Elon Musk Acquires APR Energy for $1 Billion to Power AI Infrastructure

Organized & Compiled by Shenchao TechFlow

Guests: Josh Kale (AI Analyst, Freelancer at Anthropic), Ejaaz Ahamadeen (Co-host of the Limitless Podcast)

Moderators: Josh Kale & Ejaaz Ahamadeen – Dual Conversation

Podcast source: Limitless Podcast

Elon's $1 Billion Bet Reveals the Next AI Trade

Broadcast date: July 22, 2026

Disclosure: Josh Kale is a contractor for Anthropic; the views expressed in this episode are his own and do not represent Anthropic. Both individuals state that the content of this episode does not constitute investment advice.

Musk spent approximately $1 billion of his personal funds to quietly acquire a company called APR Energy. This company doesn’t manufacture chips, rockets, or AI models—it does one thing only: mounting gas turbines on trailers and delivering them to your data center campus to power your GPUs within days. The fact that someone who has championed solar energy transition for two decades has now purchased fossil fuel power equipment reveals the true bottleneck in the AI arms race.

This episode uses the acquisition as a starting point to break down two ongoing rotations: first, the pullback in memory stocks (DRAM, HBM, NAND) after a sharp rally, despite no change in underlying demand; second, capital shifting toward power infrastructure. The two hosts map out the full stack of power technologies—from mobile gas turbines (APR Energy) and solid oxide fuel cells (Bloom Energy) to gas turbines and grid equipment (GE Vernova), and small modular nuclear reactors (Valor)—arguing that power will be the next frontier for AI investing in the second half of 2026.

Regarding Musk's acquisition logic

One gigawatt of power is enough to supply 750,000 households and is equivalent to the output of a nuclear reactor. This amount of electricity can power approximately 600,000 H100-class GPUs.

We became aware of this transaction because one of APR Energy’s investors was required to disclose $50 million in personal returns. This was the only lead.

He chose to use his personal funds rather than purchase under xAI or Tesla, likely for tax and structural reasons.

Regarding power bottlenecks

In 2023, data center electricity demand was approximately 23 gigawatts, doubling to 46.5 gigawatts by 2026. However, only 12 gigawatts of new capacity were planned this year, with only 5 gigawatts actually completed.

Connecting electricity to the grid requires high-voltage transformers, grid infrastructure, and various permits and regulatory approvals. The entire process takes 5 to 7 years—I’m not exaggerating.

New York State has just banned a batch of data centers, which will delay local data center construction by five years.

Regarding the pullback of memory stocks

In July, the average price of DRAM rose nearly 20%, but memory stocks fell by 20%. Prices are still rising, yet the stocks are declining.

SK Hynix has secured full-year supply commitments for 2027. Thirteen to fifteen customers have locked in 40% of next year’s projected profits. Demand is overwhelming.

Micron dropped 24% in one month, but its forward P/E ratio is only 7x, with revenue growth of 350% and a gross margin of 85%. Such high profitability is rare in the hardware sector.

Korean investors were over-leveraged by approximately $1 billion, causing the market to lose $1.5 trillion. The fundamentals remain intact; this is merely an oversold condition.

About electricity as the next trading signal

Electrons are more valuable than dollars. Throughout all of human history, there has been only one trend: more energy equals more productivity, which equals more innovation, which equals more prosperity.

This reminds me of what memory trading looked like before it became "memory trading." History doesn't repeat itself, but it often rhymes.

GE Vernova is the TSMC of the power industry. Orders are booked through 2031.

Microsoft has signed a 20-year power purchase agreement with the Three Mile Island nuclear plant, taking 100% of its output. Locking in power for two decades is far more imaginative than issuing 20-year bonds with a 3% yield.

Model independence of energy demand

Whether your model is open-source or closed-source, expensive or inexpensive, cutting-edge or non-cutting-edge, you will need energy and GPUs.

China has just released Kimi K3, a 2.8-trillion-parameter open-source model. Already, 56% of tokens on OpenRouter are being consumed by Chinese open-source models. But Jensen Huang isn’t worried, because whoever builds models still needs to buy his GPUs. The same goes for electricity.

What did Musk buy for $1 billion?

