AI is too important for the rest of the world to say we're only going to use the API calls from three companies in the US. There's a bunch of people secretly cheering that this is exactly what you'd want to have happen to design the regulatory review process for AI model releases. Intelligence is going to be multifaceted. It's not going to be one or two models doing everything.
Today on Billions, I'm sitting down with Aaron Levie. He's the leader who rapidly transformed Box from a pure SaaS business into a cutting-edge platform for AI agents. We dive into the high-stakes battle for sovereign AI, government regulations, and why the narrative that SaaS is dead might be completely wrong. Aaron, super excited to have you on.
Super excited to have you on. Like I know you are like you have very interesting take about AI. I'm curious to get your last take, you know, on what happened with Anthropic recently because I think you said it was like a big win for open and lightweight models. So can you maybe elaborate and your thoughts on the topic?
Yeah, so I think what first of all, the favorable model is kind of an unprecedented level of intelligence and capabilities. So clearly Anthropic designed an incredible model and trained an incredible model. But the US government effectively enacting an export control which meant that non-US citizens aren't able to use the model is really a completely new precedent and moment in AI regulation where we've never seen that before for any kind of frontier model being held back legally from being available in the market. And the implications are first of all, it's almost entirely impractical for any AI lab to be fully responsible for the nationality of the customers that are using it. At a minimum, it would completely obliterate the API business of Anthropic because the idea that you would be able to control the end customer through your partner ecosystem is very difficult. So it effectively has the impact of making the model unavailable to the world. And so the risk is if you were either a major app developer and you want to be able to make this technology available to your customers or you're a country anywhere around the world that wants to ensure that your ecosystem is able to build on top of leading frontier AI, this poses an all new risk, which is at any moment, can that AI model get pulled back from one government or due to one government. And we just never had that precedent before. AI has been this very fast-moving, very market-based ecosystem over the past couple of years. And so I think first of all, the US government probably very quickly needs to improve the situation with Anthropic. I think Anthropic is quite motivated to resolve whatever the particular issues are. But even when that happens, it's easy to think in a year from now or 3 years from now or 5 years from now, how many more versions of this might we see in this landscape that will play out. And so it makes you wonder while if you're a country Germany, France, Japan, UK, and you have the capability to be at frontier near frontier AI from a talent and an infrastructure standpoint, do you need to be building these kinds of models yourself and make sure that you have them available to businesses and users within your country. I think it really introduces that question in a much bigger way. And we've already seen examples of this. China obviously developing great open source models. Mistral doing the same. But I think this highlights the increased importance in that dynamic and the idea that you can have sovereign AI or AI that you control. Now, there's one caveat to all of this though, which is you're not going to take that big of a sacrifice or a discount on the model's intelligence for that benefit. So if frontier from US in a closed model environment stays 10, 20, 30, 50% better than the next best model that you could make sovereign, I think people will take that risk and they will use the frontier model at all times. So this only works if open weights models or sovereign AI can remain close to the frontier and much closer than it has historically. And I think we're seeing some examples with Kimmy and DeepSeek and a few of these new models coming out of China and Nemo Tron. These models are doing a great job staying at the forefront, but they're still not Fable level. And that's going to be the big question for the ecosystem over the coming months and years.
Do you feel like it's only going to be a matter of performance like staying, let's say, as you said, 20, 30% behind or also a matter of timing? Because what we're seeing right now is like when the first model went out, the first lab to catch up maybe took like 12 months or a bit more than a year. But now I feel like models are catching up within a couple of months. So do you think it's only a matter of difference or also a matter of how fast can other models catch up on the best models.
Well, I think there's a big debate and if you watch Dwarkesh talk to Dario or Jensen, you sort of see examples and contours of this debate which is is the frontier and the open and the best in class open weights model, how close are these tracking versus are we seeing a divergence? Are they six months behind or three months behind or one month behind or how much is it really just a factor of how much compute you have for inference and test time compute as an example because there's different ways to drive the workloads in these models to get better results. I am generally actually optimistic that the frontier open weights models will be able to stay relatively caught up. I think it's of such economic and strategic importance for countries like China, maybe individual companies like Mistral or even Nvidia to make sure that is maintained and we have not yet seen a breakthrough that is so either proprietary or undiscovered by the rest of the ecosystem that keeps this completely closed off. I think there's obviously a conversation around RSI and what self-improvement look like and does that lead to an accelerating takeoff like whoever is able to design these systems that self-improve first then obviously there's a compounding dynamic to that. At the same time, it's hard to imagine China not wanting to stay in the race in a very big way. It's unlikely that China says, 'What? We're totally fine with Anthropic and OpenAI the only vendors of AI in the world and we know we'll be closed off from this ecosystem.' So at some point you can just throw enough money at this problem. China could just say we're going to spend 50 billion dollars and just make sure that this is of such critical national security importance that we have to make sure that we're throwing compute at this problem. We're going to sequester the top researchers and scientists and make it happen. I think it's unlikely that that wouldn't work and I haven't seen any evidence that suggests otherwise.
So your review like when it comes to China is what regarding the chips because for so long there's been kind of an embargo on chips and it was not possible for them to get them and they were getting it through like Singapore or all these kind of things. Like what's kind of your view because it feels like eventually I mean in the US it has always been very pro open market and now it feels like things are moving away from it. So I'm curious to get your take.
