TensorWave raises $350M Series B to build AMD-exclusive AI supercomputing infrastructure

Uh, what a fantastic feeder for a mafia. The SpaceX mafia is something we're going to be hearing a lot about in the coming years, especially post IPO, which happens tomorrow. So, tune in. Our next guest is Jeff from TensorWave. He's the co-founder and CGO. How are you doing, Jeff? Welcome to the show.

Hey, doing good. John, how are you

doing? Oh, great.

Thank you so much. Since it is your first time on the show, uh, please introduce yourself and tell us a little bit about the company.

Yeah, I'm Jeff Dartuk, co-founder and chief growth officer, like to say chief GPU officer here at TensorWave. Our focus is on um building out AI infrastructure exclusively with AMD.

Okay. So, uh I mean chief growth officer that feels like a sales revenue goal. chief GPU officer feels like an operations, construction, uh, manufacturing, racking, buying like the the other side of the business. Do you touch both? How deeply integrated are those two?

More GPUs and powered shells, more revenue.

That's right. Everybody needs GPUs and powered shells and uh getting access to them and I I'm the gatekeeper.

Okay. So, lay down the AMD thesis. Uh, we can go in that, but let's just start with like how you became interested in AMD. Why now? Why? What's the traction? What's the evidence that you're seeing that this is the right choice for your business?

Yeah. So, when AI took off back in 2022, 2023 and there was the big supply constraint um around getting access to Nvidia chips. Uh the company that we had started previously uh we were deploying FPGAAS at scale and so we're working with XYlinks and Altera chips and primarily uh Alterara primarily XYlinks and when Xylink got acquired by AMD our previous company was um working very closely with AMD. they would send us all of the latest and greatest chips and uh we'd get them deployed and debugged for them at scale. And so that's how we forged this relationship early on with AMD. And we saw as soon as there was the supply constraint on the Nvidia side um the opportunity when AMD announced their GPUs back in 2023 for us to be the first to go to market um with their chips at scale.

And how are you talking to your customers about what AMD can offer? Like we've seen uh semi analysis has uh inference x formerly inference max and had some really unique data showing that for certain models they haven't been able to benchmark everything especially the proprietary models but for certain open source architectures AMD is actually maybe more cost-effective on a token per dollar basis but uh what data are you leaning on what questions are people asking you and then I'm super interested to understand um how companies are thinking about uh where the integration if they do want to migrate from CUDA uh over to AMD like what goes into that how costly is that it was seen as impenetrable impossible now it's looking more and more feasible but what are companies saying uh about the trade-offs involved in being on AMD

yeah so in the very beginning uh when the supply constraint issues happened people were just looking for more options and AMD um was the only option that was available and um so at the very beginning people wanted to just try it and see if it worked and so AMD when they first launched um launched with the emphasis on scaling out uh scaling out inference at scale

and so being able to support inference and support some of the hyperscalers um was their first focus and then as more labs started getting on and testing it um they realized that there were some gaps in the software um you know it seemed like the hardware story had gotten out ahead of the software story and so that's where we were able to uh help lean in uh with AMD to help fill some of those gaps by giving them access um and helping debug things along the way. And so uh yeah, it the the frustration that customers had around, you know, not wanting to be locked into one vendor, not giving all of their margin to Jensen forced them and they were willing to look at other options that were available on the market. And so, uh, we were there, you know, every step of the way to help fill some of those gaps as, uh, customers were considering what is that leap going from CUDA to Rockum on AMD.

And, and the big public story that played out around that time, you're probably referring to George H. Hot, flagging some bugs, eventually getting on the phone with Lisa Sue, going full founder mode in the CEO seat, and watching Rockm develop and actually crush a lot of those bugs. Obviously, some people have flawless experiences. Some people probably still run into problems every once in a while. That's going to happen. But, uh, a really good turning of the tide, really good different way to interact with the developer community. Um, I'm interested to know uh like how how big do you have to be to go multi-platform now? Do you have to be one of the, you know, trillion dollar labs or are there smaller companies that say, "Hey, we have this, you know, this model that we're fine-tuning. It's pretty good. It's built on some open- source architecture and we'd like the flexibility to not be locked into one chip. But is that a million-doll endeavor like an, you know, a year of 10 uh researchers time? like how are we bucketing this idea of of uh diversifying the the options for a company?

