为什么天然气将成为 AI 的下一场重大短缺

Natural gas is well supplied today, in 2026 and 2027. When you get to the middle of 2028, we start to break very materially below where gas available in storage has ever been before. By 2029, we drop below all known historical storage evidence. You're going to start to see a knife fight to secure physical natural gas in 2028 like we really haven't seen before. And the biggest losers of this would be the US consumer.

So, Matt, the last time we did this was, I think, during COVID. It's kind of crazy that it's been 6 years. I've always loved talking to you about energy markets. You've been working in this space for 20 years. You're about as encyclopedic on this stuff as anyone I've ever met.

But you've also been acutely studying the current energy situation in the US, rebuilding, in a way that you'll describe, from the well level up, a picture of what's happening, especially as AI is creating all this new demand through data centers and so on, of what has been going on over the last 18 months of concerted effort. You've reached a fascinating and somewhat scary conclusion.

1. What Drives the Deficit

I'd love you to just start with a conclusion, and then we're going to talk through how you came to this conclusion, who the winners might be, the losers might be, and what's to be done about it. But before we get deep into all the component parts, just tell us what you found after 18 months of study.

We are headed into a place where we see an historic deficit in natural gas supply available in the United States, which does portend some pretty serious consequences. Natural gas, which is over 40% of US power generation, has become—or is imminently going to become—the most important fuel in the country. It's overtaking petroleum, given the amount that we now use for generation.

Our work suggests that 2026 and 2027 natural gas is appropriately supplied. But as we get into 2028, and you plug in this compute, assign gas to very specific assets as they're plugged in, and continue to export LNG as we're planning to do with known projects, we start to eat into our working gas storage, which is the nexus of supply and demand in the country.

I think we will come to the conclusion that the upside risk in the price of natural gas is both unbounded and convex. Where you will feel it most acutely will be electricity prices in 2028, 2029, and 2030, based on our work.

How much of this is just attributable to data centers? Just purely, we're building a lot more data centers, and that's just for AI. Is it that simple, or is there something else going on as well?

The die was cast long before AI compute came to the scene. If I may set the stage a little bit, US gas was plentiful starting in about 2010, when shale started to really come to the scene and change things. We had been importing natural gas to satisfy consumption on top of what we produced domestically. Shale started to be very productive, surprised to the upside, and became this abundant resource.

As natural gas became more abundant, we started to export it. Starting with Cheniere, we've gone from that early Cheniere export to today, when we're exporting about 15 Bcf a day of nameplate US export capacity. Now, that 15 billion cubic feet is on a base of about 110 to 112 Bcf a day of natural gas production in the US. So, if you think about it, it's become about 12% to 15% of the US's daily ability to supply the market.

As this abundance continued, more and more facilities and projects have been announced. So, as of today, we're scheduled to export up to 35 Bcf a day by the end of 2030, and in that case, the die has been mostly cast. To build an LNG project, you need various approvals. They're project-financed; you site and permit them many, many years in advance. Most of these projects that get you from 15 or 16 Bcf a day of nameplate capacity to 35 are well on their way.

That's the primary incremental demand driver in the country over the last 10 years and will be, at least, for the next 5. We had moderate population growth during the 2010s and into the 2020s. We went through a period of stagnating electricity demand, with energy efficiency and some other things driving down electricity demand, while you had more demand for gas-driven generation. But it's really been in the recent past where compute has started to pull incrementally. Before that, you had LNG as the main demand driver.

Now let's put those together. I just shared that we're going to go from about 15 to 35 Bcf a day of incremental LNG exports. After evaluating every producing gas well and the entire pipeline, processing, and gathering system, we have the capacity to add about 20 Bcf a day of gas production. Even without AI compute, we had sources and uses matched between our ability to deliver new natural gas from Appalachia, Haynesville, and the Permian and that which is supposed to leave the door through LNG.

Now, in AI compute, there are so many different power-generating ideas in order to power compute, given this value of time to power that folks talk about so much. It sort of goes from the large-scale, most-efficient assets, which are GE Vernova combined-cycle, all the way down through the distributed-generation assets, which we'll call fuel cells. We'll add Bloom, Caterpillar, and Solar Turbines. There are various local, field-level, behind-the-meter assets that are also relevant.

We have had to assign, with an outside partner, probabilities to all of this stuff.

All that stuff.

What we've gone about doing is starting with our base case, which we'll call P50. Everything with a probability of 50% or more—50% being that they have some approvals, usually a PPA, someone planning to buy power from them under contract, and usually some sort of interconnection agreement or an interconnection agreement in process—those are the assets we've taken seriously in our base case.

We'll call that the P50 level, and when you do that, it's about 5 Bcf a day of very credible incremental natural gas demand associated mostly with AI compute. Importantly, there are multiples of what we're considering seriously in our base case that have been proposed that will consume natural gas.