Well, if you and I think a lot of very smart people believe this. If you believe that AI is both a national security threat and opportunity and it's of critical strategic economic imperative, that can lead some people down the path of well, then you actually want this to be a closed environment. You need to keep the chips. You need to keep the malls. And I do think it's of national security level of importance but there's this interesting tension which is if you're also not the exporter of the technology then the ecosystem effects start to favor the non-US companies and Jensen made this case in the Dwarkesh podcast about a month ago. I actually I probably lean more toward Jensen's side of this argument, which is, if you just play this out, AI is too important for the rest of the world to say, 'Oh yeah, we're only going to use the API calls from three companies in the US.' It's just too critical of a technology. So sovereign clouds have to get built out. GPU clusters are going to get built out globally. Countries are going to want to train their own models. That's an inevitability. So if that's an inevitability, then you have a choice. Do you want your technology stack? Do you want your architecture to be the architecture that powers all that? Or do you want to force the rest of the world to have to go and design another architecture? And if that happens, who stands to gain the most? Obviously China, because they will just eventually stand up enough compute, enough infrastructure, and make that happen. Now, I think this is a very worthwhile debate, and I am only like 80% confident in my opinion. The 20% where I'm not confident is you could possibly squint and argue that America is just so far ahead in compute and chip design, and now we're starting to figure out how to do fabs. And you've got Elon with terrafab, and we know how to do the orbital data centers maybe in the future, that maybe there's one scenario that you play out, and that is that we do actually just have the API for intelligence, and that we control the access to it, and we can make sure that only our allies are using it in the right ways, and we can prevent the bad actors. I think there's certainly a scenario that looks like that. That's what many of the AI safety ecosystem is hoping for as the outcome. And I totally grant there's some percentage possibility of that. But I kind of default to this is such critical infrastructure and so important that it will cause other countries to have to just go their own direction if the US isn't enabling that.
Yeah, personally I'm more on the side of people should let Jensen win, you know, like he has a chip should like sell them everywhere and
What's interesting is you can already now see the downstream impact. So Jensen was the first to experience this but Anthropic is now the next to experience it. It's the same exact thing. So if you already believe in the precedent that you shouldn't export a chip to China or some other country designing leading AI then it stands to reason you probably also aren't going to let the leading AI model that could help train the next model or help crack systems be available to those countries or certain ecosystems. So we now have, because of the chip export controls, a lot of the foundation for now what we're going to be dealing with. And again, the risk is that creates a very unstable environment for countries globally. Now again, as long as we have the best models, maybe you can just power through that but it creates a lot of incentive for other countries to want to respond in kind.
Yeah, and do you feel like looking at Anthropic marketing in the recent months? I mean, it's been pretty common for models to make this huge announcement that their next model shouldn't be released because it's too dangerous. And we saw that with OpenAI. So obviously what it does is everyone gets excited about the model.
Anthropic did that six months ago and now they've done it again with Metis and Fable. Like do you feel like their marketing went a bit too far and it's just like going against them or do you really feel Yeah.
No, no, no. I was just wondering because like David Sacks like was sometimes working with the White House like what he was saying is that essentially they emailed the Anthropic team and they told us, 'Hey, can you please deploy a patch on that specific model?' And apparently the Anthropic team hasn't been that responsive. So we're hearing kind of two sides of the story and I think in the marketing of Anthropic it's been very real from the start that they want to be seen as the good guys versus evil Sam Altman, which I think is not super fair when you look at the global environment. So yeah, I would love to get your take on this.
Yeah. I you know, this is one of the more complicated topics right now because so let's start with the easiest part. Unquestionably the rhetoric that has come out of Anthropic and other safety researchers created the environment that we're now in where if you scare the crap out of the government by saying this is the most dangerous thing we've ever seen and has all these risks and then the government sort of, we can debate how compelling the particular jailbreak is but if they eventually discover that there's a jailbreak on what you claim is now the safe version and then it defaults back to the unsafe version or it can do something that it perceives as unsafe, well, that's clearly a byproduct of creating the atmosphere where everybody is concerned about this. So unquestionably it's coming from that tone and that rhetoric. So that's the first layer. So that's quite straightforward. It's pretty obvious and so we're dealing with now the consequence of that. The interesting thing that I don't know though is, is it plausible that there's a scenario here where actually Anthropic prefers, weirdly, prefers this outcome. Which is now we have a precedent that has been set that models at a certain threshold of capability do have to go through an approval process. They do have to go through a review process. They do have to be approved and vetted by the government. That's not so different from what the regulatory lobbying has been pushing for from many of the AI safety researchers. And so interestingly, if you study the journey of this whole space, the top AI researchers in safety and regulations and controls probably would have designed a system that says the government is going to approve the release of models at a certain capability level. We're going to do lots of red teaming. We're going to get in a room and we're going to agree that the thing is safe to release. And nobody had created the catalyst previously to just build that environment. So I think there's one other layer which I hold out a little bit of percentage of my thought process on is maybe this is exactly what Anthropic wanted to happen. You know, we're all dealing with the day-to-day drama of, 'Wow, I can't believe they responded that way.' Or I can't believe Amazon is the one to like, we're caught up in all the little leaks and the messages. When actually you zoom out and you're like, 'This is sort of how you would design the regulatory process of AI model releases if you were a deep AI safety research-oriented person that was really concerned about these kinds of risks.' Now, I happen not to be in that camp. So I think this is a bad outcome. But I could totally be convinced very easily that there's a bunch of people secretly cheering that this is exactly what you'd want to have happen to design the regulatory review process for AI model releases.
