Analysis

From Bitcoin Mining to AI Data Centers: Does the Big Pivot Actually Pay Off?

co-written by eustella AI16. July 2026, 08:00

Bitcoin is stuck in a bear market – yet the shares of many Bitcoin miners keep hitting new highs. The reason: these companies are increasingly leasing their power and data center capacity to the AI industry. Multi-billion-dollar deals with Microsoft, AWS, Anthropic, and CoreWeave have repriced the entire sector. But does the transformation actually pay off for the companies? A look at the numbers shows the answer is considerably more nuanced than the stock rally suggests.

Why Miners Are Switching in the First Place

The economics of Bitcoin mining are currently broken for many operators. According to the Q1 2026 report by research firm CoinShares, producing one Bitcoin cost publicly listed miners an average of roughly $80,000 in the fourth quarter of 2025. Per mined coin, companies were losing around $19,000 according to a CoinDesk estimate from March – back when Bitcoin was still trading between $68,000 and $70,000. The price has since fallen to around $64,000, widening the gap between production costs and market price even further. The hashprice – the revenue miners earn per unit of computing power – fell from $36–38 per petahash per day in Q4 2025 to roughly $29 in Q1 2026, according to CoinShares, a five-year low. At these levels, the analysts estimate that 15 to 20 percent of the global mining fleet is operating at a loss: any machine less efficient than an Antminer S19 XP paying more than 6 US cents per kilowatt-hour for power is unprofitable at these prices. CoinShares therefore expects further capitulation among high-cost operators unless the Bitcoin price recovers meaningfully.

Structural problems compound the squeeze: the April 2024 halving cut the block reward to 3.125 BTC, and the next halving arrives in 2028. Transaction fees, once around seven percent of miner revenue, have shrunk to roughly one percent – partly because institutional products like ETFs are “locking up” large amounts of Bitcoin for the long term and on-chain activity is declining.

At the same time, demand is exploding for exactly what miners have in abundance: power. Hyperscalers like Microsoft, Google, and Amazon are desperately searching for data center capacity for AI workloads. The bottleneck isn’t hardware but grid connection: in key US markets, it can take more than four years for new large-scale consumers to be connected to the power grid. Miners solved that problem long ago – they sit on secured power capacity, land, and cooling infrastructure. Industry estimates suggest they can bring AI-ready facilities online up to 75 percent faster than greenfield builds. Bernstein analysts have accordingly taken to calling miners the “power landlords” of the AI era.

The Scale: $135 Billion in Two Years

A recent Bernstein analysis shows just how large the shift has become: over the past two years, Bitcoin miners have contracted out roughly 7 gigawatts of power capacity to hyperscalers, specialized AI clouds (“neoclouds”), and chipmakers across 19 deals – with a total value of more than $135 billion. And that’s just the beginning: the contracted 7 gigawatts represent less than a quarter of the sector’s planned power pipeline of around 30 gigawatts.

Among the most prominent deals:

In early July, TeraWulf signed a 20-year lease with AI company Anthropic for its campus in Hawesville, Kentucky – roughly 401 megawatts of capacity, with expected contract revenues of about $19 billion. That’s more than the entire company is currently worth on the stock market (around $12 billion). IREN secured a five-year, $9.7 billion contract with Microsoft; Hut 8 signed a 15-year, $9.8 billion lease for a 352-megawatt site in Texas; Cipher landed a 15-year AWS contract worth $5.5 billion. Core Scientific, the pioneer of the model, has contracts worth more than $10 billion running with AI cloud provider CoreWeave. Most recently, CleanSpark announced its first AI colocation deal: a 20-year, $6.6 billion lease in Georgia.

The market is rewarding all of this: Bitcoin has lost roughly a quarter of its value since the start of the year – from about $87,500 in early January to around $64,000 most recently. A basket of mining stocks, by contrast, had gained more than 50 percent by early June according to 10X Research, with TeraWulf up over 70 percent – and that was at a point when Bitcoin’s year-to-date loss stood at only 17 percent. CoinShares expects that listed miners with AI contracts could derive up to 70 percent of their revenue from the AI business by year-end, up from roughly 30 percent at the start of the year. At TeraWulf, HPC leasing already overtook mining revenue in the first quarter of 2026. Just how heavily the market is rewarding the AI story shows in another CoinShares figure: miners with secured HPC contracts are valued at 12.3 times their expected next-twelve-month revenues, while pure-play Bitcoin miners trade at just 5.9 times – the market is paying more than double for the AI exposure.