Bloom Energy's 6 Series, its latest-generation fuel cell, will take 150 million cubic feet a day of gas per gigawatt. The market has been assigning a high probability to their obtaining 2 gigawatts a year of manufacturing capacity, and that's likely to ramp to 5 gigawatts, which there isn't gas for unless you take it from something else.

In the extreme case, how high does that number get?

If you start to move it down and say P30 or P0, that number can more than double and be 12 to 15 Bcf a day by the early 2030s if unmitigated.

Twelve to 15 Bcf a day in the extreme case? Just shut off the exports. Who cares? We didn't export natural gas for a long time. People domestically are not going to tolerate skyrocketing energy prices, especially when they think the simple solution to this is just to stop shipping it out of the country and use it for ourselves. Why is the solution not just to shut off exports?

It's more complicated than that because of contract law. There are rules. There are really good reasons why we're exporting, and these projects have tens of billions of dollars of project financing and contracts attached to or associated with these LNG projects. As the US will be about a third of global gas supply in several years, our allies and other FTA—and increasingly non-FTA—countries are reliant on that supply.

What's FTA?

Free-trade agreement. So, the answer is that it's both because it's a third of global supply, which is really important for the rest of the world, and domestically there are contracts and investments. It could be stopped, but it would be very complicated.

Yeah. Let's step back for a moment. You have to have a starting place for a base case, which typically starts with signed contracts. What do the words on the page say? What's allowed? What's not allowed?

When we set out to build the firm, we've had about 16-plus months to start to model almost every asset at a time, at the atomic level. Along the way, there are numerous constraints, rules and regulations, and contracts. When we set out to build this, it was about acknowledging those constraints for what they are, assuming that contract law would be followed.

As we go through and build all of this, it can flex up and down based on the choice to send less LNG out of the country or out of our terminals, for instance, or to slow AI compute growth, which is one other solution. We're not really willing to propose that because we know that there's insatiable demand.

It's not popular to say, “Slow AI compute growth,” but to the extent that would happen, that would be another lever to reduce the pull or strain we expect in the system as the decade goes along.

2. Why Supply Can’t Catch Up

I'm just going to try to ask really simple questions here because, just so as not to minimize it, it's your view that in the bad-to-worst-case scenarios, this is a full-blown crisis. This is not a small thing. This is the story in the country.

I want to make sure the whole reason we're going into all this detail is that, in this scenario, it's really, really bad, and it's really bad primarily, I guess, through prices. Maybe you can continue to articulate why we don't necessarily want this specific outcome and what we can do about it.

But help me understand: underneath the United States right now, or North America, there's a certain amount of gas, just objectively, and I'm trying to understand how much of this is that we are literally going to run out of the gas that's under the ground versus it's just a problem of how quickly we can find out where it is, get it out economically, process it, store it, transmit it, use it, et cetera. Those seem like 2 separate problems: literally how much there is, and then what we can do with it and about it. Is any element of this problem that there's just literally not enough of it?

Starting with what we'll call resource in the ground.

Yeah.

There's tremendous data availability. We can measure where we are in the exploitation of most of the major gas-producing basins. They would be Appalachia, which is primarily the Marcellus plus the Utica; Haynesville, which is a key swing basin; then, of course, the Permian and, to a lesser extent, the Eagle Ford. These are oil-directed plays where the decision to drill and produce is driven by oil, with gas as a byproduct.

In each of these plays, there are some stacked-pay zones where well-penetration output can be measured with a lot of data. What that allows us to do, when you digitize the acreage controlled by each one of these companies with polygon shapes that use a bunch of lat/longs to drop in and associate a well with an area that's controlled, is figure out what's left.

The reality is there is gas, and we've, as a part of our analysis, produced the gas that is logically captured and can be produced from wells from existing acreage positions of all these companies. There is gas. We're assuming it gets developed here. That's how you get to our 20 Bcf a day of growth.

It's unbelievably cool that we can literally know at this precision what is underneath the ground, often deep underneath the ground in hard-to-reach places. It's a technology story, right? That would be fun to tell sometime, but it doesn't sound like the actual problem is that we are literally running out of the stuff underneath the ground.

We've also had a history of just finding new stuff that we didn't know existed before. So it sounds like the problem is more our ability to serve the demand in this kind of time frame, not that we're literally going to run out of the resource over the next 20 years.

So it's a little more complicated than that. We get through most of the existing captured inventory of companies in the next 4 or 5 years. If you think about bringing on a new well, it has a decline rate, and each well, as it's stacked on an existing company-wide portfolio, declines. A lot of these companies' decline curves are maturing some, and so they—

They're pretty steep in natural gas, right?

