You look at the trends: open source is going to become more and more popular, and it's going to close the gap. But the other thing is that it's not necessary to close the gap. I think the opportunities for the infra startups are always trying to identify some of the core problems and solve these problems and excel at that. But post-training is going to continue to grow because it's so effective at improving the capabilities of already capable AI models for excelling in particular tasks.
Ion, it's very nice to have you.
Thank you. Very glad to be here. So I think the message I delivered today was several points. I think number one is that infrastructure is becoming increasingly complex for many reasons, because the hardware is becoming much more complex, applications are becoming much more complex, the speed of innovation at the hardware level and application level has increased dramatically, and because people want to have ultimate efficiency, because of the very high cost of running these AI workloads, everything is tightly coupled, everything is integrated, so you need to develop everything together as a big, big, big system. So that's kind of number one.
I think the number two I would say is the fact that there had been a lot of discussion about AI breaching different systems, cybersecurity, and I think there's a question about how dangerous is AI for our society in general, for humans. And I think my position there is that we should take a step back and remind ourselves that at the end of the day, AI is a system, a tool, and yes, it's very complex and it's going to behave in ways we do not expect. It's not the first time we had complex systems before, like the internet and large distributed systems, which behave in ways you do not expect. And we should really focus on understanding as much as possible from all of these attacks and use that understanding to improve the safety, to guide the systems to achieve the goals we want. But we should never forget that, I think at the end of the day, they are tools, and these tools are there to improve our productivity and, in general, improve our life.
Yeah. To address your first point, we have seen so many vertically integrated companies that optimize every layer of AI stacks. Where do you think are the opportunities for AI infra startups?
So I think, you know, the opportunities for the infra startups are always trying to identify some of the core problems and solve these problems and excel at that. I think that you need to focus also, you have bigger opportunities when you are going to focus on new emerging problems, right? So there are things like, and many companies are doing that, what are these emerging problems, emerging workloads? I think clearly one is about post-training. We've seen that for the past year or so, actually many companies going into that space, but post-training is going to continue to grow because it's so effective in improving the capabilities of already capable AI models for excelling in particular tasks, right? And it's also driven by the fact that right now there are a lot of very competitive open-weight models, many coming from China, which you can use to post-train to significantly improve their capabilities and exceed, for certain tasks, the capabilities of the best proprietary models.
The other thing, I think the other area will be data, especially multimodal data, because right now you are going to have more and more multimodal models and applications, and I think that having better ways to process this kind of data together — video, images, text, sound, audio — I think it's going to be an increased challenge. I think the third one, which kind of connects the two, and you see that, it's in the robotic space, right? In the robotic space it's still early. Many of these companies, they still build their models. So having maybe an open-source stack or a stack which is working well for robotic applications, where you don't build, you know, trillion-parameter models, it's a little bit more tractable. I think that's kind of very, very important and it's an opportunity.
Another one I would say will be related also to data. There's a lot of challenge just moving the data around right now. It's like when you do, even when you do training, because of the scarcity of the GPUs and AI chips in general, even if you are a big company, you are going to get resources and GPUs from different cloud providers. Maybe you are going to build your own data centers. So then you need to put this together, right, and to make it easy to use and to train across, you know, different data centers, different clusters and so forth. And there is both how you train it, how you provide this a little bit more portability — this is a little bit what SkyPilot is doing — and how you move data very efficiently, right, between this kind of location, because a huge amount of data, right, you train on tens of trillion tokens or even more. So I think these are some kind of aspects.
We have seen, for example, the AI labs, they are really competing head-to-head and they're accelerating their pace of releasing new models. What do you think of the evolving landscape of those models, like frontier models versus open-source models? Do you think they will eventually catch up the gaps? What is the moat, defensibility?
Moat? First of all, as you know, I am a strong believer in open source. A lot of things I've done, it's open source, and so therefore my opinion will be colored by that. To answer your question, I think the moat of these, the biggest moat is capital, I think that, as you can see, if others can match their capital, or even if they have even less capital, but they have enough, they can create competitive models. That's why you do not have, even to this day, you do not have a lab which is absolutely dominating, right? It can dominate for a few months maybe, but then things are equalizing, right? Everyone even forgot, you know, forgot that at the end of last year, so it was just a few months back, arguably the best model was Gemini.
So you remember that. So from the proprietary models, right, so you see how fast things are changing. And now in February and so forth it was, and for several months it obviously was Claude, right, is Opus, then Opus, and then Astra, pretty good model right now for coding. So it's maybe a little bit better now, you have Opus 5.5, also very good, I'm talking about coding. So history shows that no company so far has proven that it can build a moat other than the capital, which is very important, and you have capital and then you have revenue, right? So that's a huge moat. But in terms of technology, even if you look at Chinese models, right, it was DeepSeek, right, in February last year, right, everyone was talking about that, amazed everyone, but right now it's like you have GLM 5.3, Kimi K3, which arguably are better than DeepSeek for some tasks.
So, and this makes me also think, because the labs have at least haven't been able to accumulate this kind of a technical moat or something like that, at least a visible one, that open source still has a chance. And I do hope that also some people will say that, you look at the trends, that open source is going to become more and more popular and it's going to close the gap. But the other thing is that it's not necessary to close the gap, because once you have, for every task there is a threshold. If the model is good enough, right, it's going to solve that task, then being 10 times better for that task, 10 times more capable for that task would not matter. Okay, you know, it's like a car: if you get a car to go safely 100 miles per hour, this is what you need. The fact that a car is going to go 300 miles per hour, it doesn't really matter for that, if your task is not to race but to go between Santa Clara and San Francisco, try to drive. So I think because of that you'll have more and more models, and many more open models being perfectly, you know, good enough or more than good enough for more and more tasks, right? And I think this is what, you know, some people like, some companies like Vercel say that, well, in the past few months it's like what they, from their router, AI, you know, routers, they see that it was a kind of a transition, I don't remember exactly the numbers, but from 70%, you know, a few months back being proprietary models, is now 70% it's open-source models.
