Co-Founder, Chief Executive Officer & Director, BOX INC
Cisco President Jeetu Patel and Box Co-Founder and CEO Aaron Levie at Semafor Tech: First Principles
Cisco President Jeetu Patel and Box Co-Founder and CEO Aaron Levie joins Semafor tech editor Reed Albergotti at Semafor Tech: First Principles, hosted June 10, 2026 in San Francisco.
About Aaron Levie
Aaron Levie, co-founder and CEO of Box, has been speaking about the state of enterprise AI adoption, describing it as a "tale of two cities." He has said that agentic AI has taken off in software engineering, but that most CIOs still cannot make agents work safely and at scale in everyday knowledge work. Levie has argued that the gap between the two is due to factors including data access, verification, and the need to redesign workflows, stating that "the right context at the right time with the right guard rails is still a critical problem for agents to work with." He has also said that the role of the CIO is becoming more important, as they are now "providing work to your organization" for the first time. Levie has discussed the economics of AI, noting that "we're well past the point where the IT organization can hold the entire budget of AI" and that companies should re-engineer processes to get the full upside of agents rather than just augmenting existing workflows. He has described the current moment as a "commercial and economic race" and has said that the next three years will create the next wave of giants in the industry. Levie has also stated that Box is focused on building an AI platform to help companies tap into unstructured data and work with their "entire agentic ecosystem."
Transcript (59 segments)
✨ AI-enhanced transcript with speaker attribution
H
Host0:00
I'm excited to introduce our first speakers. Please welcome Aaron Levie, co-founder and CEO of Box, and G2 Patel, president and chief product officer of Cisco. I think I see them.
A
Aaron Levie0:21
Hey guys, how you doing? Great to see you.
G
G2 Patel0:23
How you doing, man?
H
Host0:24
Hey, how are you?
A
Aaron Levie0:27
All right.
G
G2 Patel0:28
All right.
H
Host0:28
What's up? It's great to see both of you. Thank you so much for doing this.
A
Aaron Levie0:33
Good to see you. It was hard to take a couple minutes away from Fable, though. I'm surprised we got all these people in one room. There are tokens to be spending right now.
H
Host0:46
Did you give it a task while you were here?
A
Aaron Levie0:48
Yeah, we got to check in on it in a couple minutes.
H
Host0:50
I really hope everyone has their agents in loops right now, because that would be really inefficient if you don't. So yes, actually let's start with Fable. That's a good...
G
G2 Patel1:00
I have a question. Does Max use AI now? Max is the son.
A
Aaron Levie1:03
Max is my son. We're working on it. Okay.
H
Host1:08
These guys, they know each other. They used to work together. That's what's going on here.
A
Aaron Levie1:12
We'll try not to...
H
Host1:12
Which I love, I love the banter.
G
G2 Patel1:15
He actually told me when I left, he's like, 'It's so great you left. I'd never work with you ever again.'
A
Aaron Levie1:20
Only because at 2 in the morning, we would be debating existential strategy. So it's nice to get a little bit of sleep.
H
Host1:26
Right? Well, now you have your agents doing that for you. But speaking of Fable, this is super interesting. So there was a huge backlash today on Fable from AI researchers who use AI coding agents to build their software, which they use to train open-source models, which then compete with Anthropic and other frontier model companies. And the allegation against Anthropic is that they are secretly sabotaging that code. The model can tell that they're doing that.
A
Aaron Levie1:56
Yeah.
H
Host1:57
That they're doing that.
A
Aaron Levie1:58
Well, it's actually not even secret.
H
Host1:59
But not even secret. But you don't know how it works, right? They're just noticing the output and complaining about it. So I'd love to know, are you concerned that there's going to be anti-competitive behavior from these companies in the guise of AI safety?
A
Aaron Levie2:13
I think I would far prefer that they didn't mitigate that in the model, because obviously AI progress is something we want as much as possible. But if you understand the DNA of Anthropic, you can totally understand why they would do that. It's anti-competitive, but it's tough to use that term because it's their product. They can choose how their product can be used to weaponize against them. I think that's a reasonable competitive dynamic if you're using somebody's product to compete with them. That's probably up to the core product to decide how that gets leveraged. The great thing is this is such a competitive market that OpenAI can easily not do that, and that would become the product that more researchers end up using. So I think these things can be worked out through the competitive forces of the market. I love the complaints that researchers have, and I also understand Anthropic's position, and I think these things work themselves out in the battle of the market.
