Phaidra raises $50M+ Series B to build AI agents that autonomously run data center infrastructure

measurable. Say goodbye to the headaches of out of home advertising. Only ad quick combines technology out of home expertise and data to enable efficient seamless ad buying across the globe. Do we have our next guest? We have our next guest. You can see him right there on the on the big. There we go.

Welcome to the show. How are you doing Jim? Good to meet you. Hello John. Hello Yord. Uh doing very well over here. How about yourselves? We're doing fantastically today. Uh bit of crazy day with Sora too taking over the news. Uh so we spent a lot of time talking about that. But what's new in your world? Yeah.

Well, uh, we raised a series B, which is of course the big news in our world over here. And how much founder? Yeah. So, we raised bit north of $50 million to build new AI agents for AI factories. AI agents for AI factories. Let's go. We've been waiting for this. What does that mean? Yes.

Jensen says AI factory is a token factory. It's just a data center. Or do you mean AI factory as there's robots walking around making iPhone cases or something like that? Yeah. Now, we we mean it in the Jensen sense, right? So purpose-built mega scale data centers that exist to convert electricity into tokens, right?

They're they're they exist to generate intelligence. That's what it means. So concretely, we make AI agents that are capable of autonomously operating and optimizing physical mission critical infrastructure.

What are the top things that kind of break when you're running a large data center that normally a human has to step into? Normally it's time consuming.

Now they're leveraging an AI agent to maybe just work while they sleep or kick off a, you know, a a problem solution so they can have the agent go and work and solve some of the problem for them. Uh what's the what are some examples of like problems that you're solving concretely? Yeah. Yeah.

So you have to pardon me if I get a little bit nerdy. I'm I'm a mechanical engineer by get nerdier. Go to town. Yeah. So I think one of the the the important things to realize is that AI factories today look nothing like traditional data centers, right?

Like when I started in the industry, you know, I joined a much smaller Google back in 2010, right? And our concept of a large data center were these 30 megawatt modular data centers. And at the time we thought to ourselves, who's ever going to use 30 megawatts of compute, right?

And now obviously the industry has changed a lot. Now we talk about gigawatt scale data centers, right? Purpose-built for AI workloads, i. e. AI factories. The most important thing to realize about AI factories is that they're orders of magnitude more complex. There's a lot of stuff in there to orchestrate.

A lot of pumps, a lot of chillers, a lot of liquid cooling CDUs, right? And they have very nonlinear interactions with each other. So what our AI agents do is they have a global view of everything that's happening across AI factory.

They can absorb tens of thousands, hundreds of thousands of sensor trends in real time and they act like a general on the battlefield, right? They're issuing AI generated command signals that are sent to the local control system for automatic implementation execution. Right?

So it's really what we would call AI enabled supervisory control, right? For these very large AI factories, where do you think the So if you're doing the orchestration and management, where is the uh physical work going to come from? Is that being met by, you know, somebody that's just on call in a data center today?

Do you think that robots will eventually do that work? Data centers are great and that they're sort of a flat controlled environment. It's kind of a a good setup for, you know, not necessarily a humanoid robot, but just a robot on wheels that can go around and and make adjustments.

Um but but what's your kind of vision for um once your agents decide, you know, what needs to be done within their system? If it can't be done purely with software, how does it get done physically? Yeah, totally. It's a good question.

So, you know, to be very clear, DI agents are not capable of replacing like the physical labor that needs to happen. So, things like turning wrenches, right? Taking a chiller offline, cleaning it, that's stuff that, you know, you need humans for today, right? Maybe eventually we'll see robots, right?

Taking care of that sort of stuff. But what we do is more of the the software level orchestration of all the of all the uh equipment, especially the mechanical cooling systems, right? Um that happen in these large AI factories.

Uh but I think you know one of the important things to also rec you know recognize is that the industry is growing so rapidly right now that there is just a massive massive massive shortage of skilled labor you know expertise in the industry right Jensen talks about how like electricians and plumbers are going to be two of the hottest trades moving forward because of the scale of the AI factory buildout in the industry that I come from you know you see a lot of folks with white hair and you don't see a lot of folks our age in the industry right so there is a massive shortage a lot of people retiring, they're taking knowledge with them.

