With the proliferation of artificial intelligence use across every sector and industry, proper planning and resource management has become increasingly important for data center developers and hyperscalers. While data center-focused acquisitions are bringing in billions of dollars, increased public pushback recently resulted in New York becoming the first U.S. state to issue a one-year data center moratorium.
A recent study from the Kansas Health Institute found that U.S. data centers’ electricity usage could more than double by 2028, and the National Electrical Manufacturers Association projects 300% growth in U.S. data center electricity consumption over the next decade.
However, despite increased pushback and load growth expectations, Pado AI founder and CEO Wannie Park told ESG Dive that legacy data centers — older, traditional on-premises computing facilities — provide an opportunity to optimize existing technology rather than create new builds. Pado, which spun out of LG Electronics, provides energy management software for data center developers.
“There is so much sort of underutilized capacity on the power . . . side of things, so when the entire narrative is ‘grow, grow, grow, break ground’ [or] ‘oh my gosh, moratoriums, the community is pissed off,’ just recognize that none of that is affecting the legacy data center market, which has thousands and thousands of sites and incredible amount of underutilized power,” Park said. “There should be thoughtfulness in terms of how to allow the U.S. markets to use that underutilized capacity to catch up.”
Editor’s note: This interview has been edited for length and clarity.
ESG DIVE: How have the conversations Pado has had around AI and energy management evolved over the past couple years, and how has that coincided with the spinout from LG?
WANNIE PARK: When we started this company and started chasing problems to solve, I think the fundamental thing was to find where those peaky volatile energy loads were. At least in my career, I think I've seen [volatile energy loads like this] twice, and this is the third time. We saw similar patterns in each.
The first time was probably back in 2006 to 2008, when the Obama administration unlocked funds to modernize the U.S. grid, and that created all kinds of havoc. Not that different, honestly — maybe at a smaller scale than what we're seeing in terms of data center development and moratoriums and community engagement, all that stuff.
The second time was when all the solar and EVs were coming onto the market, and there was a term that the industry dubbed the duck curve. When people come home and they start charging their cars or solar goes dead, like there was a volatile sort of pattern when it comes to the electric load. Same sort of thing.
Data centers are exactly the same. You get all these inference jobs — something happens in the World Cup finals or whatever — and you get this spike. It's volatile, and so that was the nexus of “we're seeing this third wave of peak volatility when it comes to power, is the industry ready to solve it?” And the answer was no.
So, when we went to LG in terms of infrastructure, or even other companies, we basically said, your hardware, your cooling systems, your storage systems, all that infrastructure is going to be so incredibly valuable. You just don't have connective tissue. So, that was how we got our foot in the door, and that's how we got funded and really activated the company.
What have you learned through software deployments so far, and what surprised you the most during that process?
There's a couple of things. I'm used to going into markets where everyone's doing it, there's legacy solutions and things like that. This is not the case with AI. We went into this thing thinking, “Man, there's going to be a lot of folks that are just completely on point.”
What we've actually found is that the industry has been slow to change, and what they tried to do is retrofit existing technologies to solve this new need. That need, through Pado's lens, would be what we consider either grid-aware compute power, meaning how do you manage your computing, your GPUs based on power — power coming in, power on the grid, renewables. All that stuff did not exist when we launched, so that was very surprising.
The second piece is just sort of the acknowledgment by the industry saying “We are very heavily underutilizing our assets that we're spending billions upon billions of dollars as an industry on.” So, the best example I can give you — there's research out there, and we have validated it ourselves — let's say you were to buy 100 Graphic Processing Units for your data center. The nameplate value would be a certain amount of power, a certain amount of compute, things like that. However, they're only getting 12%, or only 12 of those chips are actually being utilized at their maximum level. So, there's like this 80-90% headroom of underutilization, and the industry is trying to find a way to address that.
But historically, the industry has responded by saying, "Well, we just need to build 10 times more data centers,” versus let's take that 12 and get it up to 20, and then get it up to 30.
It's not easy building a data center, and we're not talking about little baby data centers. We're talking about data centers that are like the size of a city when it comes to a power footprint, and all the hurdles you have to go through seem like that's a very hard way to grow a market versus [working to get more of a return] from existing capacity and assets.
Where do legacy data centers fit in?
Legacy data centers are generally data centers that are 100 megawatts or smaller. But the real sweet spot, frankly, is probably between like 15 to 50 megawatts.
Historically, data centers started in the IT closet, blowing fans at the server to keep things cool. So we've evolved from there to like something that's more traditional, and a lot of it was driven by the cloud. Then the question was what is this thing doing? And it was typically doing enterprise-type work.
Now, all of a sudden, well [there’s money and growth] in AI stuff. And so, for these legacy data centers, they don't have the latest gen chipsets. What they have is maybe an older generation, or they may even have just traditional CPUs. To process and deliver inference jobs — which would be like a query, like you're asking ChatGPT [or another model], "What's the weather like tomorrow?” The existing infrastructure can handle most of that, and you don't need to invest billions of dollars in the next gen cooling systems. Maybe it's not as sustainable, but as a bridge, you can still do the air blowing.