H
Host3:32
Before you became president, you were running security products for Cisco. You're still in that purview. You're also a big proponent of open source. So how do you feel about what's happening in that space?
G
G2 Patel3:46
I think America needs to have a very strong open source strategy. In the absence of that, even if there's no technical limitation of using a proprietary model within your environment, there will be restrictions. You need a strong investment in open source in addition to proprietary models. Nvidia is doing a good job with Nemo and their portfolio. In the past 96 hours, a lot happened: new image models, new video models, new language models came out. I think that will be self-propelling because there's enough demand. Now, the costs of tokens are far higher than the actual value these tokens generate at scale. The big risk is if you don't create an equilibrium, people pull back on using tokens, which is not good for anyone. You need an intelligent routing layer and small language models. Not every query needs to go through a frontier model with 10 trillion parameters. As you redirect to small language models, open source will be used. Token economics will dictate that open source is very important, good for the national interest of America, and good for corporate customers to have choice.
H
Host5:32
Aaron says it's fine for Anthropic to defend their turf, but we know that's not what they're going to say if they get a call from regulators. They're going to say, 'We're doing this because it's safe. We're worried about AI safety, and this model's too dangerous. It's going to create a bunch of security flaws.' Are you worried? Now, Mythos already started this debate, but is it that much of a step change in security if they allow this to run free?
G
G2 Patel5:58
We were in Project Glasswing. We were the first of the companies to get access to the Mythos program that had 11 companies initially, where they made the model available without guardrails. Our product teams have been working hard to patch vulnerabilities before Fable came out. I feel there is a massive step function of vulnerabilities that will be found. The time from when a vulnerability is announced to when an exploit happens is now a matter of minutes. Patching is hard; it takes about 45 days, and only 20% of vulnerabilities ever get patched. 80% go unaddressed because you don't have enough resources. You have to patch the infrastructure, but also have mechanisms to address vulnerabilities at the root, like using memory-safe languages because most vulnerabilities happen with C and C++. The other area is providing compensating controls during the window when you're exposed before a patch is applied. It is a non-trivial risk that you should not take lightly, and it's a national security risk if critical infrastructure like power plants or hospitals go down, especially in coordination with adversaries.
H
Host7:58
You also mentioned cost, Aaron. Your customers are using all these different models, and you have a really interesting vantage point at Box because you see what people are doing. There's been this meme about sticker shock. Is that real? Do you see that in your customers?
A
Aaron Levie8:14
It is real. As G2 noted, the scale of this... If you looked at the revenue graph of OpenAI plus Anthropic, it's 5x, 8x, 10x in some cases compared to 18 months ago. That's effectively when we figured out how to have agents actually work. You have the combination of reasoning models and the ability for agents to run these models in a loop. That goes from a task that would be a 10,000 to 50,000 token problem per task to now workloads that could be 500,000 tokens, 5 million, or 10 million tokens. So you're looking at one to two orders of magnitude increase in compute consumption the moment a person does a single prompt of an agent. That has exploded overnight, starting with coding, and then we figured out how to take that coding harness into something like Co-worker or Codex. Nobody has budgeted for this because it just showed up overnight. In Silicon Valley, that's okay because you can raise more money to fund the tokens. But if you're in a large enterprise with a CFO modeling the business within a tenth of a percentage point of cost, and suddenly 10,000 engineers start using cloud code, that's a huge bill.
H
Host10:09
So the really bullish thing is that we're even having this conversation. It's very bullish for AI that it has gotten this expensive because empirically people are willing to pay the money. You could not rip out coding agents from our engineering team; we would cease to function as an organization. So the bullish thing is it's working, and the next act is the optimization period. We're not even in optimization yet. There's probably another 5x to go before you really are in the optimization period. But as agents go from coding to the rest of knowledge work, we'll see another jump in total spend on compute as people figure out which workflows are working and where they're getting high ROI. Then they'll shift to a mode of understanding the workflow and spending, say, $10 million on a Fable-class model, and maybe in the future they'll have Fable for 30% and something like Nemo for 70% from a routing standpoint, and then they're only spending $5 million on the same thing. That's the period we have to go through over the next couple of years.
