Public Bitcoin miners have liquidated over 15,000 BTC from treasury reserves since late 2025, redirecting proceeds into artificial intelligence and high-performance computing infrastructure. The sell-off accelerated in Q1 2026 as hashprice fell to $0.034/TH/s — down 35% year-over-year — and minin...
"We are no longer a Bitcoin company. We are an infrastructure-first owner and developer for HPC/AI data centers across North America." — Ben Gagnon, CEO, Bitfarms (now Keel Infrastructure)
Public Bitcoin miners have liquidated over 15,000 BTC from treasury reserves since late 2025, redirecting proceeds into artificial intelligence and high-performance computing infrastructure. The sell-off accelerated in Q1 2026 as hashprice fell to $0.034/TH/s — down 35% year-over-year — and mining difficulty swung violently, plunging 7.76% on March 21 in the second-largest downward adjustment of 2026.
The exodus is not marginal. Core Scientific sold $175 million of BTC in January alone, reducing holdings from a peak of 9,618 to roughly 630 BTC. Bitdeer zeroed its entire 943 BTC treasury on February 20. Bitfarms rebranded as Keel Infrastructure and redomiciled to the United States. Cango, a former Chinese car dealer turned miner, posted a $452.8 million net loss and is pivoting to AI under the name EcoHash. Collectively, these companies signed AI/HPC contracts exceeding $65 billion in 2025, and by late 2026, mining may constitute less than 20% of revenue for companies that pivoted.
The structural question is no longer whether miners will diversify, but whether Bitcoin mining itself survives as a primary business line for publicly traded companies.
Bitcoin mining difficulty fell 7.76% to 133.79T at block height 941,472 on March 21, 2026. This marked the second-largest negative adjustment of the year, following a 14.72% surge on February 19 that pushed difficulty to 144.39T — the largest single increase since May 2021.
The whiplash tells a specific story: miners flooded in during late 2025 when hashprice briefly recovered, then exited rapidly when economics turned hostile. Network hashrate retreated from over 950 EH/s to approximately 943 EH/s. BTC traded at roughly $70,600 at the time of the adjustment, down approximately 50% from all-time highs.
Hashprice — the dollar revenue per petahash per second per day — collapsed from roughly $55/PH/s/day in Q3 2025 to approximately $35/PH/s/day by early December, and continued sliding into 2026. After pool fees (2.5%), firmware skims (1-2%), and realistic uptime (95-96%), net revenue lands around $42-$51/PH/s/day at the best of times. At the worst, new mining rigs now carry an ROI of approximately 1,000 days — meaning most new hardware will not pay back its cost before the next halving.
The BTC sell-off across public miners follows a clear pattern: dump bitcoin, raise debt, sign AI hosting contracts.
Core Scientific held 2,537 BTC ($222 million) at year-end 2025. By early March 2026, the company had sold roughly 1,900 BTC for approximately $175 million, reducing holdings to around 630 BTC. The company stated it expects to monetize "substantially all" of its remaining bitcoin during the year. The catalyst: a 12-year, $10 billion hosting agreement with CoreWeave covering 590 megawatts of capacity, with over 350 MW already activated. AI colocation revenue jumped 268% year-over-year to $31.3 million.
Bitdeer — a Singapore-based miner that topped global hashrate rankings at 63.2 EH/s — liquidated its entire reserve of 943.1 BTC plus 189.8 BTC of newly mined coins, bringing proprietary holdings to zero on February 20. The company simultaneously expanded a private placement of convertible senior notes to $325 million. Proceeds are being redirected to data center expansion and AI cloud infrastructure.
Bitfarms (now Keel Infrastructure) completed a year-long strategic review that concluded with a full rebranding, U.S. redomiciliation, and CEO Ben Gagnon's declaration that the company is "no longer a Bitcoin company." Stock jumped 16% on the announcement. The company is repositioning as an HPC/AI data center developer.
Cango (rebranding to EcoHash) reported $688.1 million in revenue but a net loss of $452.8 million in its first full year as a miner. The company sold approximately $305 million of bitcoin in February 2026 to repay debt and fund its AI inference services pivot. CEO Paul Yu described the move as "advancing our pivot to become an AI infrastructure provider."
Riot Platforms faces a different dynamic. Activist investor Starboard Value sent a public letter on February 18 arguing that Riot's 1.7 GW of available power across its Corsicana and Rockdale, Texas facilities could support an AI/HPC business worth $9 billion to $21 billion in equity value. Starboard pointed to Riot's January 2026 deal with AMD — 25 MW committed at $311 million in revenue over 10 years with an 80% EBITDA margin — as proof of concept. The firm warned that if Riot fails to execute, its rare power assets make it a prime acquisition target.
