In late 2023, Bitcoin miner Ben Gagnon concluded that his employer, Bitfarms of Quebec, was in the wrong business. For Gagnon, Bitfarms’ greatest value lay not in the giant racks of GPUs that generated Bitcoin, but the immense power supplies it controlled. The then 34-year-old chief mining officer who’d launched his career crafting makeshift crypto-minting rigs in Hong Kong and Taipei, reckoned that Bitfarms could make much more money by exiting the field full stop, demolishing the “mines,” selling the chips that filled them, and re-channeling its plentiful flow of megawatts to power-craving industrial users at plants newly built on the Bitfarms sites.
The big revelation came in March of 2024, when Core Scientific and CoreWeave unveiled a stunning new deal. Bitcoin titan Core Scientific would refashion a mining facility in Austin, Texas, as an AI data center, and rent the power and buildings to the neocloud. That Bitcoin-to-AI play struck Gagnon as a model that Bitfarms could follow on a grand scale, from Quebec to Washington state. “It was the big aha moment,” recalls Gagnon. “Suddenly, we saw who the right customer was.”
At the time, Bitfarms was mired in turmoil. In early 2024, the board was searching for its third leader in five years. The sitting chief agreed to depart when Bitfarms named a successor, but in May—believe it or not while still CEO—instead filed a $27 million suit for “wrongful termination and breach of contract,” plunging this Canadian exemplar of all-green, hydro-driven Bitcoin mining into legal warfare. Pounded by a sharp drop in the token’s prices, Bitfarms’ stock was languishing, and a rival pounced. Two days after the board fired the CEO, Riot Platforms unveiled a hostile $930 million takeover bid for Bitfarms at a lowly $2.30 a share, 75% below its peak in late 2021. After Bitfarms installed a “poison pill” defense, Riot launched a website dubbed “A Better Bitfarms” trashing allegedly poor corporate governance, and demanding the departure of the co-founder named as interim chief.
The board tapped Gagnon to both quell the war versus Riot, and pursue the strategy of total transformation he’d been advocating. In July of 2024, Gagnon took the helm, and over the next few months, helped broker an agreement where Riot named a director of its choice, and the co-founder left the board. Riot halted its attack and has since sold its shares.
While battling for Bitfarms’ independence, Gagnon kept his sights locked on engineering the Big Pivot. He describes the mission as “pursuing a strategy of steering the business away from Bitcoin mining, and taking it over to a pure energy infrastructure play in support of AI and high-performance computing. We’d shift our focus entirely to addressing the defining constraint of the most important technology of our lifetime.”
The venture needed a rebrand, so Gagnon picked “Keel” to evoke the structural backbone of a vessel that, as he says, “provides stability and converts energy into forward motion,” a reference to the stable long-term leases that he expects to provide a smooth and lucrative course versus the storms he encountered in Bitcoin. He’s now attempting to pull a hat trick seldom seen in the corporate world, the total reinvention of a large enterprise, including in product and nationality, on a crazy-rhythm beat. On April 1, 2026, the franchise was reborn as Keel Infrastructure and changed its domicile from Canada to the U.S. The photo on Gagnon’s X profile reprises the nautical theme, adding a whimsical accent. Gagnon grins from a porthole-sized frame, sporting his usual oversized rimless glasses, and a gold-braided yachtsman’s cap displaying his user name, “Captain Keel.”
Buying beer instead of Bitcoin
Now 37, Gagnon is making by far his biggest wager yet in a career that careened from mining Bitcoin in remote oil patches by rescuing flared natural gas to rehabilitating some of the world’s dirtiest energy sites under the banner of AI.
Gagnon is the son of two American parents who settled in Hong Kong, where he attended international grammar and high schools that offered lots of, as he puts it, “Mandarin as a second language courses.” He gained a bachelor’s degree in business and economics (and a minor in Mandarin) at the Indiana University’s Kelley School of Business in 2011, on the Bloomington campus. “I got fascinated by Bitcoin, and could have bought it for 73 cents,” he recalls. “Instead, I did what every college kid does. I bought beer.” Over the next three-and-a-half years, he reflected on that choice while watching Bitcoin climb 1,400-fold to over $1,000.
Returning home, he attained a master’s in computer security and networking at the University of Hong Kong, where he took cryptography courses that renewed his interest in the science of generating cryptocurrencies. After graduating, Gagnon worked as a network technician at a company that offered mobile applications for investor conferences and other events. After less than two years, he quit his job and moved again, this time to Taiwan. His new venture: Building and running a jerry-rigged Ethereum mine at the home of his girlfriend’s (now his wife’s) parents in Taipei. “Everyone thought it was Bitcoin 2.0, so I figured, I’m not going to miss this opportunity like I missed Bitcoin as a sophomore,” says Gagnon.