It's steep initially in natural gas, but Expand Energy and others have such mature portfolios; the replacement is less costly today than it would have been 5 years ago. When you stack all of these wells based on existing acreage onto these companies, assuming they're going to drill optimally based on their forward curve, which is depressed—and we'll talk about that—you get to this 128 to 132 Bcf a day of maximum deliverability.

So we are assuming that all of these companies develop the rest of their acreage.

But that's a flow metric, not a stock metric.

It's a flow metric. What is possible when you use known well-performance parameters to maximize production before you get to midstream and other surface-level constraints, which we'll talk about. So there's resource.

Yeah.

We are depleting the known resource. If you were to assume prices go up meaningfully, you may unlock additional basins that are legacy known basins. We know a lot about most of the rock in the U.S. There are other known gas basins, but they have been uneconomic. Furthermore, there isn't infrastructure to really accelerate drilling and activity in those basins.

This isn't software. You don't press a button to solve this. The constraints are multifold. So the first constraint is the rock. We have the ability, we think, to get to 128 to 132 BCF. I started with the highest estimate, which is 132 BCF, as a starting place, because that's how you solve the LNG exports. We'll take the under on that, but that's where you can get to.

A common pushback, as we've gone through this, is that there's plenty of resource available to us in the Permian. There's plenty of resource in Appalachia. A number of companies describe themselves as having a lot more inventory of wells to drill than we can justify with facts, and I'll just leave it at that.

But when folks meet with companies, they should ask to understand exact engineered locations on a map. Where do they have not just the ability to produce, but plans to have infrastructure on the surface to allow it to flow, for instance, and the ability within financial parameters to invest in and produce the resource in the ground?

We're fairly far along in understanding it. We've accounted for all of the major productive basins in the country, and I do not think we're likely to be surprised by some new major shale find. At this point, knowledge of those things is pretty mature. So, to say it back, there's a lot of resource, but at this rate, we're depleting the known resources quite quickly.

We are advanced in depleting the known resources, especially as we move to the next layer, which is infrastructure.

Okay. There's a bigger, longer-term question about whether this is just a parenthetical period of time between 2010 and now, when we were awash in gas. Prior to that, we really weren't, and maybe after this, we really won't be.

Let's talk now about what happens after we get gas out of the ground. It still has to get processed, transmitted, and used. What are the most important rate limiters in that part of the equation?

In some cases, it's processing the natural gas flows to the surface with natural gas liquids embedded therein. In some cases, there's sulfur or nitrogen that has to be dealt with. In some cases, it comes with oil, and so you have to have surface-level infrastructure to produce the oil, which is different.

Gas primarily has to be produced into a pipeline system, and there's a certain spec on regulated pipelines; it's kind of 1,030 BTU. That's the spec. So you have to remove enough of these other hydrocarbons to get it to pipeline spec, to be able to produce it into the system and have it consumed by folks downstream.

Processing is the first major constraint. There are a couple of basins with a little bit of extra processing; we'll fill it up pretty quickly. We do not yet have the processing necessary to get to our assumed 20 BCF production target.

Processing would be something that takes 2 or 3 years to build, at the midpoint. We generally know what processing investments are being made, and projects have been announced by Permian processors or by Appalachian processors. We know where the materials and liquids-handling throughput capacity will be in 2027–28, at least.

You really would need, in the near term—even to get to our 20 BCF of incremental gas production we're willing to estimate—you need to have more processing built.

Build stuff now.

One, build it now—imminently.

Gathering is small-diameter pipes. Gathering is what takes it from the wellhead to processing or the pipeline system. There is a fair amount of disclosure around processing systems being expanded and built, and we would posit—we have put all those on a map on top of every one of these wells at their lat/longs—in order to grow even a fraction of where we must have natural gas production go in the U.S.

Gathering has to be invested in very materially, imminently, to get to the place where we can achieve 130 BCF a day of production in the U.S.

The last one is the interstate gas pipeline system. This is where I come back to your answer on LNG—lots of rules and regulations around these things. Pipelines are monopolies for the most part. Local distribution companies that deliver gas to your stove, those are monopolies or oligopolies.

In the last 10 or 12 years, we've really built 1 interstate gas pipeline, and that was Mountain Valley Pipeline, connecting Appalachia to the Mid-Atlantic. The various environmental permitting regime changes—

Things are a nightmare to build.

Yeah. It’s been made very difficult to build interstate gas pipelines. This administration has been trying to reduce the barriers to building interstate gas pipes. We’ve started to see some more progress to that end, but there is an urgency to build more connectivity to wield gas around the country to serve this incremental AI compute load, if it’s going to happen. That’s necessary.

3. The 2030 Gas Crisis

Before we keep going through this sequence here, can you just say what you think the state will be? Let’s assume that there’s roughly inertia in the system and nobody listens to this. Hopefully, some people—a lot of people—will listen to this and have ideas. But what is going to happen in the world? What will the state of the world be like in 2030 if none of this starts getting addressed sooner than later? What’s your best guess as to what it looks like?