CleanSpark stated in Q1 2026 that Bitcoin mining investment "doesn't make a lot of sense" at current hashprices compared to AI infrastructure returns. The company is developing a 300 MW facility in Brazoria County, Texas dedicated to AI and HPC workloads.
TeraWulf has secured AI/HPC contracts valued at $6.7 billion, supported by Google-backed financial frameworks. The company has paired large-scale power with investment-grade counterparty backstops and non-dilutive financing, positioning it as an execution leader in the pivot.
The math is straightforward. Bitcoin mining generates fluctuating revenue tied to a cyclical commodity price, a difficulty algorithm that ensures perpetual competition, and a halving schedule that cuts block rewards every four years. AI hosting generates contracted, recurring revenue with long-term counterparties.
Consider the Riot-AMD deal: 25 MW at $311 million over 10 years implies roughly $31 million per year, with an 80% EBITDA margin yielding approximately $24.9 million in annual EBITDA from a relatively small power allocation. If Riot monetizes its remaining 1.4 GW at similar terms, Starboard estimates potential annual EBITDA of over $1.6 billion.
By contrast, Bitcoin mining gross margins for U.S.-listed miners sat at roughly 47% as of January 2026, according to JPMorgan — up 300 basis points from December but still under pressure. For miners operating above $0.06-$0.07/kWh with hardware less efficient than 15-16 J/TH, margins turn negative.
The contract structure difference is critical. AI hosting customers sign multi-year, fixed-rate agreements. Bitcoin mining offers no such guarantee. A miner's revenue can swing 20% in a week based on difficulty adjustments and BTC price moves.
The profitability crisis has cascaded into the ASIC hardware market. Bitmain slashed prices on S19 and S21 series machines to as low as $3-$4 per terahash — historic lows. Even next-generation hardware, such as the Whatsminer M6DS+ released in March 2026 (504 TH/s, 17.00 W/TH efficiency), faces the 1,000-day ROI problem at current hashprices.
The discounting reflects a buyer's market where marginal miners are exiting and new entrants face poor unit economics. Hardware that previously cost $15-$20/TH/s is now available at 75-80% discounts. For well-capitalized miners with sub-$0.04/kWh power, this represents a buying opportunity. For everyone else, it confirms the industry's contraction.
Not every miner is running for the exit.
American Bitcoin (ABTC), backed by the Trump family, moved in the opposite direction, purchasing 11,298 ASIC miners in early March to expand its fleet by 12%. The company is betting on a recovery in BTC price and mining margins, though it has the political connections and brand to pursue a different strategy than pure-play miners.
Marathon Digital Holdings (MARA), the largest public miner by hashrate, has maintained its HODL strategy, keeping one of the largest corporate BTC treasuries. The company has selectively added AI/HPC exposure but has not announced plans to liquidate bitcoin reserves.
The miners that stay tend to share common characteristics: sub-$0.04/kWh power costs, latest-generation hardware (sub-16 J/TH), vertically integrated operations, and a conviction that BTC price will recover sufficiently to justify the hashrate investment.
The miner exodus raises a structural question about Bitcoin's security model. The network's hashrate remains near all-time highs at approximately 903-943 EH/s, supported in part by new entrants and private miners filling capacity vacated by public companies. The difficulty algorithm's self-correcting mechanism ensures blocks continue at 10-minute intervals regardless of how many miners exit.
However, the concentration of mining among fewer, larger operators narrows the set of entities securing the network. If public miners continue redirecting power capacity from BTC to AI — and contracts worth $65 billion suggest they will — the network's security becomes increasingly dependent on private mining operations and smaller public companies that lack the infrastructure to pivot.
The March 21 difficulty drop gave remaining miners immediate breathing room: each unit of computing power now solves blocks more readily, temporarily lifting revenue per hash. This self-balancing feature is Bitcoin's most robust defense against miner attrition, but it functions by making the network easier to attack when hashrate declines — a tradeoff that becomes more consequential as public miners step away.
The Bitcoin mining industry is undergoing its most significant structural transformation since the 2021 China ban. Unlike that event, which relocated mining from one geography to another, the current shift removes capacity from Bitcoin entirely and redirects it toward a competing demand center: artificial intelligence compute.
The economic logic is difficult to argue with. AI hosting contracts offer 80% EBITDA margins, multi-year revenue visibility, and investment-grade counterparties. Bitcoin mining offers none of these. At current hashprices, the only miners with a defensible business are those with the cheapest power and newest hardware — a category that shrinks with each difficulty adjustment.
For the Bitcoin network, the implications are manageable in the short term. The difficulty algorithm will continue to absorb miner exits, and private operators appear willing to fill vacated capacity. The longer-term concern is whether a network increasingly reliant on private, opaque mining operations can maintain the decentralized security model that underpins its value proposition.
The question facing every public mining company board is now binary: is bitcoin a product or a commodity input for a data center business? The treasury liquidation data suggests most have already answered.