The Exchanges

Q I do want to finish, though, on what the next five years hold for you, and what does the next chapter of success look like? We have this interview in 2026. Where do you want to be then?

A I personally think we're at one of the most exciting points of the web, and it's Direction going forward that I've ever seen. So if you think about how much, you know, compute we have to work with, how many new amazing APIs we have to work with, how much the internet is becoming more real time, more collaborative, more multimedia driven. Software is going to be more multiplayer and more real time. So I think in three years or four years or five years from now, I think we can expect more immersive applications, better collaboration, being able to do completely new things with our data. Maybe some of that will be driven by AI. Maybe if it Some of it will be just driven by completely new formats that we work with, and our job at Box is we just want to be a platform that helps people work with their content in that future state, and we're going to continue to stay focused a hundred percent on that and not deviate from the strategy.

AI assessment note: “our job at Box is we just want to be a platform that helps people”

Q Who is going to win the enterprise race, OpenAI or Anthropic?

A Oh god, that, that's impossible. Back to the cloud piece, I think, um, you know, I think it's, it's totally fair to think about it as a race, and certainly if you're, if you're in either of those companies, you have to treat it like a race, because, because, you know, like, you obviously want 80% market share, not 55% market share. So, like, you have to treat this as a, you know, we gotta dominate, uh, that's exactly how, they, they should be executing that way, everything is going according to plan. If you compare it to other areas of compute, Um, and I, I ran this analysis recently. In 2010. 2010. Not, like, you know, maybe you were 12, but, like, the rest of us, we were just, like, in companies doing things. Uh, in 2010, AWS made five hundred million dollars in revenue. Azure had just launched, and GP, and GCP was called Google App Engine, and it had a little, like, a turbine logo, uh, with, like, wings or something. So, that was the state of cloud. Fast forward to this year and it's a couple hundred billion dollar a year revenue ecosystem, right? So, so in 15 years, right? So, and we were in that moment being like, who's going to win? AWS or Azure or GCP? What's, how is this all going to play out? And it just turns out the market was so large, like obviously it was due to their execution that they kept it going and kept it large and the competition kept up, but it, it just …

AI assessment note: “everybody kind of won. And so I, I sort of think of AI in a similar fashion”

Q Do you think Google smashed it last week?

A I do. Yeah. But last week, even to me could be kind of like condensed as just like another week of awesome AI updates. Um, I'm not sure that, that, that I, I'm not sure that I can sort of pull out like, like this is the one thing that will set off a new trajectory of AI. Like, like, You know, there were agents, there were better models, there was, you know, multimodal, like all great stuff, but what, but the, the thing that I think it underscores that I think honestly, uh, uh, cloud next did as well is like, you know, Sundar has said a message to the company that AI is the number one, most important priority of all things. It's like 10 times higher than every other level of everything else you're working on. It's gonna be like the new way to search. It's gonna be the new products in cloud. It's gonna, you know, It's going to be in workspace. And I think that message is, it was very important. It is now very clearly like the evidence of it working is now showing up. And, um, and I think, I think, you know, Google IO is merely the, the, you know, another moment of it manifesting that, that, that, that company has religion now, uh, on AI.

AI assessment note: “I do. Yeah. But last week, even to me could be kind of like condensed”

Q So does this completely discard like Moore's law? Cause as you said, we've never seen that before and it's completely atypical compared to previous progressions.

A To be fair to maybe like a Moore's law, like religious, you know, you know, zealot, um, this is a sort of a different, uh, a different variable than, than what Moore's law is, uh, which is about the density of, of, of, uh, of your chip or the performance of your chip. So So this is more about model algorithms and, and how we're actually utilizing these models. There's actually a different law, which is sort of the GPU performance. Um, and I think there's a sort of a, you know, Jensen almost kind of created a law around this, but we have seen faster than Moore's law improvement on GPU performance. So, so the good news is we get a little bit of a reset on, you know, there's been a lot of talk on, hey, is Moore's law slowing down? Um, and GPUs kind of are, are giving us that That, you know, maybe next tailwind of, of compute performance improvement, um, which is this, you know, miracle that, that, uh, that we now get to, to, uh, to deal with. But, um, but I would say I, I'm seeing no signs of AI models, slow down, anything plateau in terms of the innovation curves, uh, that, uh, that we're seeing.

