Bitcoin Mining is Changing in 2026 | with FRED THIEL

This year has been a year where we have been quietly putting in place the pieces that will be very evident to the marketplace in '26 regarding what it means when you own large amounts of power. You have land and the ability to build a variety of types of data centers, and you're able to deploy and operate tier 3 and tier 4 data centers for the most sensitive data types for the most critical customers that need to operate with full sovereignty.

Thank you. It has.

Quite a long time.

Excellent.

Yeah. Mara was originally a company that did vanadium mining, which is a mineral. So they've been a miner for quite some time, but they didn't do very well at that. That was back in 2010. Then they did some real estate and then the company did some oil and gas. Eventually, it was taken over by a group who turned it into a patent troll, Marathon Patent Group. As a patent troll, they acquire typically bankrupt companies' assets. One of the patents they acquired was a very basic patent for using a limited instruction set for voice to control a portable or mobile device. That just happened to tick the box that Siri is a voice-controlled device, as is Alexa. Marathon Patent Group sued Apple, and Apple paid a license fee. They then got very excited and sued Amazon, and that case is still going on. Marathon partnered with Fortress Investments, which has an arm that litigates patent suits, so Fortress is running that process. In 2017, a group took control of Marathon Patent Group and decided to go into the mining business. They brought on board a friend of mine, America Komoto, who was a lower mid-market finance guy who had run hedge funds. He had to first fix the balance sheet of the company. I was asked to join the board in early 2018 because I was familiar with Bitcoin and crypto, and the other board members weren't. We spent some time changing players on the board and getting everything ready. We lit up the first Bitcoin mine up in Canada in a building where, when we turned on the miners, there was such a vacuum from all the fans that it took the paint off the walls, and that paint came into the miners and trashed them. There's a great video of this somewhere on the internet from a movie we did called 'The Road to 23 Exahash'.

We had to reboot and start all over again. Long story short, we had some Bitcoin mining going at a site adjacent to a coal-fired plant in Hardin, Montana. We had problems there when a weather event caused part of that coal plant to go offline completely. I stepped off the board and into the CEO role in April of 2021 as employee number five. We had essentially no hash rate running. We decided it was time to really grow. We had been able to raise capital through an ATM, or at-the-money offering. In 2021, the Bitcoin price was on a tear, and everybody was interested in investing in Bitcoin miners. We raised quite a lot of money, placed a huge order with Bitmain for 70,000 machines, and then we just had to find someplace to plug them in. The Bitcoin mining industry really depends on three things—capital, compute, and capacity—and we were an asset-light miner, looking for hosting partners. Over the course of 2021, we started deploying some miners. We did a convertible bond offering at the end of 2021, raised $870 million, placed a very big order with Bitmain for the state-of-the-art S19 XP, and deployed more in 2022 as the market was crashing. By the end of 2023, we were the largest miner in the world. Towards the second part of 2023, we realized we really needed to own the infrastructure we operated in. You couldn't be asset-light. When you work with a third-party partner, you rely on them to operate the machines and maximize uptime. We found we were much better at that than any of the hosting partners. We were able to acquire 70% of the capacity where we were hosted for less than it would have cost us to build it.

You had Riot and Core Scientific.

In late 2021, most of them tried to go public by acquiring a shell. Riot I think actually did an IPO, but I'm not sure. They went public much earlier because they were in a different business. A lot of these were already public companies that somebody repurposed.

It was basically what's called an empty shell.

Exactly. So it was a hot idea. At the end of 2021, our stock hit a peak of $81 a share with a multi-billion dollar market cap. We weren't a huge holder. We had so much cash on the books in January of 2021 that we bought $150 million of Bitcoin on the market because we didn't have our miners deployed. We had to do something with the money.

We were kind of at a disadvantage because we didn't own our infrastructure.

They owned their own data centers. We were just hosted.

I think it's all changing. The Cipher and Iris Energy of the world are going down the AI route very quickly. The capex spend is very asset-heavy because GPUs are 10 to 15 times more expensive to build a data center with miners or GPUs than Bitcoin miners. A data center for bare metal, renting out GPUs to a hyperscaler versus a Bitcoin mining operation, might cost a million dollars a megawatt for infrastructure and miners, while the other costs 12 to 20 million a megawatt.