There is a tremendous inertia around natural gas being the primary fuel to power AI. The market has been focused on understanding the power shortage and trying to solve that, and power generation generally, which could be solar and batteries, wind, nuclear—whether large-scale or small modular reactors, or SMRs—or natural gas. Natural gas is well supplied today in 2026 and 2027, and the result is that nobody is investing in gas. In fact, EKT (likely EQT) is shutting in natural gas right now because they think it’ll be more valuable later.

The rig count and the things that we can see in real time to figure out if the market is onto this tightness in 2029 or 2030 show that no one’s onto it.

No one’s on it.

It’s not apparent today. It’s perpetuating this view that most Americans have, which is that there’s plenty of natural gas, because for 15 years all we’ve heard is that it’s abundant.

Abundant. Yeah.

And so there’s a complacency that’s developed, and we think that complacency is going to take us right up to the point where it’s too late. We do think the die has been cast. Gas, which is currently at $3.50 or $3.60, going out to 2026, 2027, and 2028, has a flat curve. In 2029 and 2030, the curve is flat because people believe the gas is abundant.

That’s despite all these AI compute announcements and despite what all of the companies are doing for their investments. Gas has lulled everybody to sleep. But what happens is these structural things start to fall in place in 2027 and 2028, and we start to draw meaningfully in the middle of 2028—as early as early 2028—on the gas system like we’ve never drawn before.

As we look at 2028, 2029, and 2030, we start to cut into US working gas storage, which is about 4 Tcf of total gas storage. There’s a range of high and low for that storage seasonally, as we draw in summer and winter and then build in the shorter months in the spring and fall. When you get to the middle of 2028, we start to break very materially below the level where gas available in storage has ever been before, historically. By 2029, we drop below all known historical storage evidence, and by 2030, we get pretty close to where we think, ceteris paribus, gas storage looks very, very low.

At that point in time, because it hasn’t happened before, we’re forced to look at where gas prices have gone during shortages. During the Russia-Ukraine conflict, gas went to $8 or $10 an MCF because we sent a lot more externally to Europe. We’ve seen various weather anomalies—the polar vortex in 2014 and December of 2022—and those prices have gone to $6, $8, or $10. But those have been transitory.

What we’re talking about are structural drivers of demand against a known possible production of gas, and they don’t match up. You draw down in a very historic way starting in 2028, to the point where the deficit gets really convex and unbounded.

Meaning, gas prices could be $20 or something like this.

I would hesitate to even put a price target on it, but at $8 or $10, we think you potentially shut off some of the US exports where there are spot cargoes and they are leaving the border to capture uplift in Europe or elsewhere. Those spot cargoes may not be lifted, and then that gas is left in the system. We’ve tried to account for that in our model.

4. Winners and Losers

But spot cargoes alone can’t solve this. You would have to get into shutting off contracted cargoes leaving our border via LNG to really start to mitigate some of this. It’s hard for us to count on the choice to shut off contracted cargoes when there’s some Japanese utility counterparty that has counted on it for its provision of electricity.

In crazy convex outcomes like this, can you tick through who you think the biggest winners and losers are?

There are some clear natural gas producer winners. Expand Energy is probably at the top of that list. They probably control 70% of the remaining core Haynesville wells, with very closely known parameters of rock where we know it to be very productive. We think Expand is far and away the biggest winner.

Uniquely, Expand is CEO-less right now. There was some turnover in the year, and they’re going through a search. The stock has plummeted over the last 6 months as part of that search, and it’s trading at 4 times EBITDA on a forward curve where no one believes what I’m telling you to be the case. Even though we think modeling the facts gets you to a much higher gas price, the stock has dropped. The assets have not changed. It has some of the highest-quality rock in the country.

The highest-quality upstream company in Appalachia is probably Range. Range has significant room to grow production and materially grow returns to investors. Those would be the upstream companies.

It’s not going to be obvious in the first pass through this equation for most, but natural gas sets the marginal fuel for the next-in-line power-generating asset in each power market. As natural gas goes, power prices go in the country.

If you think about the dispatch curve of different generating assets in the country, there are some where the fuel is free. That would be solar, and to a lesser extent wind and hydro. Solar assets, which are growing meaningfully, have been 90% of the interconnection queue with batteries in the last 10 years in terms of new assets coming on, other than gas.

Solar assets stand to benefit from a windfall where electricity prices are going up because gas is setting the margin of a plant, and the price of its fuel is increasing, while the sun costs the same. We think there are some companies positioned very well for margin expansion with no incremental capital cost.

XPLR, ticker XIFR, formerly NextEra Yield Co, has an interesting set of assets. They have a windfall coming in the latter part of the decade and in the early 2030s, ceteris paribus, because they mark their PPAs to market at much higher values without any capital expenditures. Split Away Energy would be another one in a similar circumstance. Solar assets at the utility scale, especially, stand to win.