Ejaaz: Musk just spent over $1 billion of his personal funds to buy a company almost no one has heard of: APR Energy. The strangest thing about this company is that it doesn’t make AI chips, rockets, or any AI models—it builds gas turbines mounted on trailers, which are transported to your data center campus and can power your GPUs within days. A man who spent 20 years advocating for solar energy made this acquisition. When you unpack why he bought it, you uncover a new insight into the AI sector—and also understand why memory stocks have recently been falling. The U.S. has ample energy, but connecting that energy to the billions of dollars’ worth of GPU clusters coming online this year is an extremely difficult challenge.

Josh: APR Energy has an interesting history. Founded in Jacksonville in 2004, it went public on the London Stock Exchange in 2011, acquired GE’s energy leasing business in 2013, and became the world’s largest mobile gas turbine leasing company. Since then, it has changed hands among private equity firms. Through reviewing regulatory filings, we discovered that Musk spent approximately $1 billion to acquire it. What did $1 billion buy? About one gigawatt of power. For context, one gigawatt can power 750,000 homes and is equivalent to the output of a nuclear reactor. This amount of power could run roughly 600,000 H100-class GPUs—currently the largest coherent cluster available.

Ejaaz: My question is, why is this registered under Elon Musk's personal name instead of xAI or Tesla?

Josh: I have a few hypotheses. One obvious possibility is that he wants the company to serve multiple entities, primarily Tesla. Tesla is ostensibly an independent company from xAI, despite ongoing rumors of a merger. So I think it’s more about tax and structural reasons. In fact, the reason we even know about this transaction is that one of APR Energy’s investors was required to disclose a $50 million return. That’s the only clue we have.

How severe is the power bottleneck for AI?

Josh: If we take a step back, everyone says energy and electricity are the next bottleneck for AI, but I think many people don’t truly understand the issue. In 2023, data center electricity demand was around 23 gigawatts—already an astronomical figure, and we didn’t have enough supply to meet it. By 2026, that number has doubled to 46.5 gigawatts. But the problem is, this year we’re only planning to add 12 gigawatts of power capacity, far below the 46.5-gigawatt target. Even worse, more than half the year has passed, and only 5 gigawatts have actually been completed. The bottleneck is severe; connecting energy to the grid is extremely slow.

Some might ask, isn’t the western U.S. rich in energy? Yes, but connecting that energy to the grid is extremely difficult. We need high-voltage transformers, grid infrastructure, and various permits and regulatory approvals. The entire process takes five to seven years—I’m not exaggerating; it’s a timeline exceeding six months. In a sense, Musk’s acquisition is operating in a gray area. The Clean Air Act prohibits relocating gas turbines to data centers to power GPUs; this isn’t illegal, but it’s not entirely legitimate either. After acquiring the company, he can now power GPUs under the Clean Air Act by leveraging already-permitted gas turbines without triggering any alerts. Musk is extremely clever—he’s finding ways to be the fastest to bring GPUs online so he can train the best models. Meta is doing the same thing, using Grok 4.5 and future models for the same purpose.

Ejaaz: It’s a race to power the GPUs. The issue isn’t energy itself—it’s the infrastructure. We have oil, we have natural gas, but connecting them to the required infrastructure is extremely difficult. If you think of the U.S. as a vascular system, all the power lines are the veins, and the grid is already under immense strain. I remember a decade ago, when discussing electric vehicles, people were saying that just charging all the Teslas would put enormous pressure on the grid. Back then, we barely kept up—and now we’re barely keeping up still. Adding electricity demand on the scale of entire cities on top of that is a major challenge.

The solution is a modular approach—off-grid and self-powered. There are three options: First, solar power, which requires the largest land footprint, captures sunlight to charge batteries for electricity, but has low energy density and faces significant permitting challenges. Second, nuclear energy, which is still far from being viable for data centers. Third, on-site turbines connected directly to natural gas pipelines to generate power locally. Data centers are essentially building their own grids. In the future, they might even feed excess power back into the main grid, but the immediate priority is powering the data centers—and the best way to do that is with self-generated power. This is the foundation of Musk’s investment: he acquired a company that owns these turbines, aiming to deploy them to data centers faster than anyone else to power up chips.

The pullback of memory stocks and the divergence from fundamentals

Josh: New York State has just banned a batch of data centers, which will delay local data center construction by five years. It’s insane that we have to navigate all these bureaucratic hurdles. But then again, if the idea of gas turbines on trailers sounds familiar, you might be thinking of publicly traded companies like Bloom Energy—we’ll get to them later. Before that, we need to talk about the internal fund flows in AI trading, because we’ve spent a lot of time on memory in this show. Memory is a core component for GPU training and inference, and its price has averaged a 300% to 500% increase over the past nine months. Demand has been insane. High-bandwidth memory is the most obvious example, but NAND flash memory too. However, these stocks have recently been under pressure.

Ejaaz: Memory stocks have definitely been hammered. If you’ve only held them for two weeks, it’s been tough. But if you’ve held longer, congratulations—you’re still up significantly. There’s been a back-and-forth dynamic: prices rose about 20% daily for several months, and now they’ve collectively pulled back roughly 20% from their highs. Interestingly, if you compare memory stock prices to spot memory prices, the spot prices are still rising. This month alone, DRAM average prices have increased nearly 20%. So while stock prices are down 20%, spot prices are up 20%.

Josh: The demand curve hasn’t slowed. Memory stocks are being sold off, but memory prices continue to rise—yet people seem to be growing weary of this narrative. Capital is beginning to flow toward more imaginative opportunities. The market is extremely emotional; take memory, for example—demand remains robust, prices are rising exponentially, and long-term supply agreements (LTAs) confirm this. SK Hynix, one of the world’s top two memory suppliers (only three exist globally), has already secured 40% of its projected profits for next year from 13 to 15 customers. In other words, supply for 2027 is already sold out—regardless of what happens with memory supply next year, these customers are obligated to pay.

If you look at Micron, it has dropped about 24% over the past month. Yet its forward P/E is only 7x, with revenue growth of 350% and a gross margin of 85%. It’s rare in the hardware sector to find a business with such margins. You’d only talk about a memory bubble if there were oversupply—but in reality, the fabs producing these chips are not yet oversupplied; this bottleneck won’t ease until around 2030. People are simply reacting emotionally, and I believe this is an oversell. If you want to understand why, take a look at South Korea.

Market volatility in South Korea weighs on memory stocks.

Ejaaz: Just a reminder, the two largest memory suppliers are both in Korea: SK Hynix and Samsung. Over the past two weeks, the Korean market has been in the red because many Korean investors were over-leveraged by about $1 billion. As a result, $1.5 trillion in market value evaporated. Of course, that number is an order of magnitude larger—I’m being a bit tongue-in-cheek. But the point is, the market is overreacting; the fundamentals are still there, and memory remains an important trade, though people are now looking elsewhere—perhaps toward electricity.

Josh: This is probably the rotation happening right now. People are saying, "I’ve had enough of this toy." The fundamentals are still strong, but everyone has made a big profit and might be looking for the next opportunity. The reason we’re recording this episode is because it looks like capital is shifting toward energy. Power trading is one of the trades I’m most excited about, because it’s one of the most enduring and essential elements of any societal progress. Even if all data centers shut down tomorrow, demand for electricity would remain enormous.

The power demand for U.S. data centers doubled from 31 gigawatts to 66 gigawatts in just 24 months—that’s insane. Data centers now account for 3% of total U.S. electricity consumption, up from 1%, and the share will only keep rising. Demand is skyrocketing, while our infrastructure simply can’t keep up.

Four-layer architecture of the power technology stack

Ejaaz: So who is creatively solving this problem? Whoever can power up a data center the fastest will attract the capital. If you can produce an electron for a data center at a lower cost, that’s an infinite money printer. Orders for gas turbine blades are booked out for years. It’s an extremely difficult challenge, but that’s where the focus lies.

Josh: Let’s break down the layers of the power technology stack. The first layer, I call it the "quick patch"—what Musk just did. Buy a company, build gas turbines, mount them on trailers, and park them at data centers. The All-In Podcast used a precise term for this: “behind the meter”—you position the turbine behind the meter to power it directly, which is technically compliant. The advantage is that it can be deployed in days; the downside is limited power output. Musk’s 1 gigawatt is only enough to partially support his average 3-gigawatt data center footprint, and it can only last 6 to 12 months.

The second layer is Bloom Energy, which produces solid oxide fuel cells. It’s also a large container that can be transported on-site to convert natural gas into electricity with higher efficiency and a lifespan of four to seven years. Why is everyone so excited? Because if you normally have to wait five to seven years to get a transformer, now you can get this unit sooner and train cutting-edge models faster than your competitors. Meta and a group of data centers in Mexico are already using it. But the issue remains regulatory approval—the New Mexico regulators have twice denied permits for the natural gas pipeline. You have the equipment, but you can’t get the operating permit.

GE Vernova: The TSMC of the power industry

Josh: GE Vernova has positioned itself at the center of this trade. They manufacture turbines and grid equipment, and their stock has risen 300% over three years. Orders are booked through 2031, making revenue highly predictable. Orders in 2025 doubled year-over-year to $7.1 billion. Similar to Bloom Energy, as long as you can produce electricity, people will buy it. When one company hits a wall, there’s always a competitor that hasn’t. GE Vernova is one of those companies that hasn’t hit a wall.

Ejaaz: I view GE Vernova as a longstanding pillar of the power industry. They’ve been around forever, understand traditional transformers and high-voltage equipment, and have established supply chain relationships. They also manufacture their own gas turbines. Look at their clients: a $7 billion deal with Microsoft, with OpenAI as one of their core customers. GE Vernova is the TSMC of the power industry. They’ve been growing at an average year-over-year rate of about 30%, but I believe that growth rate will become exponential over the next 6 to 12 months.

Nuclear power and long-term power supply contracts

Josh: The top layer is nuclear energy. While most nuclear companies won’t come online until 2031 to 2035, a company called Valor is accelerating this process by developing modular nuclear power plants.

Ejaaz: To add, IPPs are independent power producers that own their own power plants and sell electricity directly to the market rather than to regulated utilities. This distinction is important because handling approvals, generating power, and selling it back themselves appears to be the optimal solution. Major corporations have already begun signing 10- to 20-year fixed-price power purchase agreements with IPPs. For example, Microsoft acquired the Three Mile Island nuclear plant and signed a 20-year agreement to secure 100% of its output. The long duration of these contracts underscores the scale of this electricity trend.

Josh: Electricity is essentially a currency now. With electrons, you can power smart devices, service tokens, and make money. Locking in 20 years of electricity is far more imaginative than issuing 20-year government bonds with a 3% yield. Nuclear energy opportunities are exciting, but it’s still too early—no reactors have been brought online, and permits aren’t finalized. But when they do go live, it will be enormous.

This is what memory trading looked like before it became "memory trading."

Ejaaz: Talking about the structure and contracts of these energy companies reminds me of memory trading. Before memory became "memory trading," not many were sure if it would be the next big opportunity. History doesn’t repeat itself, but it often rhymes. Most people haven’t heard of GE Vernova yet. Electricity is harder to understand than memory—memory is easy to grasp: "AI needs memory." But electricity? "Isn’t electricity just everywhere?" That’s a more subtle question.

Josh, what are your thoughts on the bull and bear markets?

Josh: I’m bullishly forever on power trading. At its extreme, electrons are more valuable than dollars, and this will always be true. Throughout all of human history, there has been only one trend: more energy equals more productivity equals more innovation equals more prosperity. The more electrons you can invest into a problem, the better the outcome. Short-term looks good, long-term looks excellent, medium-term is uncertain—but power demand will always move upward. Memory was the first trade in early 2026, and ironically, it peaked the same week SK Hynix listed on Nasdaq. Now the question is: what’s the next scarce input? Musk answered it with a $1 billion acquisition: it’s power.

Ejaaz: I’m also bullish for several reasons. First, I like that this AI infrastructure layer is completely model-agnostic. No matter who created your model or which country it’s from, you still need energy and GPUs. Over the past week, everyone has been discussing China vs. the U.S.—China just released Kimi K3, a 2.8-trillion-parameter open-source model, and 56% of tokens on OpenRouter are already being consumed by Chinese open-source models. But Jensen Huang isn’t worried, because regardless of who builds the model, they still need to buy GPUs. The same goes for electricity—whether the model is open-source or proprietary, expensive or cheap, cutting-edge or not, you still need energy.

Second, when we talk about energy, we can’t ignore Musk simultaneously buying APR Energy on Earth for a short-term solution while planning to use satellites in space to collect solar energy. If this doesn’t demonstrate how seriously people are thinking about future energy needs, I don’t know what does.

Josh: That’s the current state of energy, electricity, and power trading. Where the next trade might be. Again, this is not investment advice. I haven’t bought anything myself, though maybe I should have. The direction feels right.