I'm Vinay Shandal, and this is "Pitching Progress." Today we're talking about powering the AI boom.
There's 100 gigawatts of existing stranded capacity that just doesn't get used over the course of the year. That's more than enough power capacity to fuel all of the AI data center growth.
That's Varun Sivaram, founder and CEO of Emerald AI, a company looking to make AI data centers more flexible.
Emerald AI, our conductor platform, orchestrates all of these different approaches: temporal flexibility, spatial flexibility, resource flexibility, in order to deliver to the grid some relief right when it needs it the most.
No need for billions in new power generation. Just a piece of software that automatically modulates AI processing to distribute the load and avoid overburdening the grid. Could this be the day Ex Machina that solves the AI industry's power woes? Let's find out.
Give me the timeline in terms of when you went from climate diplomacy, to Orsted, to startup. It feels like it all happened in warp speed. Give us the chronology.
Well look, for the last 15 years, I've been firmly ensconced in energy space, seeking to build these multi-billion dollar energy projects. At Orsted, as you know, we built multi-billion, or multi $10 billion offshore wind projects, solar, storage, all over the world. And after those 15 years, I had some pattern recognition built up, and I realized what I'm seeing in the AI revolution is unlike anything I've ever seen in this sector. And I've had a front row seat, as you mentioned. I was fortunate enough to serve as America's lead clean energy diplomat. And so I traveled the world, focused on building more clean energy supply. And I realized that just building more supply, more energy supply, was only half the equation. And if we focused only on that half, we'd never be able to align AI's extraordinary energy demand growth with goals for communities to have affordable power prices and reliable and stable grids. Just building more stuff wasn't gonna get to there. And that's why I sort of left it all. I left 15 years just focused on building more energy, and instead founded Emerald AI to make energy demand more intelligent, which I founded in late 2024, so just 11 months ago. I wanted Emerald AI to focus exclusively on making AI data centers more intelligent in how they use energy, and the best friend a grid could have.
Well, Varun, look, I mean you're not alone in seeing that opportunity. Some fantastic pools of capital and individuals have rushed in to back you. I find it enormously impressive that in less than one year, is it 11 months? You've raised north of 40 million. Tell us more about what that journey's been like.
Look, 11 months ago, nobody was talking about AI data center flexibility. This is not on anyone's radar, right? And today, AI flexibility is a very hot topic. Energy secretary Chris Wright announced that he was directing the Federal Energy Regulatory Commission, FERC, to look at fast tracking data center power connections, if those data centers are flexible. Flexible AI data centers will soon have a leg up as a matter of national American policy. And so 11 months ago, nobody was talking about AI flexibility, and I was fortunate to have these backers who really took a chance, such as Google's Chief Scientist, Jeff Dean, Professor Fei-Fei Li at Stanford, one of the pioneering minds of AI. I was very fortunate to have folks both who are luminaries in AI, including Crusoe CEO, Chase Lochmiller, and also folks who are luminaries in the energy world. John Doerr, Secretary John Kerry, Prime Minister Malcolm Turnbull of Australia, and a range of others. Kate Brandt, Google's Chief Sustainability Officer. They all made this bet that this interface between the data center world, AI world, and the power grid, is going to become not only one of the most important interfaces in history, but one of the most valuable interfaces. And Emerald AI was founded to go and seize that interface and 11 months later, we're seeing this vindication that ah, the world has kind of figured out. And in those 11 months we've been hard at work, building the solution to meet the market demand that's exploded right this moment.
You've had how many pilots? And tell us about the recent announcement in Virginia.
The announcement that we just made on October 29th was with NVIDIA, as well as a range of other partners such as the Electric Power Association, EPRI, Digital Realty, PJM, the largest grid on the east coast. We made this announcement of the world's first power-flexible AI factory, or AI data center. NVIDIA is leading the charge, and building this 96-megawatt facility. It's a large commercial facility in Manassas, Virginia. And Emerald AI and our other partners are helping to make sure that that facility can be power flexible. What that means is, when the grid locally in Virginia, or on the PJM system is stressed out, that data center can actually provide relief.
So at Aurora and the Manassas, when PJM is tight, what does power flexible actually look like?
It can ramp down its power consumption intelligently, in order to help that grid get through that moment, avoid rolling blackouts, and at the same time, protect the performance of the AI workloads that are running. These will be next-generation, cutting-edge workloads, training, fine-tuning, using inference on AI large language models for applications in robotics or healthcare. But we're gonna make sure that those workloads have stringently high quality of performance while at the same time, meeting the energy grid's needs.
Why is the grid happier to have that flexible data center, versus the one that is not flexible? Just explain the advantage it creates, and also just how that works in practice.
First of all, let me make the point that you really don't need to do this very often. On the order of 1% per year or less, is when the grid is truly stressed out, when for example, it's a hot day in the summer, and everyone's running their air conditioner. And those peak load moments are the days that really stress the grid out. One thing you could do, which is our current approach, is the grid says, or the utility says, look, this data center, if it happened to consume at the absolute worst moment, its full allotted capacity would overload my grid and potentially crash it. Therefore I'm gonna make this data center wait 10 years while I build out the power lines and the substations and the transformers and maybe some new peaker power plants to meet that worst case load. But what if the data center, this AI factory, promises that during those rare moments, the data center is willing to actually ramp down its consumption, locally. If it's willing to do that, then you don't have to build out all of that new expensive infrastructure right this second in order to serve that new load. That new load can fit on your existing grid because there's 100 gigawatts of existing stranded capacity that just doesn't get used over the course of the year, and it could be unlocked. That's across the country, and that's more than enough power capacity to fuel all of the AI data center growth using BCG's own estimates over the next five to 10 years.