G
G2 Patel11:27
Here's a crazy math. Think about if every employee in your organization uses $200 of tokens every week for 50 weeks. A company our size would need $900 million extra that's not budgeted. And that's just $200. Some people might use more. On average, an agent consumes 450% more network bandwidth than a human conducting that same task. So you've got this massive upsurge of consumptive behavior of infrastructure. These things are economically not viable until you get to an equilibrium. The way we've done it so far is like riding a bicycle. The first time you ride a bicycle, you're not that good at it. It's the same with AI. You have to get familiar with it first, then get good at it, and then get efficient with it. I think we are between phase one and two right now. That's a very token-consumptive phase because while you're getting familiar, you're doing a lot of stuff that is expensive but not really value-creating. You have to move to getting really good at it before you get efficient. But during that time, you have to have a mechanism to regulate so CFOs don't give up on this thing.
H
Host12:52
Yeah, just go raise the $900 million. It's fine.
A
Aaron Levie12:55
I mean, drop in the bucket. We'll just go to Sequoia with the success you folks are having right now.
H
Host13:03
Aaron, you mentioned the skyrocketing revenue of these companies. Those are non-GAAP financials laundered through the media. Both OpenAI and Anthropic have filed confidentially their S-1s. We're going to get the real numbers soon. You have a good vantage point for who's using what in the enterprise. So who do you think is better positioned going into the public markets? OpenAI or Anthropic?
A
Aaron Levie13:33
Okay.
H
Host13:33
An easy question.
A
Aaron Levie13:34
That's a really smart one for you to answer very definitively.
H
Host13:37
Yeah.
A
Aaron Levie13:38
I would answer that question if I were in Japan and we never saw the English translation. But in the heart of San Francisco, I'm not going to answer that question.
H
Host13:48
Are you big in Japan?
A
Aaron Levie13:49
Yeah, we are. Actually, I was there last week and I was much more willing to give market share comparisons.
H
Host13:56
I just had my agent look. I think it's very clear that Anthropic is having a huge burst of success. We're seeing it in every enterprise we talk to. Claude Code has completely taken off. There's a certain draft that happens when your most advanced users adopt a technology. The thing about AI, and G2 and I have been in the enterprise world, we're probably wired this way from day one. The moment we saw AI, it was like, 'Yeah, this is an enterprise tool.' There will be lots of cool consumer use cases in healthcare, legal, and entertainment, but it's an enterprise tool. Where do you spend money on intelligence? You spend it in the enterprise. What Claude Code did really well was win the developer, and the rest of your company will want to do what your developer is doing because that's where people value intelligence the most: your engineering team, your most technical audience. Then it follows from there. So the phenomenon we're seeing is Claude Code taking off, and then the adjacencies to that. That was the story three months ago. I think Codex is now having another moment, so Codex is firmly back in that race. It's extremely early to try to predict medium- and long-term outcomes. You should assume at a minimum we have two to three major players in this space. You can't even count out things like Cursor with the composer model. They have access to the full SpaceX cluster. The moment you think you understand the market share dynamics, a massive shift happens overnight. So I don't think you can predict winners or losers at this point.
G
G2 Patel15:54
AI is still in the game.
A
Aaron Levie15:55
Of course. Absolutely.
G
G2 Patel15:56
But does it have to be Grok versus Cursor? Cursor is a harness that runs agents and uses Colossus compute. At some point, the value probably accrues to whoever is powering the underlying infrastructure and the models, especially in a world of the Neotrons layering in. So you absolutely have to keep SpaceX in the game.
A
Aaron Levie16:24
We are early investors in Anthropic. We were the first design partner for Codex. We have a very strong partnership with Google. We work very closely with xAI. I have to say it's actually not a good business decision to just bet on one, because you're going to have three or four major players in the market. They'll play a very competent game. There's a race condition going on right now. There's no reason any of them are going to stop, and they all have enough compute resources. The way the market is evolving, Colossus 1 was given to Anthropic, so Anthropic had enough compute, but then xAI is going to use Colossus 2, and they have Cursor that they're going to buy, which will benefit from the xAI model. All of these things start to compound on each other. So it's very hard to predict what's going to come out ahead.
H
Host17:26
You both have been pretty early on that there's not going to be a job apocalypse, and it seems like the world is coming around to that idea. At the same time, you have layoffs being blamed on AI at Meta. What do you make of that?