Maybe more interestingly, as it relates to some of our discussions of the past, residential solar—which has been suffering from the first removal of tax incentives to install residential solar since the late ’70s—is really one of the only ways to protect yourself from what’s going to happen during the time of 10:00 a.m. to 6:00 p.m. once gas gets really tight in the electricity markets.

What you pay for electricity at your house—we think residential solar grows exponentially from here, even without tax incentives. For the first time, it’s very economic with where electricity prices are likely to go to install solar, especially when it combines batteries, which make the electricity much more available around the clock.

I think it’s important to say you’re an investor. You have money behind this work beyond those two categories. Are there any other surprising winners, do you think, in all of this? What about nuclear? What about Westinghouse or places like this?

5. Nuclear and Solar

What we’re talking about is a complex dynamic system where there will be choices to consume electricity or not at different times. As I go through this, I want to make sure I acknowledge that there’s no silver-bullet solution for what I described as a convex situation with natural gas and, therefore, electricity prices as the decade closes.

There’s no bridge fuel other than solar and wind, because currently natural gas is the only flex fuel to get us to when we can bring on nuclear. We’ve spent a fair amount of time as well in the nuclear ecosystem, and to us, large-scale nukes are the only solution that makes sense. That points us primarily to the AP1000 Westinghouse units.

Don’t those take, like, 5 years to build or something?

More than that, but at least they have a storied past and a very brief history. We’ve built 2 nuclear reactor units in 30 years in the US: Vogtle 3 and 4. Around that time, we also tried to build one in South Carolina called V.C. Summer, another nuclear project at the time that nearly bankrupted SCANA, which was later pushed into the arms of Dominion, and the project was shut down.

The muscle memory from trying to build large-scale nukes—especially off the back of Fukushima in 2011, Chernobyl, and Three Mile Island—is that, for 3 decades, nuclear engineers, scientists, and companies have gone to do other stuff.

The Vogtle 3 and 4 experiment, where it cost 3 times as much and took, I think, 15 years from birth to commercial service, is the recent memory of building these nuclear units. But if you go out to the 2030s, what I’m describing in terms of the gas deficit only gets worse in 2031, 2032, and beyond.

And so, in our mind, the only viable solution is to build large-scale nuclear as fast as possible, which would mean it needs to come online in 2033 or 2034, as soon as it can come online. Regular utilities, hyperscalers, and regulators should all align around that goal.

But because people don’t really believe that gas is in short supply as the decade goes along, they don’t believe in the problem, and they don’t like the solution. So they need to be convinced of the problem. The country has a history of building pipelines to solve problems that exist today, not problems that will exist in 5 or 10 years.

We’re trying to get out ahead and see where the puck’s going. Where it’s going is that we are going to need large-scale nuclear by 2033 or 2034.

You don’t think SMRs can be a solution, where you use smaller reactors to power individual data centers behind the meter and this just never touches the system?

Many of the SMRs are still science experiments. The NRC and the U.S. government are actually doing a fair number of things to break down the barriers to bringing those to fruition, to see if they work, what the cost will be, and whether they can be scaled.

But many of these SMR companies are not set up to manufacture and truly scale for the solution that’s needed to solve this problem, which is tens of gigawatts as you go into the 2030s. And so that points us to these AP1000s.

Since Vogtle 4 came online, and Vogtle 4 experienced very material improvements from Vogtle 3, China today is building 39, plus or minus, nuclear reactors. Thirty-four of them are 1 gigawatt or more, and I think a third of those are modeled after the AP1000.

We know a lot more today about building large-scale nuclear plants than we did when these mistakes were made. Large-scale, to us, can be commercialized on a known timeline, where the costs are probably better than with something where we don’t even know if we can scale the businesses yet, in terms of SMRs. Large-scale versus small probably wins in our mind.

The 2 companies most levered to that would be Kamico, which owns 49%, and Brookfield, which owns 51%. You’ll probably find that the U.S. government agrees with what I’m describing. They seem to really be lining up and trying to facilitate commitments and early procurement, which will derisk some of the supply chain and help put timelines on this.

6. Consumers Pay the Bill

When Westinghouse goes public, it’s deeply undervalued within Kamico today, so that’s an interesting one. BWX Technologies is a super-interesting company. They’re the primary supplier of nuclear power for the U.S. Navy, they significantly benefit from the coming nuclear cycle as well, and there’s lots of dollar content in the AP1000s.

Who are the big losers, do you think, in this future?

Sadly, the biggest loser would be the U.S. consumer. To the point where you take what I’m saying, and if we’re even partially right, electricity prices rise, which you can see some of on the forward curves in these different markets.