The business drew unwanted attention by using ten times the electricity of a typical home in Taipei—specifically, the phenomenon alerted the police. “They couldn’t figure out what we were doing,” he recalls. “They thought we were growing marijuana upstairs and kept raiding the house, like on a weekly basis. So we had to explain to the police every week that we weren’t doing anything illegal.” It took only a couple of intolerable months before Gagnon saw the necessity of uprooting his facility from his future in-laws’ home.
Gagnon then ran his operation in Guangzhou on China’s mainland, but shut down after two years due to a tough regulatory climate. “Let’s just say that the laws were somewhat selective for a young foreigner navigating a new industry,” he avows. In 2019, he joined Bitfarms, one of several Bitcoin miners, including HIVE, that sprouted in Quebec during the 2017 to 2018 boom, drawn by the province’s plentiful and relatively inexpensive hydroelectric power. As chief of development, his charge was expanding the business to the U.S. and other energy-rich areas of Canada.
Gagnon focused on the huge but remote Alberta oil patch. “I was always looking for the cheapest energy sources for mining, and that was Alberta,” he says. The producers were “flaring” the natural gas created from drilling, since no pipelines existed for sending it to population centers. Gagnon created the first containerized Bitcoin data center that incorporated a generator and offered 100% financing. His mobile units channeled natural gas right at the far-flung drilling sites for minting tokens. His rigs empowered the drillers to monetize energy they were burning into the atmosphere. The product proved a big winner for his new employer. “The drillers were using gas that cost 1.5 cents a kWh versus 6 cents in the U.S.,” he recalls. “So we got great prices for our equipment because the payback period for the drillers was so short.”
When Gagnon started pushing the AI remake in 2023, he doubled-down on a possible acquisition that ideally fitted the new plotline. His target was Stronghold Digital, a startup using one of the most unusual models in all of Bitcoin mining. From the mid-19th century to the 1970s, mines in Pennsylvania routinely separated the low-quality coal—scraps mixed with rocks, often with high sulfur or ash content—from the good stuff. Children called “breaker boys” seated at conveyor belts or chutes would pick out the slate and rock by hand, no gloves permitted—brutal work that often left their fingers bleeding. The legacy: Black carbon mountains scarring the towns of western Pennsylvania. For decades, the sites Stronghold owned had been performing a public service by scooping truckloads of waste coal and burning it in their boilers to supply electricity to the local grids, while pocketing big subsidies for the effort. The state government reasoned that the benefits of cleansing the countryside more than compensated for allowing just about the dirtiest form of electrical generation in America.
In 2021, a pair of entrepreneurs turned those locations into Bitcoin mining operations. But Stronghold way overinvested in chips and other equipment, and by 2023 it was foundering, hammered by crippling debt and legal disputes. Gagnon identified Stronghold’s sweet spots, one each in eastern and western Pennsylvania, as ideal locations for AI campuses. Stronghold already controlled significant amounts of power, but held the potential to obtain much more. Working with its owners, Gagnon applied to utilities and state authorities for big new batches of megawatts. Once assured the approvals were at hand, Gagnon, in August 2024, after a month as CEO, landed an agreement to purchase Stronghold for $175 million.
How power became the chokepoint in the AI buildout
Increasingly, AI leaders are viewing power as the toughest chokepoint in the AI supply chain. As OpenAI co-founder and CEO Sam Altman characterizes the issue, “The bottleneck for AI isn’t code, it’s electrons.” Adds Elon Musk, “The limiting factor is fundamentally electrical power…[that’s leading to] producing more chips than the hyperscalers can plug in and turn on.” And Greg Abel, CEO of Berkshire Hathaway, a major energy provider to AI data centers via its sundry utilities, concurs, calling access to megawatts a “big constraint” for rapid buildouts.
The backlash against AI data centers gained fresh force on Election Day, when so many candidates seeking restraints or outright bans on their spread won congressional and statewide races. That result may actually strengthen the hand of the Bitcoin contingent, since they’ve already got a grip on the energy pipelines the hyperscalers and neoclouds need, and appear willing to pay up for. Put simply, the future “scarcity value” of the crucial asset these crypto producers control likely jumped on November 3. Gagnon calls Keel’s position “bird in the hand economics.”
Keel stands among the first and largest crypto operators to follow Core Scientific into AI. But today, more than ten major players are making the switch, in part or whole. Among those that have already secured big, long-term contracts are Riot, which clinched a $9.1 billion lease from Anthropic for the AI campus at its Bitcoin mine in Milam County, Texas; Hut 8, partnering on a $7 billion contract with Google and Anthropic at its River Bend complex in Louisiana; and Applied Digital, which will serve CoreWeave under an $11 billion agreement at its forthcoming campus in Ellendale, North Dakota.