G
G2 Patel17:43
On our side, we have one product that's 100% written with AI. By the end of the year, we have half a dozen products that will be 100% written with AI, no human lines of code. By the end of next year, 70% of our portfolio will be 100% written with AI. You would think we should let go half our engineers, but all of my bottlenecks right now are human. In a world where you're writing unlimited amounts of code, code review becomes a bottleneck. So you automate code review. At some point, when you can write anything, what you choose not to write is far more important, which means product judgment is a bottleneck. I don't think you can outsource product judgment. We're still pretty far away from human instinct and judgment being easily outsourced to an AI model. You still need humans to direct it in the right way. We have a large shortage of people who understand AI well enough and are dexterous in AI, and we're paying ungodly amounts of money for those people. So it's not a matter of letting people go. But every job will get reconfigured. The gap between someone fluent in AI versus not is not going to be a 20% differential in productivity; it's going to be 50x. When you get to a 50x differential, it makes little sense to hire someone not fluent in AI. As a society, we have to invest in upskilling and training people. We're working with governments around the world to have training programs in every major country to get AI dexterity to a base level. It's like not knowing how to use the internet now; you wouldn't hire that person. It will be the same with AI.
H
Host19:57
Aaron, you call this an enterprise technology, but ultimately it's going to change our lives as individuals, whether at work or in our personal technology lives. I'd love for you to tell me, when you think about what the world will look like when we get past this phase of building infrastructure, what's work going to be like? How are we going to interact with technology?
A
Aaron Levie20:21
I think you can already see this version playing out in the valley. This is why I'm optimistic on jobs, because the fastest-growing startups are hiring like mad. What are they hiring for? The things that computers can't do. Computers can do a lot in digital domains where they have the right information, but they don't do a lot in domains with human-to-human interaction or anything in the physical world. You're going to have agents doing a bunch of work for you in the background at all times. You'll give those agents tasks, and your responsibility is to use your judgment on whether you gave the right task, whether you veered it in the right direction, whether you reviewed the work, and whether you incorporated that work into a broader workflow. In engineering, we're probably overestimating the near-term impact of AI because engineering has properties that work well with automation: it's text-based, users are highly technical, and the data is lots of code. So you see profound jumps in productivity. But then you go to the real world and talk to a lawyer, a financial analyst, or a salesperson at a bank. Their daily task list isn't all inside a terminal with text that needs to be autogenerated. It's meeting with a client, parsing their concerns, applying judgment, doing deeper analysis, and interpreting that analysis to decide whether to make an investment trade. So most of the economy's work will get a major boost from AI, but there are still bottlenecks where humans have to do additional work. Anthropic wrote a post last week on three possible scenarios for AI. In their most fast-takeoff scenario, they had one paragraph on the bottlenecks of society that don't relate to compute: how long it takes to develop a drug and get it into the economy, how long it takes for democratic societies to change with new technologies. Those are the bottlenecks. That's also why you have an optimistic job scenario: there's more time for people to figure out what that new type of work is, and some things will not be automated because they're not things computers relate to.
H
Host23:42
I think as we run out of time, I'd love to ask you a question. Are you still thinking about something like AGI on the horizon, where we hit this point where it completely blows away everything we're talking about today and goes to another level, or is that something you don't think about anymore?
G
G2 Patel24:00
Aren't we already there to some degree? It's bonkers how quickly we get used to this stuff. Everyone now complains about the seats not being comfortable in a Waymo. This thing is driving itself! It's nuts, and we're like, 'Yeah, that's kind of normal, but the seat's not quite comfy.' My ping to the Waymo team was, 'Can we get Wi-Fi in the Waymo?' Which is a very first-world problem.
H
Host24:32
I have not heard that, but I guess I don't have a nice enough car. These are Jaguars, right? They should be comfortable.
G
G2 Patel24:37
But I think the more germane question we should ask ourselves is what the strategic modes will be for a company moving forward. There will be a cultural mode around speed that's really important. There will be a mode around efficient generation and use of tokens that is secure. There will be this notion that you have to have judgment on the products you build, and what you don't build will be important. As a matter of habit, you have to be early to market and you can't be late. The speed at which you operate cannot be a secondary thought. If you're a large company and you decide to work slower, that's not a good survival tactic.
H
Host25:29
I'd love to continue the conversation, but we're done.
A
Aaron Levie25:32
But you got... I'm happy I reunited you two. It's been really fun. Thank you all so much.
G
G2 Patel25:37
It was really fun.
A
Aaron Levie25:40
Thanks, man. It was great.