As electricity prices rise, you start to think about the trade-off: Are we going to export natural gas to foreign buyers? Are we going to use it for AI compute? Or are we going to try to keep consumer electricity bills down? It’s an awful trade-off, and I think it will probably start to contribute more to the public dialogue, the NIMBYism that we’re already seeing pop up in some places.

We think AI is tremendously transformational. We’re not anti-AI, but it consumes a lot of power. We need to really focus on the 2030 to 2035 period. The U.S. consumer is probably going to pay the bill in the meantime.

Most of the solutions being proposed by the government are to consume more gas because everybody believes there’s enough. Bring-your-own-generation, or BYOG, is a thing today. That’s what the hyperscalers are being asked to do: site their data center at a certain latitude and longitude.

Well, that means more gas, not less. Every time you read a press release from Bloom or likely GE Vernova, think more gas. You can use the energy efficiency of each one of those units and understand exactly how much more incremental gas, beyond the base case that I just shared, is dangerously tight.

Another loser—and I want to be respectful here—is some of the biggest winners so far, at least in the stock market: the manufacturers of gas turbines or distributed power generation sets. When you think about those companies, it’s been somewhat boom-and-bust.

In the early 2000s, there was a boom to build as many gas plants as we could. The capacity was overbuilt, and the industry really languished for a long time until now. You’ve had tremendous profitability and equity returns come from these companies over the last 2 years.

But as you look at 2028 and 2029, most of them are adding more capacity again, just like they did in the early 2000s.

Which companies are these?

Caterpillar. Caterpillar is, I think, doubling its Solar Turbines capacity between now and the end of 2029, which I would judge to be at exactly the 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 has been topical recently because of other things that folks are talking about, such as the rare earths they use in their manufacturing, for instance. We don’t think Bloom Energy’s assets, at 2 gigawatts or more, will be able to get natural gas in competition with all of the other assets being deployed that will consume gas, given the scarcity that we see.

Those are 2 manufacturers of distributed generation, or behind-the-meter generation, that we think are probably more poorly positioned than investors appreciate. It really doesn’t make sense to us, beyond 2029 or 2030, to build large-scale natural-gas generation until we ramp up production meaningfully and can make sure we have security of deliverability and a supply of gas that’s consistent with our model.

7. AI’s Next Shortage

We could see orders slow very meaningfully for natural-gas-generating assets, even at large scale, as 2026 progresses. Those could be some of the losers. It just may not make sense to use gas for power generation for new or incremental assets after a certain point.

It seems like this whole memory-shortage thing that we’re going through right now—that hyperscalers might also be people who are in trouble here, if this is a key input to what they’re doing. Do you think that’s a big problem for them?

As we’ve been socializing this a little bit, trying to learn more and have people poke holes, we shared this with one of your recent guests. He listened and said, “Well, this sounds like DRAM 2 years ago: slowly at first, and then all at once.” The lack of investment in capacity expansion is going to come up and bite us, and I think that’s where the analogy starts.

When we think about the way this plays out and other analogies, that’s probably the best one. When we think about the costs of the hyperscalers right now, energy is budgeted to be about 10% of their costs. Depreciation is the highest. Certainly, memory and other things factor into that as well, but total energy cost 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. So we do think it becomes a much more material issue.

The levelized cost of energy, or LCOE, takes into account capital expenditures and the cost of fuel. Everybody making decisions in this moment is using the forward curve for natural gas, which is flat—mostly flat, with a little backwardation—out to the 2030s and the mid-2030s.

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 in Anthropic or elsewhere?” For us, we’re just focused on modeling objectively: When you plug in this compute and this power-generation source here and there, how exactly does it pull on the system of companies that we focus on?

If you’re 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 that data center power requirements are much lower because we make performance breakthroughs, or that AI demand isn’t what we think it’s going to be? What is this most sensitive to, such that you might be wrong?

After we did most of our work, we went on a bit of a listening tour to target conversations with who we think are the subject-matter experts in those areas, such as energy storage or hyperscaler compute deployment and energy consumption.

The common pushbacks, which we’ve spent a lot of time understanding, are around the Permian oil play. Oil is high, the Permian has lots of gas in the ground associated with it, and people ask, “Why can’t Permian productivity just fix the problems?”

Our base-case model already accounts for the 7-plus billion cubic feet per day of pipelines that are already being built or developed and that come online between 2026 and 2030 in the Permian. If there were a new gas pipeline that would come online between now and 2030, we would know about it because of the regulatory processes and the time it takes to build these pipes.

We’ve mitigated the risk of being surprised by the Permian by moving into the midstream to understand the bottlenecks and constraints. Beyond the deliverability of the resource—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. That will be one of the pushbacks: There's plenty of gas in the Permian, but I would posit—and this may be controversial—that 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. Those incentives didn't exist until Iran. In order to produce a lot more Permian gas than even these 7-plus Bcf/d 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're concerned about.

