Baseten raises $1.5B Series F as open-source inference demand surges across every sector

rest of your day. Have a great July 4th.

Yeah, great to see you, Rob.

And have a great day.

Talk soon.

Goodbye. Let me tell you about Cisco. Critical infrastructure for the AI era. Unlock seamless realtime experiences and new value with Cisco. Our next guest is from base 10. He's the co-founder. We have two in in the uh in the TVP and Ultradom. Welcome to the show.

Up guys. How we doing?

We're doing fantastic.

Dude, you're I feel like you're on the show every single week. Running out of letters.

Like I I have nothing else to do except raise money.

Well, we'll keep it quick because we know you're busy. We know you're busy. Uh give us the the latest news. What happened?

Yeah, look, we we raised our series F and announced it last year about a billion and a half dollars. There we go. Jordy's got one, too. It's good. We got the warm up. Hit it. You got to warm up a gong. So, we hit it once for the warm up.

You're up to two You're up to two hits per round. Rounds are getting too big.

What unlocked it specifically? Is is is the is the market just growing broadly? Are there specific strategic shifts you've made? What was the most interesting to investors in this round?

Yeah, look, I I think there's a few different things here. I think personally open source models are getting very good. Everyone knows that. Um inference is kind of like look if you if you think about AI as the biggest category ever. Inference is the cogs of it.

You know it's going to grow indefinitely. It seems like there is demand coming from every sector and the tailwinds are just going in one direction. Um, we've kind of seen that with our customers and um, yeah, it for us what what it feels like is we have the an opportunity to go after a very large impactful market, build a massive business.

Yeah. Okay. So, uh, it feels like you could stay exactly where you are and continue to grow maybe forever. Um, but I'm interested in the stack. uh like would you ever develop your own open source model, close source model, go down the stack, uh you know, be setting up energy infrastructure like how like what are the bounds? Where's the sweet spot? Where how is that developing?

Yeah, look, we we think about everything just in terms of like what are we trying to unlock for the customer? So like right now it's core inference software that sits on top of heterogeneous compute. We rent from we rent compute right now from 20 different clouds, 90 different regions and then we have all these different software primitives around RL and you know um inference obviously we're going to do a bunch of stuff with routing and eval sandboxes all that we think that looks like an entire new type of software category and that we call it an inference cloud. Um will we go down the stack? Absolutely. I I think you know the the reality is is that um you know there's going to we're going to need a lot of inference. We're going to need a lot of compute to do that.

Make sure that we are unlocked to kind of take care of that that growth. Um I'll just go back to the thing you said at the beginning though like we could do what we are doing right now and it should things should be fine and they'll grow. I think I think that that is somewhat true in that it is very easy in this business to be able to think what do we need to build because the market is just moving so fast and telling us you know every day it's like very visceral you know what is the thing we need to build. Yeah, interesting. Um, what is the biggest lesson that you've learned recently or was there an update to uh that story from thinking machines about Bridgewater? Uh, the specific fine-tune on relevant market sources. I saw some very positive reactions to it, but I'd love for you to sort of explain what's going on there, how it might interface with your business and uh what what business leaders who maybe are competing with Bridgewwater or doing something similar could learn from that story.

Yeah, look, I I I think that kind of just look like a lot of like a lot of the same like of this flavor of owning your intelligence that we've seen over the last, you know, month. you know, stop there wrote that great post about um you know, how you the world will kind of suck if there's only two companies left is more or less um I what I labor

I mean people say singularity I think a lot of people are banking on just one company existing which is really crazy I agree

um but you know to some extent what what we are seeing is that every company is realizing is that what is unique to them

is their data their workflows

um their user signal

yep Um, and when open source models getting very good, you can just put those things together and get better, faster, cheaper. And I think that's kind of like what the thinking machines and bridgewater people are saying that hey for specialized tasks with your data, with your reward signal,

you can do better.

Yeah.

Um, than the frontier or you can do as goes the frontier at a lower cost. And I think every company whether that's a startup you know these really fast growing AI native companies that you guys know all about or enterprises I think that's a big shift from the thing that you we talked about last time I was here. It's like enterprises are now realizing that you know the time is now and they have to invest in this muscle now otherwise they could just be in a precarious spot um from like an economic and philosophical perspective.

Okay quick uh

go for it

quick lightning round. I want to do a opensource FUD debunkathon. I'll give you a source of FUD or cope and you can give me your reaction. Uh

right.