AI assessment note: “this is a sort of a different, uh, a different variable than, than what Moore's law is”

Q You know, when you look at their releases last week, you're looking at language learning cheaters and going, fuck, did you a lingo have a business anymore? And does language learning actually have an independent category of open AI is able to provide such quality.

A Open AI is kind of telling us what they are going to become, which is. ChatGPT is going to be this universal sort of assistant interaction, you know, interface, and all the, you know, subsequent tools to, to manage that and interact with, with different AIs that kind of come together there. And then they're going to have an API business that, that will be, you know, you can almost just, you know, very quickly, easily understand. It's going to be audio, video, text, like it'll, like, it's going to do all the formats of, of information with, you know, kind of complete intelligence on them. So then you're sitting around, you're like, well, what startup should I go build? You probably don't want to do things that, that instantly could be subsumed by a horizontal chat interface. And you probably don't want to do things in the model area that you might, might just be one training run away on their end of being subsumed by a, just a more superior model. So that means, I mean, I mean, this is sort of like, to me, it's the most exciting part of software to 80% of people. It sounds like the most boring part of software, but it's like, you have to do like the, the workflows That, that eventually a human who wants to go and run a full business process has to implement that doesn't want to just do back and forth chatting with a thing. So like to even to your tutor point, um, uh, or, or the,…

AI assessment note: “I don't think ChatGPT itself is going to become a language tutor.”

Q Uh, exceptional man, but I wanted to start, and we were just chatting, I was running around the park listening to the Dworkesh and Jensen episode, and I was like, I don't think Jensen came out very well. Do you agree with me that Jensen didn't come out very well from that episode?

A Um, I, I think this is like the greatest Rorschach test of all time of, of where, uh, where, where somebody is mentally on AI. Um, I, If I, so I happened to see a bunch of the tweets before I watched it. And so I was a little bit obviously biased in advance, but if I hadn't seen any of the commentary and I had just watched it, um, I would have been very confused by the, by the commentary post, uh, you know, post, uh, interview. And, and, um, to be clear, I kind of jumped to the more, uh, salacious part of, of, you know, China and that topic, but I'm, I'm almost probably 80% with Jensen. Um, and, um, and I, I, my, my sort of, uh, kind of way of thinking through the logic actually works much closer to, uh, to Jensen. Um, you know, the, the idea that we're in some kind of, you know, kind of race, existential race where a month or two of advantage is gonna, you know, change the total outcome of, of AI progress and, and what everybody does between us and China. I, I just don't agree with, I think what we are in is a commercial and economic race. Obviously with safety, you know, built into that, there's no question. Um, and I think we actually have a lot more power globally, uh, if it's our technology stack that's powering AI. And so I, I, I kind of am more in the camp of, of Jensen on, on his, on his lines of logic. And, you know, Dwarkech kind of oversimplified a few, a few compone…

AI assessment note: “I'm almost probably 80% with Jensen”

Q Now, I want to start with one of your tweets mentioned actually you're living through the transition, you know, to cloud. And I just thought, you know, you really had such a front row seat to that. Having the experience that you have in terms of living through that transition, how do you think about what it takes to be successful with this transition in this next wave of AI?

A We are clearly in this, this window that is, um, going to be relatively temporary. Um, uh, you know, I don't know if it'll be two years. I don't know if it'll be five years. I don't know if it's already over. But basically you have these windows every, you know, more or less once a decade. There's nothing, you know, kind of particularly specific about the decade part, but we had it in the, the PC boom. So that was basically the eighties. You had it in the web boom in the nineties. You had it in the, the mobile and cloud booms in the kind of mid 2002 1010. And each one of these moments, there's an architecture shift that happens in tech that creates Really the only window of opportunity you have for new kind of platform scale, you know, large franchise kind of companies to emerge, because you need a change in the technology industry for new insurgents to be able to enter, or else basically the incumbents will, will just naturally gobble up all over the market. So we are in at one of these moments with AI, which is a, a, a period where we are going to see not only breakthrough technology, but the breakthrough application of those technologies. That is, um, as much going to be an incumbent's game as a startup's game this time around, because the incumbents have a lot of the data and a lot of the workflows, so it's even, it's even sort of more, uh, competitive, I think, than the pr…

AI assessment note: “startups will emerge that will be able to figure out a set of use cases”

Q So does this completely discard like Moore's law? Cause as you said, we've never seen that before and it's completely atypical compared to previous progressions.