Yeah, that was a question that a lot of um companies had. What is the switching cost of going from CUDA going to Rockom?

And uh we were able to show I mean in the early days right when back in back in 2023 data bricks was able to show um that you could switch from uh you can go from Nvidia and switch to AMD and it works out of the box in the early days. And so being able to showcase that and focus on the most important use cases that the majority of AI companies are using and optimizing those things over time uh was AMD's focus. And Lisa did a great job like you mentioned she did go into founder mode. Uh she lit a fire over there to make sure that her team was emphasizing building the developer community building out the resources necessary. And as Anous over there who's the CVP of AI

Yeah. He uh I love what he says. He says speed has become the mode and they've been able to move even faster now and with all these coding codegen tools that are available they're able to even move even faster uh to help fill uh and close the gap um on that uh supposed coupeote.

Yeah. A lot of times people uh sort of put AI as like a a monolith and uh I mean we were talking yesterday about uh how just in the customer support we're talking to Brett Taylor at Sierra just in the customer support world um you know uh speed might not matter for agentically going and solving a problem in the back end after a customer support ticket close. If you're on the phone with someone and you need a real-time voice model, it might actually make sense to colllocate that or distribute that. We talked to Matthew Prince from Cloudflare about this that there are certain models that need to be in certain places. Speed might be more important. Others is just raw horsepower. It's going to take an hour anyway, send it off to the big data center. But I'm interested in in the other pockets like where is is AMD poised already in the lead or poised to take the lead in images or video or voice or real time or agentic work or integration between the GPU and the CPU? Like is there something where you're particularly excited? I mean I noticed Luma AI is a customer. They obviously do creative work and I'm just wondering if there's something where you're like, oh yeah, like they've they've nailed like this particular architecture is great and like obviously there's other work to be done. There's more cost on certain workflows because those were developed at DeepMind on TPU or at some other company on Nvidia. Um but uh is there any place where you're like ah AMD's really nailed it in this in this subcategory of AI? Yeah. Well, they started off great on the hardware side um and focusing on adding uh more HPM than what you can find on a comparable Nvidia chip. And so having more HBM allows for workloads that are running larger models uh to run on fewer GPUs um and to do it faster.

Yeah. Uh so you have the CUDA moat uh as as you put it. Uh is it is there like a who who's building the bridge? like you need a bridge to go over remote. Like I imagine that there's companies, you're probably among them, that are specifically building software that can just translate any architecture from Nvidia to uh to AMD. Like how close are we to just that being one click of a button?

Yeah. Uh well, one of the things that I found as we were digging into um you know, understanding more about what Nvidia was doing, understanding what that CUDA mode actually was and and talking to a lot of the original engineers that helped build CUDA from the ground up.

Yeah.

Um one of the guys that's on our team, Greg Damos, was was there in the very beginning building CUDA at NVIDIA

and he had always said that CUDA was always built to go beyond CUDA and to to scale out infrastructure. And so um one of the things that that we've as we've talked to people they said that CUDA is not the moat at Nvidia their ecosystem is the moat and the resources and developers and libraries that they have poured into over the last 1015 years is the moat and uh that's the maturity that exists today and so as AMD is building that out and as we are supporting them we find the important thing is to focus on uh building out an ecosystem uh to support the uh to support AMD. uh we don't focus on software ourselves but we have a lot of other companies that are trying to solve that problem to create a heterogeneous stack to be able to switch um between chips and uh so bringing together those people showcasing all the different layers that are uh existing

beyond uh the Nvidia ecosystem is is what uh my current focus is on today.

Where uh where are you setting up your supercomputers? Some people call them data centers. We call them supercomputers. Yeah, we we have supercomputers all over. Um we have in Arizona, Florida, Pittsburgh, and uh have data centers going up all over North America currently.

Amazing. Awesome. Tell us about the round. How much did you raise?

Yeah, we raised 350 million series B. Here we go.

Congratulations.

Thank you so much for taking the time to come chat with us.

Yeah, great to meet you, Jeff.

Great to meet the team on the progress.

Thanks, everyone.

Have a good one.