Feeding the Beast -Inside the High Stakes Race to Power AI (Condensed Version)

Transcript

SPEAKER_01 0:00

Secondly, we need to recognize, I would say, that uh all data centers are not created equal. Some of them are going to be uh significant amounts of capacity, a gigawatt or more, coming online perhaps all at the same time, but others are going to be a bit intermittent in terms of their need for power or come online incrementally. So you need to differentiate and not just sort of think of data centers in a monolithic way.

SPEAKER_00 0:25

And the biggest biggest challenge is that the extra power needed to fund the activity, to run the activity, to energize the activity. On top of it, uh, if you take an average uh data center design compared to what was five years ago, the compute need per rack is 10 to 12 fold. For example, each rack used to be about 100 kilowatt, now it's 110 kilowatt. Um 10 kilowatt before, now it's 110 kilowatt. 120 kilowatt is the demand per rack. So you can see because of the density of the computing and the ability to uh not tolerate um any kind of loss of power, all of that leads to a power density that was unheard of. That is what leads to the actual magnitude of the demand.

SPEAKER_01 1:12

The whole system really has to adapt to this the speed. Traditionally, the regulators, I'll return to them for here for just a second, the regulators have um have been able to receive from the utilities the uh the costs that they have expended, you know, for putting in place new facilities and approve or not in some cases, or or maybe reduce the amount that the utilities could collect, generally approved if those expenses were were approved in advance. But um, the luxury of being able to wait until the money is spent and then have the rates adjust uh is slipping away, if not gone. The the regulators and the utilities are going to need to be now looking proactively at what the expenditure is going to be in the next year or two because the regulatory pro you don't have time as much time for the regulatory process to work its way through. Long hearings, these rate proceedings can sometimes last a year or two. And by that time, you know, the costs that you might have been required to absorb to build out the system could be at risk and what's being demanded.

SPEAKER_00 2:16

Everything needs its own addressal process, right? So power availability, power supply, power delivery. That is one aspect that needs serious look, right? Um, same thing goes with the water, is the other natural, naturally occurring resource that's needed. Uh, whether it's even if you go all liquid for rack cooling, you still have area cooling and uh other processes that required lots of water. So the water is the other piece that requires serious planning, serious allocation, serious estimation. Right? And then there is um there's three or four other things. One is I would consider the ability to project the model the demand, growth. What is it now? Why is it now this way? What is it like to look for, look like in a couple of years, five years or seven years from now? That will decide the ability to pre-plan some of these capacity uh constraints that we have today. Then there is this other human aspect, workforce for that area, uh, mechanics, electricians, I people who can do data center work, design, maintenance. I'm not sure if we are all fully ready to have the right amount of workforce ready and certified, ready to go to work. You might need a lot of retraining, a lot of incentive making through maybe Department of Labor. There's a this workforce piece as well, just like cybersecurity, had a big big demand for people that are trained, cyber aware. It's gonna need uh people that do this kind of work.

SPEAKER_01 3:44

You can see that. Great, sure. Okay. So what this shows is this is a projection by the North American Electric Reliability Corporation or NURC, uh, which is the entity tasked uh under FERC jurisdiction with uh maintaining and and signaling the risks to reliability as we look out into the future. And it's pretty obvious just from the red that you can see they are projecting as of uh late last year that there are high risks in significant swaths of the country, including the region of the what's called the PJM, Pennsylvania, Jersey, Maryland area, which is the largest uh electricity market in the country, if not the world, large parts of the Midwest ISO, ERCOT, um, and out of the Northwest. And most of this risk to reliability is driven by the data center phenomenon.

SPEAKER_00 4:34

And I would consider this to be a case of classic change management. So if you can draw a direct line of benefit from what's needed to what it'll mean for everybody. So if let's say compute capacity growth leads to better outcome economically, more GDP, more transactions, more stuff getting done faster, i.e., productivity. This if you can map cross cross-map those benefits, uh, it also means there's also uh an incentive for creating in um innovation, better solutions, more innovative approaches, or utilization of spare capacity that might be sitting around. How do you get them from here to here? Um, and then third part will be creating models to uh show benefit of planning, uh the uh the model of concurrent needs and how they can be met. Pretty much like how AI figures out something, right? If you give multiple factors, say, give me the best optimal solution, AI finds a solution for you, typically, right? Give us given uh best uh inputs available, can give you a recommended course of action. Probably we can put AI to work to also find better ways to do this and connect the benefits uh to what it needs. Um I would say that add a couple of points.

SPEAKER_01 5:47

Um, you know, the heating demand, the the heating heat that's produced, uh, which leads to the requirement for cooling, which is a lot of the water use in these facilities. The data center developers are also beginning to think creatively, more creatively about that. I mean, that heat can be used to um uh to create to produce more electricity, right? You can siphon off that heat and use it to meet some of your on site demand. So, you know, that's also uh I think a way that uh they're looking at adapting.