A To be fair to maybe like a Moore's law, like religious, you know, you know, zealot, um, this is a sort of a different, uh, a different variable than, than what Moore's law is, uh, which is about the density of, of, of, uh, of your chip or the performance of your chip. So So this is more about model algorithms and, and how we're actually utilizing these models. There's actually a different law, which is sort of the GPU performance. Um, and I think there's a sort of a, you know, Jensen almost kind of created a law around this, but we have seen faster than Moore's law improvement on GPU performance. So, so the good news is we get a little bit of a reset on, you know, there's been a lot of talk on, hey, is Moore's law slowing down? Um, and GPUs kind of are, are giving us that That, you know, maybe next tailwind of, of compute performance improvement, um, which is this, you know, miracle that, that, uh, that we now get to, to, uh, to deal with. But, um, but I would say I, I'm seeing no signs of AI models, slow down, anything plateau in terms of the innovation curves, uh, that, uh, that we're seeing.

AI assessment note: “this is a sort of a different variable than what Moore's law is”

Q I'm just following that lead. Jason Lemkin, my dear friend says, Why has no public company created any good agent product? Everyone creates 60% shit agents, but he's like, the one person who's done is Palantir, and no other public company has created a sufficiently good agent product. Why is that?

A Um, you know, I don't know that I can fully endorse the point, but I'll, I'll, I'll, I can give you the, because I, I would argue our agent is sort of the best agent for working with content. You know, this is a very fast moving, uh, space, And you have to be kind of wired in at a level that, that I don't think you've ever had to be wired in in tech. Um, uh, like I am, uh, you know, and, and the information sources aren't the classic ones. It's not the, it's not the roll up review two weeks later from your traditional news publication that is going to give you any kind of alpha. It's, it's, it's the practitioner who's the, you know, literally the engineer at the, you know, agent sandbox company. And their, their long form article on how they are handling, you know, memory and the harness and, you know, like, like, like if you're not wired into that ecosystem, it's very hard to then have your team, you know, be at the kind of forefront of all of what is happening. And so it just is a, it's a different pattern than what we've ever had to do. Like, like, you know, COVID was, was pretty crazy. Like we all had to Kind of like hunker down and be paying attention to daily news cycles on, on COVIDy stuff. Um, but it wasn't like a tech problem, like it wasn't hard technologically. Um, so there's like, there's, but there's not been a moment before where the speed of change and responsive…

AI assessment note: “if you're not wired into that ecosystem, it's very hard”

Q What's the biggest thing that you've lost with scale? Some people lose speed, some people lose creativity, some people lose innovation. If you were to say, no, we probably lost that. What was something that you lost at scale that you most want to get back?

A There is a premium when you're a bigger company, which is, which is when you make a decision, it needs to be a well thought out decision that you're just going to press a button and we're going to go execute and you're going to iterate and you're going to learn lots of things. But like you, because of the, because of the work it takes to drive that alignment, but it means when you do that, like you don't want to follow up a week or two weeks later and be like, ah, just kidding. Like we got to do this thing. Um, and, and so, you know, maybe, um, this is the only, you know, in the 30 seconds I had to think about your, your question, you know, when, when you're a 20 person company, you just get in a room and you're like, okay, here's what I think we should do. It looks like this. It should be priced like this. Let's go build it. Let's test it or whatever. And then, and then like the whole company is, is sort of like fully lined up to go do that thing. You find out that it sucks and you, and you just like quickly pivot. And like everybody was on the same page that no, no, we were just, this was just It's clearly a hypothesis. We, we don't really know for a fact what's gonna happen. We can just pivot through this. In a big company, you're like, that, that team still exists, but the problem is you have everybody else watching that team saying, please tell us when, like, the thing is …

AI assessment note: “when you're a 20 person company... You find out that it sucks and you just quickly pivot”

Q and you being young at the founding, My question is, I had Joe Fernandez from Joy Mode and Klaus on the show a long time ago now, actually, but he said that serial entrepreneurship was overrated, which I thought was a very interesting, quite contrarian statement, given the praise that we lord on serial entrepreneurs. Given that Box was your first company, would you agree that serial entrepreneurship's overrated? Yeah.