So in a peak alert, what slows, what moves, what are the knobs you're turning, Varun?
Some of these things are what we call batchable workloads. It doesn't really matter whether they complete right this second, like fine-tuning a large language model, let's say, or a deep research query that could be completed in 30 minutes or 35 minutes. You can take advantage of some of the inherent temporal flexibility to slow some things down, or pause some workloads, and reduce power. In addition, there will be some other AI workloads that are quite time-sensitive. You may be able to move those workloads from the stressed data center to a different data center that doesn't have that local grid in a stressed out condition. And by doing that, combining temporal flexibility with spatial flexibility, you can make data centers quite flexible. And on top of that, by the way, you might have local onsite resources like batteries. And those, what I call resource flexibility, can offer you even more. Emerald AI, that's what we do. We do the math. We have an AI running that behind the scenes autonomously to take care of it all, so the user doesn't notice, and the grid gets the relief when it needs it.
Look, a lot of startups show up with these concepts on paper. You've done a couple of pretty impressive road tests, but small scale. October 29th, big announcement. Tell us about what you're testing, what you need to prove, what needs to pan out for this technology and software to really begin to scale.
The announcement we made with NVIDIA of what's gonna launch next year. The world's first power-flexible AI factory that's a commercial scale, nearly 100 megawatt facility in Manassas, Virginia, called Aurora. And Aurora is gonna benefit from all of this road testing that we'll have done across the United States, and the United Kingdom, so that once we get there, we're going to be working with the regional grid and utility entities such as PJM and the Electric Power Association, EPRI, and their DCFlex initiative to test that this commercial scale facility can actually perform the way that it needs to to support the grid. We'll first prove this in what's called simulation mode, and then we'll graduate to actually flexing the power consumption. And this is a tight integration between NVIDIA, which has extremely advanced software, and Emerald AI, where we have worked very hard in close collaboration with NVIDIA's engineers, and NVIDIA's technology team to integrate Emerald software right into the stack, into what's called NVIDIA Mission Control. The goal is for future AI factories that are designed to NVIDIA and its ecosystem partner specifications. There will now be this reference, designer reference architecture for power flexible AI factory. So it's not just about this first commercial deployment, it's about the next thousand.
Varun seems to have some solid reasons to be optimistic about Emerald. They've got a piece of technology that can solve a trillion-dollar problem, and a partner in NVIDIA that's trusting them to implement it right now on a several billion dollar piece of infrastructure. But when it comes to business, nothing's a sure thing. And when it comes to working with large US utilities, nothing comes easy. To give us a better sense the value proposition, and challenges ahead. I turn to my colleague, Stevan Jovanovic, one of BCG's leading experts on power and utilities.
Stevan Jovanovich, welcome to the pod, man. I'm so excited to have you on this episode.
Delighted to be with you, Vinay.
So Stevan, look, when I talked to Varun, he is like, "Guys, don't worry about building new capacity. I got this. I'm going to help you access the existing capacity, so all these power demand curves that are out there, yes, we're gonna need that power, but that power already exists, and therefore the billions of dollars lining up to put CapEx in the ground to build new generation, we do not need." 'Cause I know, Stevan, you're out there helping IPPs build more capacity! So do you agree?
Let's just talk about the simple facts, right? So we have, in the United States, it used to be 1,000, probably now 1,200 gigawatts of installed capacity, right? That's like gas, coal, nuclear, a bunch of renewals, right? So 1,200, give or take. We need, various projections of between 80 and 100 gigs extra for data centers. The average utilization of the US power grid, probably between 40 and 50%. So if you were to perfectly levelize the load, we'd have 500 gigs or 600 gigs of stuff running continuously 24/7. And we've already built 1,200. And data centers are looking for another 100, right? So we could 5x what we need for the data centers if we could perfectly levelize the load. And so if you can avoid needing to run in those peak hours, there's ample existing generation capacity. And the crux of the problem is, you have a incredibly expensive chip, right? By the time you put that chip in, you just want to run every single second of the day. And so the only reason now that they're even entertaining flexibility is because they're being told that it'll take them four, five, seven, or 10 years before they can interconnect.
And forget interconnect, is also you can't build behind the meter because of political reasons, or because like the queue to get a gas turbine. And so yeah, I get it, yeah.
Right, what is it worth for me to build one on these, or come to market a year earlier, versus the penalty I take from not running 100% of the time? And I think we're finding that the answer here finally is penciling in, where a little bit of flexibility, that is worth it to sacrifice not running all the time. As soon as the supply side, so to speak, catches up, that business case falls apart.
But that catch up only happens if the power generation assets come online fast enough, which we all know is riddled with all kinds of challenges, yeah.
Exactly. And that's like 10 plus years away, right?
