Together AI raises $800M Series B as open-source models hit enterprise-grade performance

seconds online, in store, on mobile, on social, on marketplaces, and now with AI agents. Our next

feel like I'm going to need these for the next one. We definitely going to need those because we have the co-founder and CEO of Together AI in the waiting room. Let's bring him in to the TV van, Ultradom Pimple. How are you doing? Welcome to the show.

Doing great. Thanks for having me on the show, guys.

Thank you so much for taking the time on such a busy day. Um, please

first time, right?

Yeah, I think so. Right.

This is the first time. Yes, I've been a fan of the show. So,

took us took us way too long.

Long overdue. So much to talk about. Uh, first, uh, give us the news. I want to hit the gong and then I have a million questions for you. So tell us what happened.

You know our our business which really serves uh open weights models and powerful AI at dramatically cheaper costs has been growing tremendously. We've grown uh you know uh from 100 million in bookings to 1.2 million over the last year and we raised $800 million in our TV to expand.

Congratulations.

Thank you. uh a Ramco Ventures fascinating partner makes a lot of sense with energy. Um but I'm so interested in um what is ju just the state of open-source AI. There's been two narratives. One was uh there was this chart that showed closed source growing a little bit faster seemed like there was an acceleration and open-source was going a little bit slower. I believe this was from the government. They published this the US government. Uh, but then most recently with GLM 5.2, it feels like open source is catching up again. Has there been as much of a roller coaster race, a roller coaster ride inside the company as it feels like from the outside or what has progress felt like in your business throughout the last couple years? Yeah, I I I I would say we one I would say open source models have really closed the gap with closed models and this is why you're seeing all this excitement around uh open open weights models and this has been a process that's been going on. So I you know the the the first and the second derivatives of progress have been have been you know the gap have been negative and I think now you're at a point where the the open weights models are able to really address the largest workloads of agentic long range uh you know long horizon tasks that have been deployed in enterprises over the last 6 months. So there is one a real need for these uh models. there is uh you you know a real reason to sort of try to optimize your cost because you're trying to deploy AI in a much broader sense now it's doing real intellectual labor and making a big difference to uh you know the progress in in companies so I I think there's a demand side of it which has uh uh shown up in the last 6 months and the progress in open models now they are very very uh you comparable to the closed models. How do you think the uh developer landscape of the open-source models ecosystem will evolve? Because uh as I sort of understand the history, Mark Zuckerberg made a big splash. He's going all in on open source with Llama, some really solid progress there. Eventually, it stopped making sense to give everything away for free. Went a little more closed source. We'll see where that goes. But then China as of last year just deepseek moment Alibaba Quen there's so many different Chinese developers and is that a cultural decision? Are they just pro- open source broadly? Is this geopolitical strategy? Uh what is the motivation and how long do you think that opensource stance will last or will we see China go closed source at some point? You know, I would say that it's really a function of the structure of the market. In China, that market is structured around open models. All these companies are making revenue. They provide APIs, they provide applications, right? They they are they are businesses that are uh competing with each other and are generating revenues and the structure of that market is models are released in in the open and the structure of the market in the US so far has been the models are closed. uh so I I I think there's a little bit of game theory to that which uh is uh you know these markets are sort of set in this particular shape but I I think at this point things are changing quite a bit you you have you know an extreme demand for tokens tokens are becoming like a fundamental resource like energy or capital or bandwidth and you have a more modular market that's forming right so you can you have harnesses where you can take uh uh you you can point them to a closed model or an open model and they continue to work. So you're starting to get this market formation which allows a place for open models and I think what that will turn into is you will see more open models in the United States. I you know we're talking to many companies who are working on uh very ambitious road maps uh for producing open models. So there's just going to be uh the existence of that market worldwide. When you talk to customers, are they almost always coming to you with a workflow that they prototyped with a closed model, prototyped with a frontier model, and they've got it working and now they want to cost optimize or do you talk to customers who are just so allin on open source they don't even care about the frontier? Uh because I've seen you know there's companies when you know GBD4 came out they were like oh I this can process uh OCR documents and just organize text data very effectively. Now we're in a coding boom that but we get these spiky intelligence unlocks every once in a while and then the cost ramps up because you're like ah I can use this everywhere and then people want to cost optimize. But our c our company's typically coming to you with a uh a workflow that they've already prototyped and they feel like is confident they want to match the performance but at a better price.