All of these converts are essentially re-purposing the same land and power, or sometimes more of the latter, that they deployed to produce Bitcoin for the AI buildout. But they’re taking different approaches. The crypto-to-AI crowd falls into two categories, both in how completely they’re shifting, and the templates they embrace. On the first score, several participants are taking the Keel route by totally exiting Bitcoin for AI, including Core Scientific, Cipher Mining and Genesis Digital. Another group that encompasses Riot, Hut 8, HIVE Digital Technologies, and TeraWulf are adopting a hybrid model where they both channel power to AI campuses and continue to mine Bitcoin.
On the second criterion, the two main business models are “compute provider” and “developer” or co-location (co-lo) purveyor. As a compute provider, the former or hybrid Bitcoin outfit delivers a turnkey project. It commissions a contractor, often a REIT, to erect the buildings, and typically rents the structures long-term from the builder-partner. It then packs the cavernous boxes with GPUs and other gear that constitute by far the most expensive part of the project. The owner’s product is the full computing package that it then leases to the hyperscalers or neoclouds for running the likes of Google’s Gemini and Anthropic’s Claude.
The developer or co-lo arrangement works much the same way, except that the site’s owner doesn’t supply the hardware, and hence shoulders much lower up-front costs. Instead, it’s the hyperscaler or neocloud, a Microsoft or Crusoe, that brings the GPUs and other equipment. The company branching out from Bitcoin mining both supplies the structures that it leases from the REIT or other contractor, and the power that runs it––MWs secured on long-term utility contracts form its essential business. The still-active or former Bitcoin miner rents the “powered shells” to the compute providers, for operating, say OpenAI’s ChatGPT or Microsoft’s Copilot. In today’s market, the developer garners a big markup on the scarcest item in AI.
The reasons so many are crossing from Bitcoin to follow either one or both of these strategies: Each megawatt of power at work generates far more cash flow at an AI campus than at plants churning codes to win the signature crypto coin. In a recent interview on the Schwab Network, Christian Lopez, head of blockchain and digital assets at broker-dealer Cohen & Co., estimates the numbers, over five years, at $1.7 million per KW for Bitcoin and almost five times as much, $7.9 million, for an AI campus. What about revenues? Sarah Friar, CFO of OpenAI, posits that each 1 GW (1,000 MW) fueling an AI data center garners $10 billion in revenues. Once again, the number can be one-fifth that amount for a Bitcoin mine.
Data center pioneers divide into two camps. The “compute” competitors (Bitdeer of Singapore, IREN of Australia and Canada’s HIVE Digital, Hut 8 to some degree) spend massively to buy the chips and rent them out, in exchange for higher rents, shorter leases, and more risk. For example, HIVE has a $350 million, 5-year deal with an as yet unnamed AI stalwart for campuses on Bell Canada sites where it’s supplying the hardware. “The cost of the GPUs is $190 million off the bat,” says HIVE CEO Aydin Kilic, who predicts the project will be multiple times as profitable per KW as Bitcoin mining.
By contrast, Keel is a pure-play developer. It doesn’t buy and install the GPUs and other gear that constitute as much as 70% of data centers that cost roughly $50 billion a GW to create (the other 30% covers land, the buildings and costs of connecting or providing onsite power). “We’re strictly in the power business, and make money on the difference between what we pay the utilities, and what the compute providers pay us on their leases,” says Gagnon.
Keel’s edge: It sits on an enormous amount of power. It has 2.2 GWs that are either approved for use, or in the last stages of the approval process. That’s more than the 1.9 GW that OpenAI was reportedly marshaling at the close of 2025. Gagnon acknowledges that Keel as a developer won’t collect revenues nearly as large as the miners turned compute providers, but insists that his cash flows will be more secure, and that the relatively low cost of the power and physical centers as a share of the whole makes customers willing to pay big premiums to get MWs in a hurry.
In part, that’s because as of now, those AI campuses are sumptuously profitable. “The gross margins for data centers are 85%, and of the 15% in costs, only a max of 2% to 4% goes to leasing the power and buildings,” Gagnon explains. “So if a customer spends twice as much, their margins may ease to 81% or 83%. The GPUs in their warehouses are depreciating fast. So they’re saying, I’d rather take a hit to 83% in gross margin and start generating revenue in 6 months than wait 12 months for a margin that’s 2 points higher. And that’s what we’re seeing.”