We don't think the Permian solves the problem. The other possibility is that you can locate a bunch of behind-the-meter, local Permian power generation, which is happening. But we're modeling what's 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.

The only other one—and this is hard to handicap for—is technologies that can disrupt and structurally change the need for and consumption of natural gas. It often leads us to focus on battery technologies. There are sodium-ion and other battery technologies that are currently not commercial, but folks are working on them and people are getting a bit more enthusiastic about them.

The vast majority of economic battery deployment today is lithium-ion. It has a fairly fast discharge cycle, and those are being deployed in earnest across the system. 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 it 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.

8. Solutions and Global Stakes

If you were czar for a day and you had to decide everything that gets started to solve this problem, what are all the things that you would do to most solve and mitigate this?

If I were the U.S. government, I would find a way to build, entirely from beginning to end, 2 to 4 AP1000 nuclear plants—nuclear reactors. That would derisk 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.

With the hopes that we get more China-like and build 30 of them.

Currently, there are 10 to 20 envisioned by the U.S. government through different groups in terms of nuclear reactors coming online, but no one wants to be first. I think we're close to a few stepping forward. But if the U.S. has said, “Hey, we have $260 billion to spend at the Loan Programs Office”—now, the EDF, it’s $260 billion, I think, to spend by the end of 2018—“we need to build 4 nukes with this,” that would derisk this materially, and I think you would see it jump-start the nuclear equation.

Energy and resource availability, egress, and knowledge, we think, are the biggest bottlenecks to productivity and deployment of all this incredible technology that America has really been the leader of developing. That's why, when we set up our firm and started building a team, things were happening pretty quickly in AI, and especially with regard to power. We didn't set out to understand the now. It's important to understand now, but we set out to understand where the puck is going, and where it's going looks like it will meaningfully diminish growth if not dealt with. That's why it brings me to the nuclear solution as maybe the most viable long-term.

My comments on solar are probably the most important thing that I would do. I think everybody should get a solar system on their houses. It won't be perfect. It will deliver electricity when there's sun out. It may not deliver electricity when it's cloudy, but it's a way to protect yourself from very high peak power prices from 10 a.m. until 6 p.m., which are the biggest part of your bill.

I would incentivize people to study their state's rules and try to put incentives in place to really, really start to grow residential solar faster than what's been a stagnating industry over the last year after some incentives were removed. That would be one place to reinvigorate solar incentives because they're going to be needed in a few years.

Could we build a giant pipeline from Canada or something? Try to tap our neighbors to help us solve this problem?

Canada is an interesting partner of ours. They have the capability of delivering about 11 or 12 Bcf a day, usually in January, periodically, and there are meaningful pipelines from Canada to the U.S. Largely, it's seasonal, and it helps us solve winter, but otherwise they're net importers most of the year. We're not really set up to take from Canada year-round.

There are a few reasons for that. One is that Canada has limited storage. It's about 0.7 Tcf, about a fourth of U.S. storage. If I think about this on a 3- to 5-year-plus basis, Canada by far has the deepest and richest resource of economic gas in the ground, but it's been trapped behind pipe.

I would build a 1- to 2-Bcf-a-day, at least, pipe into the U.S. Midwest, the MISO power market, and then wheel it around—MISO, PJM, SPP—and try to satisfy this demand because I really don't want to see demand slow. I really don't want to see consumers' bills go up.

We've really focused on the U.S. here. What are the implications of this for the rest of the world?

Well, the U.S. has become the leading provider of natural gas, with our exports going from virtually zero, starting in the mid-2010s, to 15 Bcf a day now. That'll be 35. Many countries in the world are building gas-generating assets that are dependent on our delivery of that gas to them.

There are lots of political-conflict-related tensions or bottlenecks today around Russia and Ukraine. Russia used to be one of the biggest deliverers of gas to Europe, for instance, and we filled the gap. To the extent we cannot deliver our 30 to 35 or more Bcf a day of gas to the global consumer, rebalancing will be required.

You probably impact Europe meaningfully, and they're left with a trade-off of taking Russian gas or much, much higher-cost U.S. gas because they can't produce enough domestically to satisfy their need. Asia is a large consumer, the largest consumer of U.S. natural gas until Russia-Ukraine, and they probably will be, at some point, the largest consumer again with some of the outages in the Middle East.

We potentially hurt important allies at a time when we really want them to be allies if we can't send them the gas they need. It is pretty important that we don't curtail LNG, although that will certainly be one of the levers, as we go out to the late decade, that we will be forced to think about to deal with rising electricity prices in the country.

9. The Coming Gas Knife Fight

Is there anything that we haven't talked about that has surprised you in this year-and-a-half-long analysis of trying to understand the state of things and where we're going? Obviously, we've covered the big conclusion, which is scary and hard to deal with, even if we start acting now. Anything else that surprised you, either in your work or in people's reaction to it, as you've started to share it?