Yeah, I I I would say they're both use cases. There are companies that have prototyped something with closed models. Once they deploy it into production, they realize this is like is absurdly expensive. they can't quite scale this and deploy it into the uh you know deploy it into their companies and they start benchmarking open models and uh I think they're increasingly finding that uh open models have become pretty strong on these benchmarks and and there's a sort of substitution uh uh happening and and there's really a stampede of this uh that companies uh starting from AI natives now to digital natives are deploying open models at uh you know really really rapidly. You also have other use cases. I think you know companies that have sensitive data, this is really their strategic asset in in many ways. It's uh uh it's what the what the intellectual property of the company is and there is increasing amount of discomfort especially as the models are getting smarter that by using closed models they're sort of giving away their strategic advantage by teaching the model how to run their business. And I I think that's becoming a bigger threat and we're starting to see companies more sort of philosophically choosing open weights and AI that they control. And I I I expect we'll see more of this.

How are you thinking about regulation? It feels like GLM 5.2 who sort of reignited this open-source debate where you have anthropics mythos and fable sort of held up back and forth banned unbanned released uh released to just Americans released to just a few companies uh and open going through some similar stuff with 5.6 six soul and some other models. And uh all of those debates, they make sense in isolation to me. I I don't want a cyber weapon or a boweapon if if it's capable of that. Uh at the same time, if you could just go on Hugging Face and download the bioweapon creator, uh I probably don't care that there's a closed source version of that. It's a different discussion. How have you dealt with the regulatory questions that are increasingly popping up around AI safety and AI limitations? I I think these are important questions and you know the one of the things that we feel is uh these models especially open weights models they're software right you can you can inspect them you you can benchmark them you can study them in a way that you can't do with those models

and that is going to become a much more important aspect of regulation where we uh you know frankly I think some of these risks um are are could be real they could be rhetoric they it you know we just don't know because

there isn't good quantification of this which I think is going to become more and more important for us to understand how to you know think about a regulatory framework that uh is sensible and is uh you know if these models do become dual use how do we sort of protect uh uh you know protect against bad effects of these models while leveraging the value that they create I mean they're immensely valuable this is the most transformative technology ology we've had uh ever and uh we you know we really want to make it abundant rather than uh limited to a few people.

Completely agree.

Do American labs focus on open source? Are they fundamentally disadvantaged because they're going to have more scrutiny around distillation? Hm.

Yeah, I think I I think that's a that that's a good question and you know I would say uh uh uh perhaps the role of distillation is overstated these models if you look at what's happening

saying it's code.

Yeah, I I I think it it is a little bit because you you know the the the technical details of it. you you do need the probabilities of tokens and all the information that APIs don't provide to do effective you know on policy distillation a lot of the recent improvements are because of reinforcement learning and the fact that you AI has become an industrial platform you have a transformer architecture that if you look at any of these models they look very similar inside so you have this common architecture you have training tools you have RL tools that have become standardized, which is why you're seeing this like rapid improvement. You can now start with machinery that works instead of having to invent it. And I think you will see that American labs that are building models will have more success than the labs that tried to do this two years ago.

Yeah.

Yeah.

Makes sense.

Makes sense.

Thank you so much for coming on the show. Congratulations to the whole team.

Congrats to the whole team. Yes. And have a great rest of your day. Have a great July 4th. Have a great weekend. We'll talk to you soon.

Come back on soon.

Thank you. Talk to you soon.

Have a good one. Let me talk about