BREAK
Keel is shepherding three “priority projects” closest to clearing all hurdles and to signing leases. One of them is a former Stronghold facility: Panther Creek, 85 miles north of Philadelphia. Keel acquired another Pennsylvania site in a separate transaction. It’s located at the opposite end of the state in Sharon, near Youngstown, Ohio. On the two sites combined, Keel’s in the process of securing over 450 MWs from the local grids. The third “priority” is a hydro project in Moses Lake, Washington, ringed by Seattle, Spokane and Portland. Gagnon marvels at the sight of the famous Boeing repair field almost adjoining the Keel site. “You’re driving to our facility and you’ll see 100 different 737 MAX, all painted different colors, what a fun sight!”
Gagnon points out that approvals fall into three categories: zoning, environmental, and development. At Moses Lake, Keel’s scored all three, and the site’s now under construction. At Sharon and Panther Creek, it has secured both zoning and development, and is working through the environmental permits, the final step. The most difficult requirement, he states, is zoning. “It’s the one that’s most political, most community-centric, and most emotional,” he says. “Projects can get shut down if you don’t behave the right way with community leaders. We got really good at advocating for Bitcoin mines that are a much harder sell than AI campuses.”
The second ex-Stronghold site is the biggest of them all, the 700-acre Scrubgrass facility near Sharon, and roughly an hour north of Pittsburgh. There, Keel’s applied for a massive 1 GW of power, 40% of the total capacity it’s seeking. At Scrubgrass, Keel is still working through all the initial permitting stages for securing power.
Gagnon says the biggest challenge isn’t attracting customers, it’s getting the final environmental permits in place. But he’s already encountered delays. In May, Gagnon guided to obtaining leases at all three priority projects by the end of 2026. But on the Q3 earnings call, he said that no deals will be signed this year, and declined to give dates for signing customers on those projects.
The rub: Keel’s now generating puny revenues since it’s now left Bitcoin mining and isn’t yet making money from AI. Gagnon’s sitting on a big cushion, having raised almost $600 million in convertible debt last year, and another almost $300 million by emptying its Bitcoin trove. Still, Keel’s burning cash fast. Will its trove prove big enough to fund crossing the span to AI on time? Investors are cautious about Keel’s prospects, to put it mildly. Keel’s stock dropped around 20% on Gagnon’s announcement of no deals this year, and currently has a paltry market cap of around $1.3 billion, about the same as when Riot bid back in 2024.
On the plus side, Gagnon’s collected a portfolio that potentially features some of the best AI sites in America and Canada. He notes that the big customers are now showing a strong preference for facilities near big population centers, just where he’s concentrating. “The difference is that until now, there’s been so much AI model training that wasn’t for customers, so it didn’t matter where the campus was located. The normal things that would have guided a developer to find a project and select a piece of land were thrown out the window in the interest of scale and time.”
Now, he says, the business is moving strongly towards “inference” where the compute providers sell to banks, research labs, call centers and other customers. “That’s where most of the money will be made, not on the retail side,” he says. “As the market moves to inference, more and more of the traditional data center qualities will come into play.” He notes that compute works much more efficiently when it’s near the customer than hundreds or thousands of miles away in a remote rural area. “It’s so important to get the fiber close to where people live and businesses are located, and in the frenzy, that’s been forgotten,” says Gagnon. The big distances cause sluggish responses that lead to poor user experiences in fields such as autonomous driving and interactive chatbots. Plus, he says, in the weeks leading up to and following the midterms, “we’ve seen a big uptick in level of demand and prioritization of our sites in Pennsylvania, driven that the support of data center had changed in many states. The Democratic candidates who won in Pennsylvania ran on platform favoring responsible data center development, the opposite the scenario in most other states.”
“The differentiator will be location, location and location,” says Gagnon. “This is a real estate industry and location is hugely important in real estate. That’s why you have so many cloud data centers in Loudoun County, Va. near D.C. and Quincy near Seattle. We’re near New York, Philly, and Pittsburgh, and have a great position in the Pacific Northwest, where there’s a big shortage of AI campuses. If you’re close to where the demand comes from, that should justify a premium going forward, just as it has in the past for data centers.”
Gagnon is especially bullish on the union of AI and the Keystone State. “Pennsylvania built this country, it had all the resources needed to launch the industrial revolution, steel, coal, railroads, natural gas. Texas was the first mover in AI data centers, and now Pennsylvania is the second mover. And we’ll be number one in Pennsylvania.”
It’s a long way from waste coal to GPUs. And Gagnon’s got a plausible story: The big guys have the GPUs, in fact too many of them. But he and his Bitcoin brethren have a lot of what the giants now need most, the power. The question for Keel is how fast Gagnon can turn it on.
This story was originally featured on Fortune.com