Good afternoon everyone. Thanks for coming here to the Nakamoto stage. Today our panel is about AI HPC as well as Bitcoin mining. How they both can help the grids or the energy infrastructure for better. And we are joined by two companies who are at the front lines of this transition between Bitcoin miners to AI HPC and everything in between. And we will be talking about a lot of things, but I will start with a quick round of introduction. My name is Nishant. I run a consulting firm called Bloxbridge. We own a media firm called The Energy Mag which covers Bitcoin mining space as well as now increasingly the energy infrastructure space. I'll let you take it, Matt.
Thank you. Matt Schultz. CleanSpark is a Bitcoin miner. We recently announced that we plan to expand into AI data centers. We have 35 locations across seven states and about 1.8 gigawatts of active power capacity.
Great. Fred Thiel, Chairman, CEO of MARA Holdings. We're one of the leading Bitcoin miners. 18 sites across four continents, about 1.1 gigawatts of power energized.
So when people talk about Bitcoin mining and AI HPC, especially in the context of power, they often think of them as competing for the same capacity. They think of one versus the other, but as we have discussed before and as we read, they can complement each other in many ways. And we would like to know more about that. And so I would like to hear from the panelists here how they complement each other, starting with Matt.
Yeah. So, Bitcoin mining is very unique in the fact that it's interruptible. Anytime that there's an act of God, an event that requires power elsewhere, or an issue with temperature that drives utility costs up, we can curtail our power very rapidly and the utilities love that. In fact, Duke University published a study called Rethinking Load Growth and it says that right now today there's between 75 and 125 gigawatts of stranded capacity on the US grid over the 22 largest utilities. The challenge is that it has to be interruptible one half to one and a half percent of the time. So what that means really is if you can shut off your power for 117 hours a year, it unlocks 100 gigawatts of capacity. Well, manufacturing facilities and traditional AI data centers are generally fixed loads. They're not interruptible. But Bitcoin mining is in fact an easily interruptible load. And I know MARA as well as CleanSpark frequently curtail our loads for the benefit of the utility grids and the communities we operate in. So the idea of securing a customer, hyperscaler, a Mag 7, whoever that has a requirement for a certain amount of energy, the capacity there they have, it's what's called the PUE, the Power Usage Effectiveness. And so they publish a peak number that says we consume X amount of megawatts assuming that it's the hottest day of the year, most humid, we have all of the systems running. But over the course of a year, the average PUE is much lower. So when you think about if you're a utility, one of our sites, I'll give you an example, 250 megawatts of power. To convert that to an AI HPC data center, you take the reduction in the PUE to provide the critical load. So you have 50 or 60 megawatts of capacity that only gets used five or 6% of the time. So the idea with the fact that they're complementary is that you can mine Bitcoin in that PUE. So when the power is available, you can mine and when it's not, you curtail and you push it back to support the data center. So it's actually a very complementary process.
Interesting. So, you mentioned that the Bitcoin mining helps keep the PUE constant or fill up the gap in between what PUE you hope to keep and what you sometimes end up getting. So Fred, maybe you can inform more to the audience about how the PUE is different between Bitcoin mining as well as for AI HPC, and also touching on how the overhead is different and then there's the PUE coming from the computation itself, how power dense the computation is. Because just from a very intuitive thought you would believe that Bitcoin mining would have a very low PUE because it doesn't have any overhead on cooling or air conditioners and all that that you need for a data center for AI HPC, but actually in the end the PUE often ends up being lower sometimes because of other factors. So we would love to hear and I think the audience will find it interesting to know what are all these factors that the PUE depends on.
Yeah. So I think you have to look at the nature of the load. So, Bitcoin mining is a fixed load in the sense that you turn it on, it runs, and then you only turn it off or turn it down if there's a demand from the grid that you curtail. As Matt was explaining, if you're running training loads, training is constant, then there's a drop, then it's constant, then there's a drop, it's constant, then there's a drop. And in that drop, you're not taking power necessarily. In inference, however, you have a load that is shaped by really the number of queries you're getting. So, as an example, if you're an Anthropic user in Europe, you have a very different token allocation for when you're running after 5:00 p.m. Eastern US time than if you're running before that. Why? Because the US capacity on Anthropic's inference systems is so high. And when US offices shut down, then that load drops and Europe can consume more of that capacity. So they're trying to load balance this curve of demand that they're getting based on queries. And if you marry Bitcoin to that, it works like a perfect way to create a fixed load, which to Matt's point, the power utilities want to have just this flat line. That's their dream. And Bitcoin mining can do that. Plus, Bitcoin mining is flexible. It can shut off. So, if you marry Bitcoin mining and inference and you can load balance in the data center, you provide the best of both worlds both to the inference company because you're balancing their load, lowering their cost of electricity. As a Bitcoin miner, you're mining with a subsidized cost of electricity because they're paying you for that service and the power company loves you.