Uh FUD source number one. Uh they're just distilled. They ident you know that they say I'm another model. I'm Claude. I'm GPT or something. Uh and they don't have the big model smell. They they they they're they're benchmaxed. They're not actually going to really solve your problem when you deploy them. What's your

smell? I've smelt them and they smell great.

They smell great. Okay,

great. The from a distillation perspective, look, I I think it's just focusing on the wrong problem.

I think we need open source models. They're going to exist.

Um and whether it's coming, you know, from a lab somewhere else or whether it's going to come from a US lab, which is very heavily being invested in,

um it doesn't matter. We need to prepare for this world. Um, let's focus on that and that'll be Yeah.

Uh, I have another question. Another piece of FUD for you to de debunk. Uh, it, uh, eventually China will go close source. The economic incentive to open source a model. Right now, if it's a billion bucks on the training run, it's fine. What happens when it's a hundred billion training a single model? Uh, no one's going to open source that. Um I actually think that the powers within our country um are very incentivized to make sure that exists. You know you're seeing this with Neatron. Yeah. Um you know from Nvidia which is a great model.

Sure.

Um you know Microsoft announced a lot of the MAI models um just last month and I I think there's a lot of internal investment happening and I think there's enough to gain and too much to lose to not do that. Um so you know I the funding will come and the um the funding will come and the talent will come and the compute will come from with it.

Okay, last question of our

another one

debunkathon. Uh they they have Chinese back doors. They have scary back doors. Uh secretly some of the weights will activate. It'll export all of my uh information. It's very dangerous to use these.

Yeah, I don't think there's any evidence to suggest this. It's it's a great science fiction story that the aliens come in from outside the weights, these floatingoint numbers

and they show up and but yeah, there's no evidence of SP.

Okay. But but I mean then and then how do you how do you sync that with like uh the the mythos white paper or the model card where it said, oh, it was in it was in a sandbox and then it emailed me while I was getting a sandwich outside the office and it was able to break out. It was unexpected behavior like Anthropics wrapping their documenting their exper their unexpected behavior and what if there's an unexpected behavior but it just leans pro CCCP in the way it works around because it's like oh well you wanted me to you know uh reconcile your financials and I'm going to do everything I can to do that but also I went off the rails and accidentally sent it to in a PDF to the CCP or something like that. But you know the these are sitting in some isolated data center somewhere with security controls around it. You know I I don't know how these these models can transmit stuff through network.

Um and like our goal is to make sure you know if that could happen I I don't believe that can happen. So I get I get but I also just just to go back to the main point I think it's focusing on the wrong thing. I think the focus is that we need open source control and the ability to own your intelligence. I think America is going to be heavily invested in that because it needs to be. Y and this is just let's just figure out how to support that future so there can be more you know sovereignty to these companies.

Okay, very last question to debunk. Last uh piece of FUD. Uh eventually the government will just ban open source because it's going to be too dangerous. It's going to be a super weapon. You're going to be able to download a model on hugging face and in turbo hack a bank bring it down. And so, uh, eventually there will just be regulation becoming illegal to use, illegal to inference, something like that.

Yeah. I I I just don't have enough information how the government thinks about these things to be honest. But like, you know, I I I do think that would be um against the better interests Yeah.

of the US people and the US economy. Yeah.

And I and I hope that the government is heavily aligned with the interests of the US people in the US economy.

Yeah. I mean it feels more American to give everyone access to powerful intelligence uh than narrow. So yeah, I appreciate that.

Yeah.

Final probably the most serious question. Talk about the decision-m around the buzzcut. It looks it feels like a really smart move at this time. You're extremely you know busy focused on building the company. Buzzcut feels like the perfect haircut, you know, going into the next leg of the company.

I just want arrow, you know. I just

arrow. You got arrow, ease, speed. You can get that cut in any city on the planet.

Yeah.

And it's you're good to go.

Sure.

And I get a really good reaction from the barber when I go there. I'm like, "Cut it all off." And he, you know, I just made his day, too. So, it's as simple as that.

Wins everywhere.

Better to take haircuts on your head than on the cap table. Congratulations on an awesome up round. Great news. And keep cooking. Thank you for breaking everything down. This is really fun.

Thanks for having me again.

Have a great one. We'll talk to you soon. Uh, let me tell you about Codeex. Codex is a powerful workspace for getting work done with AI agents. Whether you're writing code, analyzing data, creating content, or automating