Let me reframe it. Bitcoin miners are really good at acquiring land and power and building very low-cost infrastructure, using containers or tents. These are tier 1 data centers. Most AI data centers need to be tier 3 or tier 4, which are very different designs. Bitcoin miners are good at acquiring inexpensive power and buying ASICs. The Bitcoin mining industry went from CPU to GPU to FPGA to ASIC very quickly, and the AI market is doing the same thing. Google, Meta, and others are designing their own ASICs or buying from Broadcom instead of using Nvidia chips. The machines in AI also become rapidly irrelevant due to a rapid technology obsolescence curve. H100s were the thing three years ago, now it's B300s, which are much faster and use less energy. They require a huge amount of power infrastructure, and every 3 to 5 years you have to rip out and replace everything, just like in Bitcoin mining. If you think about a data center, it's land and power, then you choose whether to run Bitcoin mining or AI. In AI, there are training workloads, which are high-intensity and constant, and inference workloads, which are short bursts with varying power demand. Grid operators may ask Bitcoin miners to curtail, but it's hard for AI data centers to do that. Marrying Bitcoin mining and AI in the same data center gives flexibility. In the US grid today, there is an estimated demand of 40 gigawatts of power for AI, but a study found 70 gigawatts available if operators were flexible just 2% of the time. The problem is location and access to internet and water for cooling.

Historically, that's changing. You can put data centers in space these days, and Starlink's satellite network makes it viable to do inference in remote locations. Over time, inference will look like Bitcoin mining, with near-remote and metro-level data centers and an orchestration layer balancing between them. Similar to Bitcoin, Mara's pool assembles transactions into blocks and directs our miners in 16 data centers on four continents to solve those blocks. In AI, you'll find a similar system develop around inference with front-end load balancing and orchestration.

A couple of years ago, as we were buying infrastructure and sites, we decided to also buy power generation—wind farms, flare gas from oil fields—to build a portfolio of sites where the average cost to mine kept moving downwards. If you own a wind farm, you don't pay for the wind; it's essentially free energy when it blows. Flare gas is very low cost, maybe a penny a kilowatt. For grid-connected Bitcoin mining, the average price is between 4 and 5 cents per kilowatt hour. For us, it's between 3 and 4.5 cents, depending on the time of year and curtailment. We have a mix of behind-the-meter and grid-detached power assets. In a data center, you can use power for Bitcoin mining, host an HPC customer, or inference AI. You could section off half a center for a smaller HPC customer. The future company needs somewhere to host their data and models, run them, and have traditional IT infrastructure. Most companies want private cloud because it's expensive to run things in the public cloud. 70% of corporate data is not in the public cloud because companies don't want it available and used publicly. When you upload your health records into ChatGPT, you're giving them the right to use that data. Corporations building models for seismic data, healthcare, or finance will never put that data in the cloud. Hyperscalers are focused on public cloud infrastructure, but no one is really servicing enterprise clients who want private data clouds.

If you go to JP Morgan and say you want to offer a private cloud as a Bitcoin miner, they'll just look at you and say you have no idea what you're talking about. That was one of the reasons we made the Exion investment. To be taken seriously offering private cloud and inference AI, we needed a partner with credibility, customers, and advanced technology. Exion is a subsidiary of EDF, the largest electricity producer in the world, operating 58 nuclear reactors. They run tier 3 and tier 4 data centers for nuclear data, which are ultra-secure and sovereign. These data centers cannot go down.

The deal with Exion is purely an investment in a tech company. EDF will retain a percentage, so they have a vested interest in using Exion as a vendor. With our investment, Exion will have the growth capital to address a much bigger market. The deal will hopefully close in the first quarter of next year, pending regulatory approval in France. We've been very active in France, establishing a headquarters for Europe and Middle East operations. Three of the top five electrical energy companies in the world are French, and many of their assets are in the US, Brazil, the Middle East, and Africa. There's a lot of excess generating capacity that isn't being sold, which is perfect for Bitcoin mining. As I said four years ago at Mining Disrupt, by 2028, you either have to be an energy company or partner with one, because energy companies will want to do what you do. Deploying large flexible loads to balance the grid is key. In France, nuclear power is over 60% of their energy, with the rest renewable. They have decreasing demand for energy, so they need load balancing. Europe will be an attractive market for Bitcoin miners, but it's more highly regulated, so you have to work from the top down and the bottom up.

My opinion is different from the press. In September and October, Bitcoin price was nearing $120, AI stocks were through the roof, and adding AI to a coffee would give a 10x multiple. There was huge frothiness. While there is huge demand for AI capacity, enterprises have barely started finding viable use cases. It's like the internet in the late 90s—3 to 5 years later, you had Amazon and Salesforce taking off. The market is coming, and it will be predominantly around inference and private models. Why pay for models full of consumer data when you have proprietary information? You're starting to see the rise of open-source models taking over in the enterprise market. Investors have realized that the circular OpenAI economy between Nvidia and OpenAI has serious challenges. Bitcoin mining, similarly, when Bitcoin was at $120, things were great, but then it dropped 25%. In 2021, Bitcoin was at $61,000, and by January of '23, it was $15,000—a 75% drop. The current volatility has dropped from 75% to about 50%. You have corporate buyers, DATs coming in, and top banks now custodying or allowing you to hold Bitcoin. Wealth advisors are recommending a 2 to 4% allocation. Soon, the SEC will allow Bitcoin in 401ks. But whales have started selling. If you bought Bitcoin for $10 and it's at $100k, you might want to take chips off the table. They can sell, convert to an ETF, or borrow against it tax-free.