Really, for the last 2 or 3 years, the phenomenon has been that CEOs and CFOs of companies—all of whom have really been asked by their investors, or whose products lend themselves well to deploying products to capture AI compute rent—

What CEOs have said versus what is possible from the system, I think, is an interesting study that will happen over time. The amount of capital made available to companies to make investments that are really shortsighted in the context of our work is significant. The incremental distributed natural gas genset is an inefficient, high-heat-rate, high-cost asset that really should only serve as backup generation in any context outside of this fast-time-to-power setting where AI compute needs the power.

Now, over 4 or 5 years, those assets may not even run. And so you've had tens of billions flow into these distributed power assets, all of which are short and will consume natural gas. We really haven't seen anybody, including firms we really respect, question at any point whether there will be enough gas and what the cost will be when the time comes. It's been surprising.

It’s been surprising that enough folks haven't put pen to paper to then start contracting gas to make sure they have supply certainty. We haven't seen more financial contracting. 2028 is somewhat illiquid, which is why we really haven't seen the forward curve move. And we think that's where the action starts.

As soon as utilities turn the page and start to really hedge or buy gas in 2028, and we start to see all these natural gas-generating companies start to think about securing supply, you're going to start to see a knife fight to secure physical natural gas in 2028 like we really haven't seen before. And it's been surprising that we haven't really seen any of this yet, because in 2028 and 2029, the physical market tightens materially, depending on where you are and the amount of money that's gone into unproven, untested projects—the amount of capital being raised for things that really may not happen until 2035.

Maybe that's been surprising, especially as it relates to the fact that you've got Expand Energy trading at 4 times EBITDA, a low-to-mid-teens free-cash-flow yield on a gas forward curve that is complacent about all of the objective things that we already know are likely to get plugged in. People are not really willing to look past summer heat or a slight outage in an LNG facility right now.

But in 6 months, we start to see these companies roll forward to look at 2028. In 6 months, you could start to see the forward curves really move up materially. Investors are not willing to look past a near-term, appropriately supplied gas market, but they're willing to pay for something in 2035 that is totally untested or unproven. And that's—it’s been surprising, the assumptions and the inconsistency across sectors and industries that we see, even among the 4 sectors we follow.

So maybe in closing, what would be the healthy challenge to pose to anyone out there whose business has, as an input, directly or indirectly, energy prices? How would you encourage CEOs—what question should they ask of themselves or their business?

Make sure when your assets are deployed that you understand exactly what the source of your natural gas will be. Make sure you have physical supply locked up and that you understand your counterparties and what will likely be very meaningful counterparty risk in 2 or 3 years.

Counterparty risk isn't something we've really talked about during the last couple of years in the AI boom. But when it comes to parties being long and short something that is moving a lot, imagine being short memory a year ago or 18 months ago and finding out all of a sudden you're short memory. That is what this natural gas market looks like to us—not 2 years out, but 6-plus months out.

Making sure you understand the physical provisioning of gas for your assets is important for the hyperscalers and for the buyers of simple-cycle and CCGT large-scale plants, but especially for fuel cells. We are very cynical about whether you can deploy fuel cells at scale, because there isn't the gas in the system to power those 24/7, 365 days a year. And so we treat them in our base case that I described as backup generation.

To the extent you were to deploy fuel cells as baseload generation, that only pulls forward and is additive to the convexity that I described. For CEOs and partners on projects, whether you are the E&C company trading at 25 times cash flow—which is a historically high multiple for an engineering and construction firm—and your main business is building natural gas plants, you may not be able to build or deploy more gas plants at a certain point in 2029 or 2030 because gas is much more expensive and you may have regulators asking questions.

The focus for you should be on how you do accretive M&A to backfill and diversify your business so that you're not entirely beholden to natural-gas-generating asset build. For hyperscalers, I know memory has been a pain point. Natural gas could be 20%, 30%, or 40% of their cost of doing business at a time when they're supposed to be reaching escape velocity with profitability.

[Speaker?]

Performance per watt is probably a compute metric that we're going to care more and more about. I would say that the questions or the challenges for each pocket or industry of companies are a little bit different, but it's all focused on making sure you're managing risk and understanding exactly, when your plans play out, how it can go wrong—which in this case means, what if gas is not $3.50 but $10 or more? What happens if physical gas is questioned? What happens when consumers and, therefore, regulators start to ask questions? Do you have a backup plan? What's your backup plan?

Matt, I love talking about the energy system with you. This was an especially fun one on the back of so much of your work. So fascinating and interesting. I hope, as the U.S. has been very good at doing historically, that lots of people listen and start to imagine solutions, and also create the right amount of urgency to get those solutions in place, and that we emerge from this more resilient, more capable, more efficient—all these things. Thanks so much for your time.

Thank you again for having me.