So, you talked about how Bitcoin mining helps keep the load constant. We both, you both mentioned that and you showed different angles on how that helps the grid. And the grid loves flexible loads. Why is that? Why does the grid like flexible loads?
Well, I mean the grid likes flexible loads because if you think about how the grid operates, it's like the plumbing in your house, right? If you had a certain amount of water pressure coming in and you open your taps, now you're taking water pressure off the water utility system. But if you shut the tap, now that pressure is building in your system. And so if the water company kept pushing all that water at you, your taps would explode. The grid works the same way. If the power generators just constantly generated and didn't regulate their supply to the demand, transformers would blow up. The electrical grid would just short out. And so the utility companies are constantly trying to figure out how much power will people be using over the course of the day. And it varies. There's a thing called the duck curve, which is like the belly of a duck. Electricity consumption peaks at 9:00 a.m., then it drops midday, and then it starts scaling up at 3:00 p.m. to 9:00 p.m. and then it drops again. And so they have to match generation capacity to that consumption, that offtake. And Bitcoin miners are perfect partners because the power utility can say, I need you tomorrow between 9 and 12 to take all the power you can, and then after 3, I need you to scale back about 20% of your power. And so we are a perfect partner because we can act as that demand response mechanism for them. AI can't do that. AI wants to run constantly. So AI is what's called base load demand whereas Bitcoin mining is what's called flexible demand in that it can be shaped. And Matt mentioned the Duke study, curtailing one and a half to 2% of the time a day. That's the only time when the US grid uses maximum generation capacity and if you operate outside of that there's almost 40 to 50% of US grid capacity available and so that's where Bitcoin miners traditionally operate. And I think the AI industry, if you marry Bitcoin mining and AI, all of a sudden you can solve a lot of the energy challenges but the AI industry has to adapt to a world where they're dealing with more flexible loads.
You know, to add on Fred's point, a lot of the jurisdictions we operate have hydro and nuclear and those are always on. And so there's this narrative that Bitcoin mining and data centers drive utility rates up. Well, the issue is once it's built and deployed, especially for those particular types of generation, I mean, Fred owns a wind farm. It's the same way. When the wind's blowing, the power is being generated and if it's not being used, it's being lost. So the idea is that if you can design consumption that meets the generation, now you're monetizing all of that power that's being generated for the benefit of the community, for the utility because that revenue would otherwise go unrealized. And so now it's power going to ground rather than revenue generation that can be reinvested in new capex and new types of generations. So it really becomes a flywheel if you can manage it and balance it properly.
So you both talked about curtailment and that's a very important point because as you both pointed out that AI demand is not so curtailable because people need their ChatGPTs running. They want their inferences all the time. Whereas Bitcoin mining you can just turn it off, turn it on and you can curtail. Now I wonder how anyone can measure how much curtailment flexibility is enough for the grid or the utility operator.
So the utility planners on a daily and weekly basis estimate how much demand there will be for power on their grid based on weather, based on day of the week, seasonality. They have years of history and they know that May 29th is typically a day of this temperature, so people are going to run air conditioning or not. So they have a lot of historical demand data that they can look at and so they forecast what their demand is and they then essentially purchase from generation that amount of power. But the fact of the matter is if you have renewable energy on the grid, meaning wind and solar, well, it may be cloudy that day and yet you'd planned for it to be sunny. So now you're running short of power in the grid. So maybe you then need the Bitcoin miners to curtail a little. Or maybe the wind is blowing more than you expected and now you have more energy all of a sudden on the grid and so now they're having to curtail something else. And so the use of renewable energy has exacerbated the problem for the utilities of balancing load. And if you look at in ERCOT for example, you had Winter Storm Uri a few years ago which was a total disaster for the state of Texas. And this winter you had another really bad storm but you had no disaster. Why? Because you had 2 gigawatts of Bitcoin miners that basically could shut down and give all that power back to the grid. And so they didn't have that crisis situation. And so all that renewable energy causes a lot of inaccurate estimates regarding power generation capacity that have to be modulated in real time. And you can't use a nuclear reactor to modulate your electricity because it's like you can't just put the accelerator button on a nuclear reactor and say up capacity by 10%. So you typically have to use thermal which are peaker plants which is like a jet engine in a building and you turn it on for two hours and you turn it off and then you can charge the grid a huge amount of money for that excess energy. Well, Bitcoin miners offset that. So teaching AI companies to build systems that can operate together with Bitcoin miners so that the Bitcoin miner can act as that curtailment buffer to the grid and balance the consumption of the AI load will be a huge innovation. And I think the other thing you're starting to see now is for inference especially, the data center operators are willing to come off the tier 4 data center requirement of five 9s resiliency because they can have multiple data centers that provide redundancy for each other and therefore you can save cost on build which means you can now do more of a modular model more like Bitcoin mining and I think you're going to see the two industries become more and more similar in what they do.