Does It Pay Off? The Sobering Return Math

Behind the headline billions, however, lies a considerably more sober reality – as the fresh Bernstein analysis from July 15 shows. The analysts ran the numbers on the deals’ returns and reached a clear conclusion: the most spectacular case is not the rule.

Core Scientific achieves a five-year average return on assets of 75 percent on its CoreWeave deal. But the reason lies in an exceptional financing structure: the tenant, CoreWeave, funds $750 million of the $855 million in investment costs itself through revenue prepayments – Core Scientific only has to contribute $105 million from its own balance sheet, effectively paying just $1.5 million per IT megawatt. Such constructions, the analysts write, are limited in availability and do not reflect the sector’s underlying economics.

The realistic baseline looks different: TeraWulf comes in at a stabilized return on assets of 5 percent, with Cipher and CleanSpark at 4 percent each. The reason: anyone building on their own dime pays between $8 million and $11 million per IT megawatt – AI infrastructure is many times more expensive than Bitcoin mining infrastructure, which CoinShares puts at $700,000 to $1 million per megawatt. The unlevered internal rates of return on the colocation deals sit at 8 to 13 percent according to Bernstein, against financing costs of 6 to 7 percent. That’s a solid but hardly spectacular business – more real estate than crypto gold rush.

What it offers instead is predictability: in place of volatile block rewards, there are long-running, dollar-denominated contracts with creditworthy tenants, some spanning 15 or 20 years. Under so-called triple-net leases, the tenant additionally covers power costs, taxes, and operating expenses – which is how Cipher achieves a 94 percent EBITDA margin on its AWS deal. It is precisely this predictability that the stock market currently values more highly than any hashrate.

The Price of the Transformation: Debt and Sold Bitcoin

The pivot, however, is extremely capital-intensive – and it is being financed in two ways, both of which carry risks.

First, through debt: IREN now carries $3.7 billion in convertible bonds, TeraWulf $5.7 billion in total debt, with another roughly $3.5 billion in borrowing planned for the Anthropic campus, arranged by Morgan Stanley. At Cipher, quarterly interest expense exploded from $3.2 million to $33.4 million after a bond issuance. Research firm VanEck puts the sector’s near-term funding gap for the AI buildout at around $50 billion.

Second, through the sale of their own Bitcoin holdings: in the first quarter of 2026 alone, publicly listed miners sold more than 32,000 BTC from their treasuries – more than in all quarters of 2025 combined. Core Scientific announced it would liquidate essentially all remaining holdings, while Bitdeer reduced its treasury to zero in February. Even Marathon (MARA), the largest listed holder with 53,822 BTC, abandoned its strict “HODL” strategy: in March, the company authorized sales from its entire reserve via its 10-K filing, partly because the loan-to-value ratio on its $350 million Bitcoin-backed credit facility had climbed to roughly 87 percent as prices fell.

This has side effects for Bitcoin itself: the hashrate recorded its first first-quarter decline in six years, and North American mining pools’ share of Bitcoin blocks fell from 40 to 35 percent in 2025. In mid-June, mining difficulty dropped 10 percent – already the second decline of that magnitude in 2026, a sign that unprofitable operators are actually switching off their machines. JPMorgan, citing the CoinShares data, points out that Bitcoin has now traded below its estimated production cost of around $78,000 for five consecutive months in 2026. That’s not an immediate security threat to the network – but it is a structural shift, because sites once converted for AI are unlikely to return to mining even in the next bull market.

Critics also see governance problems: a report by consulting firm Blocksbridge documented in early July extensive share sales by executives and board members of several mining companies following the stock rally. At IREN, the award of more than 18 million free shares to the two co-CEOs drew criticism over the dilution of existing shareholders.

Not Everyone Is Playing Along: The Cango Counter-Model

Interestingly, not every miner is following the herd. NYSE-listed Cango – originally a Chinese auto-financing company, a Bitcoin miner since late 2024 – is deliberately staying out of the AI training business with hyperscalers. Head of communications Juliet Ye reasons that this sector is already crowded with tech giants; her guiding principle: “What not to do is as important as what to do.”

Instead, