Podcast Notes | Musk Spent 1 Billion Out of His Own Pocket to Buy a Fleet of Power Generators, Just to Power GPUs?

Organized & Compiled: TechFlow

Guests: Josh Kale (AI Analyst, Anthropic Contractor), Ejaaz Ahamadeen (Limitless Podcast Co-host)

Hosts: Josh Kale & Ejaaz Ahamadeen Dialogue

Podcast Source: Limitless Podcast

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

Release Date: July 22, 2026

Disclosure: Josh Kale is an Anthropic contractor. The views expressed in this episode are his personal stance and do not represent Anthropic. Both parties declare that the content of this episode does not constitute investment advice.

Key Takeaways

Elon Musk spent approximately $1 billion of personal funds to quietly acquire a company called APR Energy. This company does not manufacture chips, build rockets, or develop AI models. It does one thing only: mounts gas turbines on trailers, transports them to your data center campus, and powers up your GPUs within a few days. The fact that someone who has preached solar transformation for 20 years turned around and bought fossil fuel power generation equipment itself illustrates where the real bottleneck in the AI arms race lies.

This episode starts with this acquisition and deconstructs two rotations currently happening: first, memory stocks (DRAM, HBM, NAND) moving from surge to pullback, but underlying demand remains unchanged; second, capital beginning to flow into the power infrastructure sector. The two hosts梳理 (sorted out) the complete power technology stack from mobile gas turbines (APR Energy), solid oxide fuel cells (Bloom Energy), gas turbines + grid equipment (GE Vernova) to small modular nuclear reactors (Valor), believing that power is the next frontier for AI investment in the second half of 2026.

Highlights Summary

About Elon Musk's Acquisition Logic

"1 Gigawatt of power is enough for 750,000 households, equivalent to the output of a nuclear reactor. This electricity is enough to run approximately 600,000 H100-level GPUs."

"The reason we know about this deal is because one investor in APR Energy had to disclose a personal return of $50 million. This is the only clue."

"He chose to buy with personal funds rather than under xAI or Tesla, possibly due to tax and structural considerations."

About the Power Bottleneck

"In 2023, data center power demand was approximately 23 Gigawatts; by 2026, it has doubled to 46.5 Gigawatts. But the plan to add this year is only 12 Gigawatts, and actually only 5 Gigawatts were completed."

"Connecting power to the grid requires high-voltage transformers, grid infrastructure, various permits, and regulatory approvals. The whole process takes 5 to 7 years, I am not exaggerating."

"New York State just banned a batch of data centers, which will delay local data center construction by 5 years."

About the Pullback in Memory Stocks

"In July, the average DRAM price rose nearly 20%, but memory stocks fell 20%. Prices are still rising, but stocks are falling."

"SK Hynix has already signed supply for the entire year of 2027. 13 to 15 clients have locked in 40% of next year's expected profits. Demand is overwhelming."

"Micron fell 24% in one month, but forward P/E is only 7x, revenue growth 350%, gross margin 85%. Such high profit margins are rare in the hardware sector."

"Korean investors leveraged approximately $1 billion excessively, and the market evaporated $1.5 trillion as a result. Fundamentals are still there, this is just oversold."

About Power as the Next Trading Clue

"electrons are more valuable than dollars. There is only one trend in all human history: more energy equals more productivity equals more innovation equals more prosperity."

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

"GE Vernova is the TSMC of the power industry. Orders are already booked until 2031."

"Microsoft signed a 20-year power supply agreement with the Three Mile Island Nuclear Station, taking 100% of the output. Locking in 20 years of power is much more imaginative than issuing 20-year Treasury bonds with a 3% yield."

About the Model-Agnostic Nature of Energy Demand

"Regardless of whether your model is open-source or closed-source, expensive or cheap, frontier or non-frontier, you need energy and GPUs."

"China just released Kimi K3, a 2.8 trillion parameter open-source model. 56% of token consumption on OpenRouter has already been spent on Chinese open-source models. But Jensen Huang is not worried, because regardless of who makes the model, they have to buy his GPUs. Power is the same."

Main Text

What Did Elon Musk Buy for $1 Billion

Ejaaz: Musk just spent over $1 billion of personal funds to buy a company no one has heard of. It's called APR Energy. The strangest thing about this company is that it doesn't make AI chips, build rockets, or do any AI models. What it does is gas turbines, mounted on trailers, transported to your data center campus, powering up your GPUs within a few days. A person who spent 20 years talking about solar energy made this acquisition. When you deconstruct why he bought it, you will find this reveals a new trading clue in the AI field, and also explains why memory stocks have been falling recently. Energy itself in the US is sufficient, but connecting this energy to the billions of dollars worth of GPU clusters coming online this year is an extremely difficult task.

Josh: APR Energy has quite an interesting history. Founded in Jacksonville in 2004, listed on the London Stock Exchange in 2011, acquired GE's energy leasing business in 2013, becoming the world's largest mobile gas turbine lessor. Afterwards, it kept changing hands between private equity firms. We discovered through searching regulatory filings that Musk spent approximately $1 billion to buy it. What did $1 billion buy? Approximately 1 Gigawatt of power. For reference, 1 Gigawatt is enough for 750,000 households to use electricity, equivalent to the output of a nuclear reactor. This electricity is enough to run approximately 600,000 H100-level GPUs, which for now is already the largest coherent cluster.

Ejaaz: My question is, why is this thing under Musk's personal name, rather than xAI or Tesla?

Josh: I have a few guesses. One obvious possibility is he hopes this company serves multiple entities simultaneously, mainly Tesla. Tesla is theoretically independent from xAI, although there have always been rumors of merging. So I think it's more for tax and structural reasons. In fact, the reason we know about this deal is because one investor in APR Energy had to disclose a return of $50 million. This is the only clue.

How Severe is AI's Power Bottleneck

Josh: If we take a step back, everyone says energy and power are AI's next bottleneck, but I think many people don't truly understand this problem. In 2023, data center power demand was approximately 23 Gigawatts, which was already an astronomical number; we didn't have enough supply to meet it. By 2026, this number has doubled to 46.5 Gigawatts. But the problem is, this year we only planned to add 12 Gigawatts of power supply capacity, far below the 46.5 target. Worse still, now more than half the year has passed, and actually only 5 Gigawatts were completed. The bottleneck is very severe; the speed of connecting energy to the grid is very slow.

Someone might ask, isn't the US West rich in energy? Yes, but connecting this energy to the grid is extremely difficult. We need high-voltage transformers, grid infrastructure, various permits and regulatory approvals. The whole process takes 5 to 7 years, I am not exaggerating, it is a cycle of over half a decade. This acquisition by Musk, in a sense, is walking a gray area. The "Clean Air Act" stipulates you cannot move gas turbines to data centers to power GPUs; this isn't illegal, but it's not entirely legitimate either. After he bought this company, he can power GPUs under the "Clean Air Act" framework as permitted gas turbines, without triggering any alarms. Musk is very smart; he is figuring out how to become the fastest person to get GPUs online, so he can train the best models. Meta is doing the same thing; he is doing the same thing with Grok 4.5 and subsequent models.

Ejaaz: This is a race to power GPUs. Infrastructure is the problem, energy itself is not. We have oil, we have natural gas, but the infrastructure needed to connect them is very difficult. If you imagine the US as a vascular system, all transmission lines are blood vessels; the grid is already under huge pressure. I remember discussing electric vehicles ten years ago, saying how much pressure just charging all Teslas would put on the grid. Back then we barely kept up, now we are still barely keeping up. To add city-scale electricity consumption on top of this is a big trouble.

The solution is modular solutions, not connecting to the grid, bringing your own power. There are three paths: the first is solar, largest land footprint, relying on absorbing sunlight stored in batteries then powering, but density is not high enough, and permits are hard to get. The second is nuclear, but far from being available for data centers. The third is bringing your own turbines, connecting to natural gas pipelines, generating power directly on-site. Data centers are actually building their own grids. Maybe in the future they can power the main grid in reverse, but currently the core demand is powering data centers; the best way is bringing your own power. This is also the basis of Musk's investment; he bought a company owning these turbines, to transport them to data centers, lighting up chips faster than others.

Pullback in Memory Stocks and Divergence from Fundamentals

Josh: New York State just banned a batch of data centers, which will delay local data center construction by 5 years. It is crazy that we have to go through all this red tape. But having said that, if gas turbines on trailers sound a bit familiar, you might think of listed companies like Bloom Energy. We will talk about that later. Before that, we have to talk about capital rotation within AI trading, because we spent a lot of time in the program talking about memory. Memory is a core component of GPU training and inference; prices averaged a 300% to 500% increase over the past 9 months. Demand is very crazy. High bandwidth memory is the most obvious, NAND flash is too. But these stocks have been suppressed recently.

Ejaaz: Memory stocks were indeed hammered. If you only held for two weeks, it was indeed uncomfortable. But if you held longer, congratulations, you still made a killing. There is a feeling of back and forth tug-of-war: previously rising 20% every day for several months, now collectively pulled back about 20% from highs. Interestingly, if you compare memory stock prices and memory spot prices, spot prices are still continuing to rise. Just this month, the average DRAM price rose nearly 20%. Stock prices fell 20%, spot prices rose 20%.

Josh: The demand curve has not slowed. Memory stocks are being hammered, memory prices are still rising, but everyone seems a bit fatigued with this narrative. Capital is starting to flow to more imaginative places. The market is extremely emotional; taking memory as an example, demand has not shaken, prices are rising exponentially, and Long-Term Agreements (LTA) also prove this. SK Hynix, one of the top two memory suppliers globally, there are only three in the world, 13 to 15 clients have locked in 40% of next year's expected profits. That means 2027 supply is already sold out; regardless of what happens to memory supply next year, these clients have to pay.

If you look at Micron, it fell approximately 24% in the past month. But forward P/E is only 7x, revenue growth 350%, gross margin 85%. Businesses with this kind of profit margin are extremely rare in the hardware sector. If you want to say memory bubble, that would have to wait until supply exceeds demand. But the fact is, fabs producing these chips are not yet oversupplied; this bottleneck won't open until around 2030. People are just venting emotions; I think this is oversold. If you want to find a reason, look at South Korea.

South Korean Market Turbulence Drags Down Memory Stocks

Ejaaz: Reminder, the two largest memory suppliers are both in South Korea: SK Hynix and Samsung. The South Korean market has been red for the past two weeks, because many South Korean investors leveraged approximately $1 billion excessively. The market therefore evaporated $1.5 trillion in value. Of course, this number is an order of magnitude too large; I am joking a bit. But the point is the market is overreacting; fundamentals are still there; memory is still an important trade, it's just people are looking for other things, maybe power.

Josh: This is probably the rotation happening. Everyone feels, "played with this toy enough". Fundamentals are still strong, but everyone made a lot of money, possibly looking for the next target. The reason we recorded this episode is because it looks like capital is flowing to energy. Power trading is one of the trades I am most excited about, because it is the most enduring, most necessary thing in any social progress. Even if all data centers shut down tomorrow, power demand is still huge.

US data center power demand doubled from 31 Gigawatts to 66 Gigawatts within 24 months; this is too crazy. Data center share of total US electricity consumption rose from 1% to 3%, and will only continue to go up. Demand is climbing vertically, while our infrastructure simply cannot keep up.

Four-Layer Structure of Power Technology Stack

Ejaaz: So who is solving this creatively? Who can power data centers the fastest? Capital will flow there. If you can produce an electron for data centers at lower cost, that is an infinite money printing machine. Purchase orders for gas turbine blades are booked for several years. This is a very difficult problem, but the focus is here.

Josh: Let's sort out the several layers of the power technology stack. The first layer, I call it "Quick Patch", is what Musk just did. Buy a company, build gas turbines, mount on trailers, transport to data centers and park. There is a very precise term in the All-In Podcast called "behind the meter" (behind-the-meter power supply), meaning you park the turbine behind the meter, powering the meter, technically legal. The benefit is online within days; the downside is limited power; Musk's 1 Gigawatt is only enough for part of his average 3 Gigawatt scale data center, and can only last 6 to 12 months.

The second layer is Bloom Energy, making solid oxide fuel cells. Also a big box, can be moved on-site, converting natural gas to electricity, higher efficiency, can last 4 to 7 years. Why is everyone so excited? Because if you originally had to wait 5 to 7 years to get transformers, now you can get this thing earlier, you can train frontier models faster than competitors. Meta and a batch of data centers in Mexico are using it. But the problem is still administrative approval; New Mexico regulators rejected the natural gas pipeline permit application for the second time. You have good equipment, but cannot get operating permits.