Our hope is that these AI agents in addition to providing optimal functionality to these AI factories can also augment the labor force, right? We we communicate our AI agents as virtual plant operators, right? They're virtual members of the operation staff.

You can't see them or hear them, but just like me, they're permanently working from home, right, and doing a lot of stuff in the background. Can you explain to me what the software stack looks like for all the different pieces of a data center?

Like uh we we've talked to a lot of uh folks who are building software for traditional manufacturing and they talk about uh Seammens has a control system that has some sort of archaic API and then they write uh you know a wrapper for that API and then they can create a dashboard or they might be able to trigger some things on that.

Do do does a chiller have an API that can be accessed over Ethernet or something like how do all these systems tie together? How open is it?

Like whenever you're building an AI agent, people will comp you to like cursor, but cursor is a fork of VS code an open-source software and then all the code is stored in GitHub, very easy to scrape in and download and work on.

Uh, I imagine some of these tools are either closed source because that's part of the business strategy of that, you know, industrial company or they're just so old that they haven't gotten around to actually writing an API. So, walk me through all that. Yeah. Yeah. It's it's a great question.

So, you know, frankly, the the data center industry hasn't seen a lot of innovation a really long time, right? Until Nvidia came out with, you know, the latest generation of chips. And now all of a sudden, right, there's real innovation, you know, real creativity in the data center industry again.

Before is really just copy and pasting the same old stuff over and over again, right? A large part of that is the closed ecosystem effect that you were just describing, right? If you look at what the traditional incumbents are doing like the seammens, the shide electrics, whatever, right? Um, you know, yes, right?

There are protocols that exist like Modbus or you know or opcua, MQTT, whatever protocols that exist that act as the APIs right that enable these uh uh large mechanical equipment to coordinate with each other. However, that tends to exist within a walled garden.

So each of these major automation vendors, right, like the Schneider Electrics of the world, they make it very difficult for you to integrate, right? There's no such thing as, you know, an an easy way for third parties to access the data and analytics that are coming off of these mission critical systems.

So, a large part of what we do is integrating with a large variety of these building management systems or these SCADA systems to ingest that data and make intelligent decisions off of it. This seems like something that you can't just pick up from a couple Google searches. Like, what were you doing before?

How did you actually learn that this problem existed? How did you get up to speed on all these different systems? Yeah, so uh in 2010 I was a brighteyed and bushy tailed college grad joining Google.

And back then Google was still like small enough and young enough that new college grads had outrageous amounts of responsibility, right? So I was fortunate enough that I actually uh helped design some of Google's first, you know, large cooling systems that go into their data centers. Wow.

Then um in 2013 I became the person responsible for Google's energy efficiency programs for its data centers. The primary metric is pee or power usage effectiveness. Back even back then like Google was already consuming billions of dollars a year in electricity.

So it is in our incentive to optimize the energy costs right and that's grown by orders of magnitude since then in 2016 something called Alph Go happened. If you remember that's kind of what made Deine famous. My uh co-founder Veta was one of the Alph Go engineers.

I saw what happened and I thought to myself, if we can create AI agents that beat the smartest, most intelligent humans at complex games like Go, then surely we can teach these same AI agents to play other games like let's reduce the energy consumption of Google's data science management.

So yeah, so I reached out to a guy named Mustafa Sullyman who was the co-founder of deep mind. Right. Let's go. Chad is loving you by the way. said, "This guy's cracked. This guy rips. This guy ripped. Everybody loves you. Thank you for joining the stream. " Almost like your entire life was leading up to your moment.

Overnight success. So, it was I mean it was just a cold email. I was totally not expecting anything of it. But I pitched this idea of like what if we take Deep Mind's reinforcement learning agents and use them to control big ass infrastructure like Google's data centers. So, we did it. It worked.

It worked surprisingly well. I actually ended up joining DeepMind for three years. I I led a business vertical called Deep Mind Energy, right? We applied reinforcement learning for all sorts of things like commercial building HVC.

Then my co-founders and I realized reinforcement learning is actually really good at industrial scale control and optimization. So we started Fedra. Well, congratulations. What a run. Very excited for you and uh glad to see you have a nice series B to go. I got a feeling you you'll be back on the show in not too long.

So, congrats on all the progress. Great to meet you. Have a great rest of your day. Thank you very much. Have an awesome day, you guys. We'll talk to you soon. Well, if you want to optimize your wrist, head over