Exactly. Now you have a huge pool of liquidity in the market. Bitcoin is a highly liquid asset. You're seeing the IPO stage where initial investors cash out and new investors come in. Most of the selling has been from people who held Bitcoin for less than 150 days, buying on the way up. When Bitcoin falls out of favor with retail because they're chasing the AI bunny, there's pressure. I have huge respect for Michael Saylor; his strategy of buying and holding Bitcoin and creating the first real debt was brilliant. But over time, you have to keep that machine fed. When you trade below your NAV, you become less attractive.

I don't doubt that Michael will come up with something new. He attracted lots of copycats who diluted the strategy. A lot of latecomers have suffered, like Metaplanet. Over 50% of our Bitcoin holdings are mined, and the rest is purchased. The DAT model will have one or two remaining. I think MicroStrategy will continue to exist—Michael will adapt and find new ways to raise capital. We are the second largest publicly traded holder of Bitcoin, and for us it's an item on our balance sheet. We generate yield with it, and it's there to use as a growth capital source maybe someday. As long as Bitcoin generates appreciation greater than what we can get with other capital, we'll keep it there.

There are three drivers: capital, compute, and capacity. The longest pole is capacity, which takes the longest to get. You need land and power. Back in the last cycle, Bitmain controlled the market and kept pricing high by constraining supply. In 2022 and 2023, the price of miners came down, and many miners bought machines in 2024. Now, several miners are at or near our scale. Many miners are hurting from that 2024 bulk purchase that hasn't accrued the value they hoped. Global hash rate has continued to grow. Tether is buying every small miner they can with Electron, but it will be a challenge for them to integrate. For small miners, there will be consolidation. If you want to service hyperscalers like Microsoft or Google, you need a lot of power. Small miners aren't going to do that. You need to have long-term power agreements and be able to pivot. We'll see some companies that focus on being real estate plays for AI, and others that focus on Bitcoin mining. The market is changing, and it's going to be exciting.

Small miners have 25, 30, 40, 50 megawatts.

They can't get into the HPC game. They can't get into a hyperscaler game. They just don't have the capital. They don't have the ability. What they can do potentially is essentially offer up their land and power to somebody who has the capital who would want to build inference AI sites because inference is just a different type of animal and it can be done.

I think you have to look at it like a real estate developer does. That's how we have reorganized ourselves around asset management. Every data center is an asset, and how do you optimize the value of that asset? There are all sorts of things you can do with that asset. You can run it as a Bitcoin mining site, you can run it as an HPC site, you could run it as an inference AI or traditional IT workload site. It depends on where it's located, price of power, curtailment, etc. We have also always been a technology first company. So we lead with the technology that enables us to do what we do. When we started running inference in Granbury, we're running air cooled inference in containers the same way we run Bitcoin mining. Actually in Granbury today, we have more immersion Bitcoin mining than we do air cooled. But if you can deploy inference AI like you deploy Bitcoin mining in containers, it means you can do it really fast. Because the biggest problem is if you're going to build a building, put network, put cooling, put racks to run training or inference the way most people do it, it's hugely expensive. Just the building, the refrigerator if you would, is $12 million a megawatt versus $400,000 a megawatt for a bunch of containers that run Bitcoin mining. If you can not have to do that and you can then focus your capital on buying the right GPUs or the right ASICs and you have the right software stack above that that lets you monetize every compute cycle and optimize every compute cycle, then essentially a mining site that is a hybrid of mining and AI, you can do arbitrage on every electron you're operating. You can say, okay at this site my power cost is maybe 2 cents a kilowatt hour for these hours of the day. I'm going to run Bitcoin mining all day long and then when my price starts changing I'm going to run more inference loads.

The only thing that matters today is owning power. Jensen Huang said it two or three days ago in an interview. He said if you look at the pyramid of Nvidia and AI in the world, the US is really strong on chips and data centers. What the US sucks at is building a grid and power generation. And China is eating the US's lunch.

China has grown their national grid with such speed, like they do with many things. It's top down. In the US, it's very messy. We've had a permitting regime which has been very difficult to live with the past administrations.

They're trying to change it. The FERC and the DOE, it's difficult because when a new administration comes in, they put new people at the top, but all the people that actually are responsible for doing the work are career civil servants. Those people don't change from one administration to another. It's like saying to the people who have been responsible for reviewing your tax returns, 'Well, now we want to be nice to taxpayers as opposed to really find every last penny.' It's like, 'Well, no, that's not my job.'