GE Vernova: The TSMC of the Power Industry

Josh: GE Vernova places itself at the center of this trade. They produce turbines, grid equipment; stock price rose 300% in three years. Orders are already booked until 2031; revenue is very predictable. 2025 orders doubled year-over-year to $7.1 billion. Similar to Bloom Energy, as long as you can produce electrons, someone will come to buy. Once you hit a wall, there are always competitors who haven't hit the wall. GE Vernova is that company that hasn't hit the wall.

Ejaaz: I like to view GE Vernova as the old pillar of the power industry. They have always been there, know how to handle traditional transformers and high-voltage equipment, and have supply chain relationships. They also make their own gas turbines. Look at the clients they signed: $7 billion deal signed with Microsoft, OpenAI is one of their core clients. GE Vernova is the TSMC of the power industry. They average about 30% year-over-year growth, but my judgment is, in the next 6 to 12 months this growth rate will become exponential.

Nuclear Power and Long-Term Power Supply Contracts

Josh: The top layer is nuclear power. Generally nuclear power companies won't come online until 2031 to 2035, but there is a company called Valor accelerating this process, making modular nuclear power plants.

Ejaaz: Supplementally, IPP is Independent Power Producer, owning power plants themselves, selling electricity to the market rather than serving regulated areas. This distinction is important, because handling approvals yourself, generating power yourself, then selling back, appears to be the optimal solution. Large companies have started signing 10 to 20-year fixed price power supply contracts with IPPs. For example, Microsoft; they took over the Three Mile Island Nuclear Station, signed a 20-year agreement, 100% of output belongs to Microsoft. The time span of these contracts is very large, indicating the scale of this power trend.

Josh: Power is now essentially a currency. You have electrons, you can light up intelligence, serve tokens, make money. Locking in 20 years of power is much more imaginative than issuing 20-year Treasury bonds with a 3% yield. Nuclear energy opportunities are cool, but still too early. No one has reactors online; permits haven't been settled. But the day it comes online, it will be very huge.

This is What It Looked Like Before Memory Trading Became "Memory Trading"

Ejaaz: Talking about the structure and contracts of these energy companies reminds me of memory trading. Before memory became "memory trading", everyone wasn't too sure if this was the next big opportunity. History doesn't repeat, but it rhymes. Most people haven't heard of GE Vernova. Power is harder to understand than memory; memory is easy to say, "AI needs memory", but power is "isn't electricity everywhere?" This is a more subtle problem.

Josh, Bull/Bear View?

Josh: I am bullish on power trading forever. Extremly speaking, electrons are more valuable than dollars; this point will hold true forever. There is only one trend in all human history: more energy equals more productivity equals more innovation equals more prosperity. The more electrons you can投入 (投入 should be translated: invest) into a problem, the better the successful result. Short-term looks good, long-term looks very good, medium-term hard to say, but power demand is forever upward. Memory was the first trade in the first half of 2026; ironically it peaked just the week SK Hynix listed on Nasdaq. Now the question is, what is the next scarce input? Musk answered this question with a $1 billion acquisition: it is power.

Ejaaz: I am also equally bullish. Several reasons. First, I like that this AI infrastructure layer is completely model-agnostic. Regardless of who made your model, which country, you need energy and GPUs. Everyone has been discussing China vs US this past week; China just released Kimi K3, a 2.8 trillion parameter open-source model; 56% of token consumption on OpenRouter has already been spent on Chinese open-source models. But Jensen Huang is not worried, because regardless of who makes the model, they have to buy GPUs. Power is the same; regardless of whether the model is open-source or closed-source, expensive or cheap, frontier or non-frontier, you need energy.

Second, when we talk about energy, we cannot not mention Musk buying APR Energy on the ground for short-term solutions, while planning to use satellites to collect solar energy in space. If this doesn't show how bullish people are on future energy demand, I don't know what can.

Josh: This is the current state of energy, power, and power trading. Where the next trade might be. Again, disclaimer, does not constitute investment advice. I haven't bought anything myself, although maybe I should. Directionally feels right.