Invest Like the Best with Patrick O'Shaughnessy
· Patrick O'Shaughnessy, Matthew Smith
[00:00:00] S0: What kinds of companies are these? [00:00:00] S1: Just like Caterpillar. Caterpillar, I think, is doubling their Solar Turbines capacity between now and the end of '29, which I would judge is at the exact wrong time when people may be questioning whether they even want to deploy those assets because the gas is much more expensive than they planned. Bloom Energy, they've been topical recently because of other things that folks are talking about, the rare earths they use in their manufacture, for instance. But for us, we don't think that Bloom Energy's assets at two gigawatts or more will be able to get natural gas in competition with all of the other assets that are being deployed that will consume gas given the scarcity that we see. So those are two, we'll call it manufacturers of distributed generation or BTM generation that we think are probably more poorly positioned than investors appreciate. It really doesn't make sense to us, beyond 2029-2030, to build large-scale natural gas generation until we ramp up production meaningfully and you can make sure we have the security of deliverability of supply of gas that's consistent with our model. We could see orders slow very meaningfully for natural gas generating assets, even at large scale, as '26 progresses. And those could be some of the losers — it may not make sense to use gas for power generation for new or incremental assets after a certain point.
· Patrick O'Shaughnessy
[00:00:00] S0: It seems like this whole memory shortage thing that we're going through right now, that hyperscalers might also be in trouble here if this is a key input to what they're doing. Do you think that's a big problem for them?
· Matthew Smith
[00:00:00] S1: So as we've been socializing this a little bit, trying to learn more and have people poke holes, I've shared this with one of your recent guests and he listened to this. He said, well, this sounds like DRAM two years ago. Slowly at first and then all at once. The lack of investment in capacity expansion is going to come up to bite us. And I think that's where the analog starts. When we think about the way this plays out and other analogs, that's probably the best one. And when we think about the cost of the hyperscalers, right now energy's budgeted to be about 10% of their cost. Depreciation's the highest. Memory and other things factor into that as well, but total cost of energy is supposed to be about 10%. If you plug in all of this compute and it's gas-powered, and we think gas could double or triple structurally, even without weather, it could end up being 20 or 30% of the cost of compute by 2029. We do think it becomes a much more material issue. Now, the levelized cost of energy, LCOE as it's referred to, takes into account CapEx, it takes into account cost of fuel. Everybody who is making decisions in this moment is using the forward curve for natural gas, which is flat. A little backwardated, a little contango, mostly flat out to the 2030s in the mid-3s. That is a very attractive low-cost fuel for the hyperscalers to commit to when they're focused on solving everything else. Like, how do I get compute in place to manifest this revenue growth at Anthropic or elsewhere? For us, we're just focused on modeling objectively — when you plug in this compute and this power gen source here and there, how exactly does it pull on the system of companies that we focus on?
· Patrick O'Shaughnessy
[00:00:00] S0: If you were forced to play devil's advocate in all of this and come up with the set of circumstances such that this is all much ado about nothing and we're sitting here in 2030 and gas costs $3, what do you think is the most likely reason? Is it data center power requirements are much lower because we make performance breakthroughs, or AI demand isn't what we think it's going to be like? What is this most sensitive to such that it might be wrong?
· Matthew Smith
[00:00:00] S1: So after we did most of our work, we went on a bit of a listening tour to target conversations with who we think are maybe the subject matter experts in that thing. So energy storage or hyperscaler compute deployment and energy consumption. And the common pushbacks, which we've spent a lot of time understanding, are Permian oil play — oil's high, Permian has lots of gas in the ground associated with it. Why can't Permian productivity fix problems? So our base case model already accounts for the 7+ billion cubic feet a day of pipelines that are already being built or developed that come on between '26 and 2030. If there were a new gas pipeline that would come on between now and 2030, we would know about it, because of the regulatory processes and the time it takes to build these pipes. So we've mitigated the risk of being surprised by the Permian by moving into the midstream to understand the bottlenecks and constraints. So beyond the deliverability of the resource or the gas in the ground itself, how much can actually get to market and either leave via LNG export terminals or be consumed in Texas or nearby? We've already included that in our base case model. And so that will be one of the pushbacks — there's plenty of gas in the Permian. But I would posit this may be controversial: there was plenty of oil in the world before the Iran conflict surfaced. Eventually there will be oil aplenty again. That's why it was at $55 a barrel before the Iran conflict. In order to make more Permian natural gas, you also have to be incentivized to make more Permian oil, and those incentives didn't exist until Iran. In order to produce a lot more Permian gas than even these 7+ Bcf pipelines being built that we're already modeling, you'd need much, much higher oil for longer, which only exacerbates this consumer crisis that we are concerned about. So we don't think the Permian solves the problem. And then the other one is, you can locate a bunch of behind-the-meter local Permian power generation, which is happening. But we are modeling what has been announced and proposed, and if it's going to consume local Permian gas, that means it's not going to make it into the pipeline downstream. We can accommodate that with our model.
· Matthew Smith, Patrick O'Shaughnessy
In any case, we always try to think about technologies that can disrupt, change structurally the need and consumption of natural gas, and so it often leads us to focus on battery technologies. There's sodium and other battery technologies that are currently not commercial, but people are getting a bit more enthusiastic about. The vast majority of economic battery deployment today is lithium-ion. It has a fairly fast discharge cycle. Those are being deployed in earnest across the system, and yes, we are also modeling known battery deployments as a part of modeling this generating system across all fuel types. A step function battery technology change could be something that would affect you, but that would affect some of these pieces that I described — winners and losers — in meaningful ways. It would be a watershed moment that I would welcome, because it would solve a problem that we're pretty concerned about. [00:00:00] S0: If you were czar for a day and you had to decide everything that gets started and to solve this problem, what are all the things that you would do to most solve and mitigate this? [00:00:00] S1: If I were the US government, I would find a way to build entirely — from beginning to end — two to four AP1000 nuclear reactors. That would de-risk the supply chain. It would invite in and really open up doors to folks who want to see somebody do it first before they do it. [00:00:00] S0: With the hopes that we get more like China — build 30 of them.