A Well, you know, I guess to be fair, it was my first company that maybe had, you know, some form of success, but had definitely tried lots and lots of other ideas prior. So I tend to not get too sucked into, you know, some of the maybe kind of quote unquote best Best practices or pattern recognition that I think we tend to pontificate about it. You know, there's amazing entrepreneurs that have done it once. There's amazing entrepreneurs that have done it five times, and it all comes down to the idea, the team, and do you choose the right market at the right moment? Those tend to be the factors more than whether it's your first or, or end company.

AI assessment note: “Those tend to be the factors more than whether it's your first or, or end company.”

Q Do you really think so? With the greatest of respect, we are seeing the eradication of kind of lower ranking legal positions.

A And that, that is a different issue, which is how do you do the next generation of mentorship and apprenticeship when AI does automate the maybe, ah, traditional tasks that those workers are doing? Big question. A big question facing every bank in the world, every law firm in the world, anybody who had a sort of an apprenticeship model, I don't doubt that that's a real issue. But that's different from the constraints that, that all of this work ends up resulting in that you still have not been able to automate. We had a customer conversation two weeks ago, and, and, and this is just going to sit with me forever. I mean, I always have this example. They've automated or they're working on automating patient referrals when, you know, when you want to go and see the radiologist or the, the high end doctor for whatever issue you have, they're automating that, which is awesome. So now you don't have to be on the phone for, you know, a week or whatever. Well, guess what? You can automate anything. But if it still is 18 months out before an appointment is available, what, what your ultimate constraint is still the healthcare institution and the amount of doctors we have, and actually the amount of, of, of real labor we have across those organizations. So yes, maybe, maybe you, you don't want to, you know, stake your career on being a frontline, you know, customer service service rep in…

AI assessment note: “I don't doubt that that's a real issue. But that's different from the constraints”

Q You know, when you look at their releases last week, you're looking at language learning cheaters and going, fuck, did you a lingo have a business anymore? And does language learning actually have an independent category of open AI is able to provide such quality.

A Open AI is kind of telling us what they are going to become, which is. ChatGPT is going to be this universal sort of assistant interaction, you know, interface, and all the, you know, subsequent tools to, to manage that and interact with, with different AIs that kind of come together there. And then they're going to have an API business that, that will be, you know, you can almost just, you know, very quickly, easily understand. It's going to be audio, video, text, like it'll, like, it's going to do all the formats of, of information with, you know, kind of complete intelligence on them. So then you're sitting around, you're like, well, what startup should I go build? You probably don't want to do things that, that instantly could be subsumed by a horizontal chat interface. And you probably don't want to do things in the model area that you might, might just be one training run away on their end of being subsumed by a, just a more superior model. So that means, I mean, I mean, this is sort of like, to me, it's the most exciting part of software to 80% of people. It sounds like the most boring part of software, but it's like, you have to do like the, the workflows That, that eventually a human who wants to go and run a full business process has to implement that doesn't want to just do back and forth chatting with a thing. So like to even to your tutor point, um, uh, or, or the,…

AI assessment note: “I don't think ChatGPT itself is going to become a language tutor.”

Q There's many points I want to touch on in that one. Do you think we're still in the experimental budget phase for enterprises? We mentioned there about how we can change functions and optimize them. Do you think we're still in the experimental budget phase and how will the best and the worst enterprises engage and adopt with AI?

