00:44:28 · Vincent Weisser
[00:44:28] S3: Exactly. Yeah. I think that there was a lot of recent talk about this. Obviously, like folks like Satya and Alex Carr were like, also making making those those points and arguments over the last few weeks. And I think it's increasingly something I think where, like, a lot of, like, enterprises want to basically build this, like, compounding data mode and flywheel, where ultimately, I think the future that we are starting to see is a lot of these companies want to build basically self improving agents. Mhmm. And the way to do this, I think, is generally you you have an RL environment for use case, you scale on it, and then you ultimately deploy that into production, and then continuously improve with the user interacting with that agent. It's almost like the the Tesla autonomy, like levels, but like for knowledge worker agents. So I think like the the ramp example I think was one that people really resonated with for basically them being able to within a week and less than 50 of of training spend Wow. Outcompete a frontier model Yeah. At automating spreadsheets and and finance at a fraction of the cost of of running the cheapest models. Sure. This is like they were able to outperform Opus at a fraction of the cost of Haiku and and like 30% faster, like much more tailored to the use case. So I think in many ways, like, these, like, specialized agentic use cases are a big unlock. And I think in many ways, you still need the god like a big mega model to to do the orchestration or planning. But then I think oftentimes execution, I think, happens can happen with these specialized sub agents. So I think, yeah, those have been extremely, I think, like, fruitful and and and useful for a lot of the companies adopting and and training their own agents. Then I think the other big trend is just like an explosion in in neuro entrants that are, like, taking model training seriously. Obviously, hundreds of Neolabs really got stood up in the last probably twelve to twenty four months. So we started powering a lot of those, like, with our full stack, including with also large compute clusters. So, like, NVIDIA also joined us as part of the round, and we've been like, we've since day one basically operated large clusters. And and this is in in large part why we need so much money. Right? It's like to basically be standing up more and more GPUs.
00:46:42 · Vincent Weisser, John Coogan
Like right now, we run like 15,000 GPUs and scaling to like over 30,000. Okay. And so so this is, like, a lot a lot of the capital needs even. [00:46:50] S1: And walk me through when you're pressing the huge cluster button. Is that because there's broad demand for specific models that you need to inference at scale, and so there's just a lot of inference demand? Or is it that for the level of training that you're doing, you need to be cluster scale to even kick off the training run? [00:47:12] S3: Yeah. So basically, it's like for for all the needs, like, from training to inference, you need obviously compute. In a lot of ways, like, our customers, like, are doing a lot of these, like, large scale training runs with us, like, leveraging our tools, like, including our compute. Right? So it's like pre training, fundamental and parameter models, like, for what we did with the RC, like Q4 last year, or really large also post training runs and RL runs. Right? So it's like increasingly basically, we abstract all of that info away. So basically, people can just, like, hit a button Mhmm. To do a post training run without having to, like, log into GPU cluster and and managing the bare metal GPUs. So we basically abstract a lot of this, like, GPU complexity away for but but yeah. Like, for that, we basically also ramping up our own clusters to basically just, like, serve hundreds and and thousands of customers, like, in parallel on those.
00:47:12 · Vincent Weisser, John Coogan
[00:47:12] S3: Yeah. So basically, it's like for for all the needs, like, from training to inference, you need obviously compute. In a lot of ways, like, our customers, like, are doing a lot of these, like, large scale training runs with us, like, leveraging our tools, like, including our compute. Right? So it's like pre training, fundamental and parameter models, like, for what we did with the RC, like Q4 last year, or really large also post training runs and RL runs. Right? So it's like increasingly basically, we abstract all of that info away. So basically, people can just, like, hit a button Mhmm. To do a post training run without having to, like, log into GPU cluster and and managing the bare metal GPUs. So we basically abstract a lot of this, like, GPU complexity away for but but yeah. Like, for that, we basically also ramping up our own clusters to basically just, like, serve hundreds and and thousands of customers, like, in parallel on those. [00:48:04] S1: Yeah. So what you're saying is, like, there might be a situation where I have a I have a 50 k budget like you gave with the ramp example. I have a 50 k budget to do some post training, and and I need to deploy that 50 k into, like, thousands of GPUs, but only for a short amount of time. And so Exactly. The budget's actually going to be relatively small even though if I were to buy all of that, it would be it's way more than 50 k, millions, millions of dollars. So you sort of absorb some of that cost, orchestrate it, abstract it away, and then you can deliver for the customer. Just a slice of it.
00:48:04 · John Coogan, Vincent Weisser
[00:48:04] S1: Yeah. So what you're saying is, like, there might be a situation where I have a I have a 50 k budget like you gave with the ramp example. I have a 50 k budget to do some post training, and and I need to deploy that 50 k into, like, thousands of GPUs, but only for a short amount of time. And so Exactly. The budget's actually going to be relatively small even though if I were to buy all of that, it would be it's way more than 50 k, millions, millions of dollars. So you sort of absorb some of that cost, orchestrate it, abstract it away, and then you can deliver for the customer. Just a slice of it. [00:48:38] S3: So so basically, I think there's this, like, also big unlock, like, also on the inference side, right, where customers don't need to, like, front load hundreds of millions and and CapEx spend for large data center commitments. [00:48:49] S1: Yep.
00:48:49 · John Coogan, Vincent Weisser, Jordi Hays
[00:48:49] S1: Yep. [00:48:50] S3: So we basically also increasingly saw a lot of inference. So it it really is the full stack, and I think a lot of the compute like, we we have, like, quite a a clear, like, visibility into the demand just because, like, people come with us with large requests or or large contracts to basically do larger training runs and deployment. So we we basically just, like, need to secure the compute to, yeah, to be able to fulfill all those. And I think, like, right now, we would probably be at, like, many orders of magnitude more compute like compute, but also revenue if compute wouldn't be as constrained as it is right now. So it's like right now, I think, like, in the in the current environment, it's like, basically GPUs are pretty sold out [00:49:29] S1: Mhmm. [00:49:30] S3: For the next few months. And for almost like the rest of the year, especially if you want to like stand up data centers in in The US. So I think this is actually like I think the the the main bottleneck to to scale right now for anyone in AI. And this is I think what we've been able to really scale through extremely well because we've basically partnered with almost every data center out there since, like, over two years. So basically, you're able to send up, like, data centers in in different geographies and orchestrate them globally. [00:49:59] S1: Yeah. [00:50:00] S0: Any predictions, forecasts for American open source over the next six months?
00:50:00 · Jordi Hays, Vincent Weisser, Rodrigo Liang
[00:50:00] S0: Any predictions, forecasts for American open source over the next six months? [00:50:07] S3: Yes. No. Really great question. I we joined also, like, NVIDIA to help them train Nematron. So they're doing the Nematron Alliance and Coalition. So we are one of, like, think 10 or so partners. So it's like so like like, I can't I can't speak about some of that like all the things there, but I think, like, in general, both on that front and other things we're saying [00:50:27] S4: We won't [00:50:28] S0: tell any we won't tell [00:50:29] S3: any. Wait. [00:50:30] S0: Wait. Anything Just between us. Between us. Just the three of So