be deploying near-infinite amounts of compute at looking through every single possible trade they could be making with every amount of research they can create. The average life sciences company is obviously going to fan out agents in parallel to do an unbelievable amount of research against every pathway that they can possibly think of. You're going to obviously have agent swarms go out and get executed in AI research itself.
There was a great tweet by Andrej Karpathy this weekend that kind of walked through the funnel of AI use cases. And on the upper end of the funnel, you basically have like non-consumption, then right below that you have, you know, 'I'm talking with a chatbot,' and then you kind of go through the funnel and you really only have probably on the order of thousands to tens of thousands of people that have even deployed some form of agent swarms on doing some kind of activity that is sort of productive. So that is how early we are in this journey. Now, it's not like everybody on the way up in the funnel are going to do agent swarms for their tasks. But you can kind of imagine at least it goes up by 100x from here. And this just gives you a little bit of a sign of how both early we are, but then ultimately how compute and infrastructure bottleneck we're going to be for most of this work in the near future.
Yeah, no, completely agree when you say how early we are, especially when it comes to things like agent swarms. I know we hear about it and hear how researchers are using it, but I'm really curious how it's actually being used at the enterprise layer. And so I want to get your perspective. I mean, with your work at Box, you're constantly talking to companies, you're constantly talking to enterprises, and you start to see trends about how they're engaging with agents. Even just what I mentioned earlier around the hype of personalized agents—that is more consumer, and it's still barely that adopted when you think about the larger context of the entire market who could potentially be adopting this. But what you alluded to earlier is the compute is likely going to come from how enterprises are adopting agents and how they're adopting swarms, how they're doing research. So what specifically are you noticing, especially when it comes to maybe the customers on Box?
Yeah. So our use cases, as you can imagine, are really about processing unstructured data at scale. So imagine if you could go deploy, you know, 10,000 or 100,000 agent executions on your contracts or your invoices or your research data. What if agents could read through all of that information, extract out the key data points and materials from that content to be able to either automate a workflow or generate new insights or connect up to some other agent that's doing some other kinds of task. So the use cases that we see are really around agents being deployed against processing all of that unstructured data at scale across healthcare, financial services, life sciences, the public sector, federal government, etc. But we already are seeing limits on—basically customers have an unsatisfiable amount of interest in being able to deploy agents on their information, and the only metric that really they're now kind of constrained by is just like what is the cost of the compute.
And then can they wire up their data and their workflows in a way that actually lets them go deploy agents at this. So as you have the cost of AI come down on a per-token or per-task basis, you will see a dramatic increase in agents being able to get deployed against these workflows. And then you do have a bunch of infrastructure changes that enterprises have to go through to modernize their data environments, to better secure their information, because as you have agent swarms working on enterprise information, obviously being able to secure and protect that data becomes incredibly important. And that will become one of the next rate limiters in agentic adoption: is do I even have the infrastructure that lets me safely deploy these swarms on my data where I know what they're doing, I know what data they've accessed, I know what work has been done, I have the ability to kind of like pull the ripcord if I don't like the work that's happening. Satya kind of alluded to this idea of—you have to almost kind of think about them as insider risk, which is at any moment they could leak your data, steal information, make the wrong decision. So they have to be quite contained and quite controlled. But once you can get that architecture right, then you actually have incredible upside on the amount of just throughput of your workflows and what you can do with these agents.
I want to shout out our sponsor, Lovable. You know, the app you've been meaning to build, the internal tool, side project, or product you'd otherwise lose a weekend to. With Lovable, you get software you can ship today—that includes an editable codebase you can inspect and change, plus two-way GitHub sync. Lovable also handles the backend and infrastructure, including managed Postgres, object storage, hosting, payments, and more. So you don't have to wire it all up yourself. And through Lovable's MCP server, you can create and deploy your project directly from the agents you already use. Turn ideas into software people love with Lovable. Lovable.dev. Now on to the episode.
So what might this infrastructure actually look like? How can you get this architecture fully right so that it works for agent employees? You're not just interacting with humans who might need access to this data, but now you have agents—and not only agents, you have swarms of agents who are all trying to interact with this infrastructure.