A I think there's two types of spend happening. Um, there's probably the bulk of the dollars that you see in the headlines. I would argue are likely experiments, uh, where, you know, you'll have some kind of hit rate success rate, um, that, that happens. And then those experiments graduate into production spend. And I think we just don't have a, an accurate kind of pie graph yet of what's in the production category versus what's in the experiment category, but it would be, it would be probably too generic to say it's all experimental and it's certainly not accurate that it's all production. Um, and so the exact sort of split of those two things is, is sort of hard to diagnose at this point. Um, I've just been on the, the road. Uh, we've done maybe about a dozen or so AI events, um, uh, throughout the US, um, you know, in the past quarter, the vast majority of companies have, have a meaningful number of AI experiments happening with, with lots of different, you know, kind of areas of their business, lots of different applications. Um, uh, but the vast majority also have, you know, uh, areas that are in production already. So unfortunately it all would get lumped into the same Kind of category of AI spend by the time the CFO, you know, gets the AI bill, but we, we are seeing real production and, and lots of experimentation.

AI assessment note: “we are seeing real production and, and lots of experimentation.”

Q model, it's got like, 85% of the way there. I speak to many of the best early-stage and more mature, you know, West Coast-based companies, and they say, hey, we use frontier models to set where we can be, and then we use open-source Chinese models to get as close as we can to that frontier benchmark. Is Silicon Valley being funded by a generation of open, CCP-funded open models?

A I mean, that, that must be kind of empirically true. Um, I, uh, I, I, I don't have the same kind of like, uh, oh, that's so scary, you know, kind of element. Now, obviously, again, holding out some, some element of risk of, of some, some backdoor weights that, that can get triggered at some moment or some parameters, but like, like, like I'm not, I, I just like, that's not how I'm perceiving it, but, um, uh, but yeah. And, but also that's, Yeah, I would say that's kind of orthogonal to my point about like the best frontier model still will go and do the wrong thing. Uh, and so thus I have to be in the, I have to be in the workflow loop to make sure that I review its, its work.

AI assessment note: “that, that must be kind of empirically true.”

Q Can I ask, you mentioned the bet there between kind of web-based versus P to P based. Can you tell me about a bet that Didn't go to plan and you made that was wrong. And were there any big takeaways for you from that?

A Yeah. So I would say that there's been two kinds of things, I guess, that are sort of bets that have either been wrong or, or more alternative scenarios that I would have preferred from kind of how we executed. So some bets are just, you're betting on a particular product and for whatever reason, it doesn't work. And usually a consistent pattern. And when certain features or products don't work and often they frequently have a, an element of Us as sort of product managers and people within the company brainstorming about how amazing it would be if we could do X thing with Y technology, as opposed to thinking about what are the customer problems that really need to be solved right now by our customer base at this particular time and working backwards from the customer problems and then delivering a solution to those customer problems. So usually there's a tendency where the idea itself is, is what is so interesting as opposed to the problem is so important for customers. That's sort of category one where you, you sort of can get it wrong sometimes. And it is a very subtle thing. These aren't like binary obvious moments because you can kind of squish together different data points to make different arguments when you're actually, you know, inside doing the development. And then I'd say the other scenario where, you know, I have some, you know, times where we kind of regret certai…

AI assessment note: “So I would say that there's been two kinds of things, I guess”

Q finish on something a bit off script, but you're a phenomenal CEO. You're a public company CEO. The pressure that you have on you is intense. You're also, like, married and have a great relationship. Biggest advice on marriage when it's super, now I'm being serious, when it's super stressful, it's hard, and you also have to show up and be a great husband. What's the advice on marriage? Uh,

A It feels dangerous if I actually acknowledge the great husband, uh, piece and other, other parts that were embedded in that. Um, that, that feels like you need like a full three 60 eval. Uh, I will, uh, I'll just say from my perspective, and I'm very lucky to have, uh, an amazing wife and family and, you know, you, you are, you're in a grind in one of these roles. And, um, and so obviously having, uh, a strong support base, um, uh, you know, helps a ton. Um, we try and make time, you know, for, for the fun, you know, side of, of life, uh, as much as possible, but, uh, obviously that gets constrained in, in the kind of window that we're in, but I've been with my wife for, I don't know, 15 years or so, uh, 16 years, and so she's seen the whole, the whole grind, uh, all, all the way, and, uh, she has her own set of grind, uh, in her business, and so it's, it's just lots of fun, so.

AI assessment note: “We try and make time, you know, for, for the fun”