Box CEO: AI Agents Aren't Going Away, We Need More Memory, Tokens Are Here To Stay

Hey, thanks for having me. Appreciate it. And congrats to the Knicks. Great work.

I am a Warriors fan. I mean, I'm probably a fair weather fan of anything, but I'd say my default is Warriors. Yes.

We definitely have more than our share of trophies, I think.

That's a very big question. So first of all, easily the most exciting and dynamic period I've ever seen as a player in software. There are a few different dynamics going on in the space. Clearly, there's the infrastructure buildout which has its own trajectories, the semiconductor and memory stack and core infrastructure players. Then you have the vector of what's the future of the software stack, with a lot of dynamic conversation around where software gets compressed and where it expands. And then you have the AI labs and the model layer. All three are interdependent components. The more that the cost of infrastructure and the chip, the less margin there is for the other two. The more that there's open source at the AI lab level, the more value goes to the software and routing layer of AI. So these are three interdependent ecosystems where as any one player moves, the rest has to adapt. But easily the most exciting time because of the rate of progress and innovation and what it means to everybody's business model in the space.

Yeah. So if you think about an agent, let's separate two types of agents. There's agents for engineering work, where the primary data that agent needs is within your codebase. That's the predominant set of data it will work with. Then there's the rest of knowledge work, which inevitably will be 90 to 95% of all agents. The data those agents work with will include CRM data, email, ERP systems, but the bulk is unstructured data in the form of files and enterprise content: contracts, marketing assets, research materials, financial documents, HR records. They'll also store their own data as agentic memory. All that information needs to be stored somewhere, accessed by agents, and incorporated into workflows that both agents and users can interact with. Agents effectively need a file system to store and access data. Importantly, that file system needs to be the same one that human users interact with, because I need to get the agent data I want it to work on. For us, this is a huge opportunity because we've already built a file system for humans and for applications. Now we have a third constituent: agents. The reason I think some of the narrative on SaaS and software is probably wrong in some areas is that agents will be the number one user of software in the future, at a scale of 100x more than people. That creates a huge value proposition for the layer of software that provides guardrails, data access, and workflows for agents. We're trying to build that platform for agents to work with enterprise content and data. There will be other plays: platforms for agents to work with structured data like Palantir or Snowflake, agents that work with CRM data from Salesforce or HR data from Workday. What we want to do extremely well is be a platform that works with your enterprise corporate unstructured data, all the stuff that goes into your files, and connect both Box agents and external agents in a secure way.

So let me add one clarification and then we'll get to the model piece. The only part to clarify is that this is not a case for or against the entire category of software. You do have to be somewhat specific about which players will see more demand and which might see more pressure. I'll give one small example. There's been a lot of commentary on Atlassian. Atlassian is a well-run company that powers engineering workflows across most large enterprises. There was a narrative that if agents are writing all the software, maybe they don't need the same underlying systems. But if you look at their latest earnings, not only has their growth continued to be very strong, they actually find that the companies using agents the most are growing the fastest on their platform. So it's not that agents replace engineers and you need fewer Atlassian seats. Agents need those same guardrails, and if agents produce a lot more code, the infrastructure to manage that code becomes really important. We're seeing the same thing in our business. If agents are reading, reviewing, and writing contracts or marketing assets, then the platforms that protect, secure, and govern that data and allow agents to operate within secure guardrails and workflows, the value of that system goes up. We've actually accelerated our growth rate as a result of AI. To your point about the models, this is a really interesting dynamic. First of all, most companies don't want to standardize on one AI model. We've already seen multiple examples of why that is the case. Sometimes models cost too much, sometimes there's a government action that reduces access, sometimes different models have different performance capabilities. In a world where an enterprise wants flexibility around which models they deploy, value accrues to the layer that is completely agnostic to which model you deploy. If I'm an enterprise constructing agent workflows, do I want to be stuck with one model vendor or do I want neutrality between the layer that manages context and workflows and the place where I bring in intelligence? Model routing is going to become a very big conversation. Brian Armstrong at Coinbase wrote about this: a frontier model versus an open-source model that is almost at frontier capability might have a 100x difference in cost. Even if it's a 1:10 ratio, that's a severe enough cost difference where I want something in between me and the model that can decide whether I need frontier performance or something lower cost. This balances back in favor of the applied layer, like Snowflake, Palantir, Box, or Atlassian, the layer that can route the task to the most effective and optimized model for that job, whether it's cost or performance. So we think the labs will do fantastically well, but there's another role for the applied layer of AI to give customers buffer from being stuck with one model provider or cost structure.

Yeah, I mean, the essence of your question is way above my pay grade because you brought in the trade element, but I would say you're going to need much of the supply chain constraints we read about and the dynamic between Western Digital, Seagate, SanDisk, Micron, and all these various players. First of all, it's great that that layer of the stack has gotten its time in the spotlight. These are companies that for decades have been toiling deep in the infrastructure world. I can't tell you how to trade those companies because there are supply chain dynamics that could reverse and PE multiples that I don't know exactly how the market wants to think about. But what I do know is that the demand for almost all layers of the memory, infrastructure, storage, and compute stack are only going to continue to go up and to the right. You've already seen this with the history of the PC and mobile phone. New technologies are always introduced to replace the old thing, but today we use one of everything. In my average day, I've touched an iPad, an iPhone, and my MacBook. Even the CPU is on fire because right next to my inference for the non-deterministic part of the agent's work, I need a compute sandbox that it can operate in, riding on a CPU. So every part of this stack is going up and to the right. What is the trade? Again, way above my pay grade, but the demand for this infrastructure is not slowing down. We are going to need hundreds or thousands more data centers worth of compute to satisfy the potential of what these agents will do. This is why the SpaceX play and orbital data centers are quite germane to this conversation because we're going to literally run out of energy from a provisioning standpoint, and you have these new interesting dynamics playing out.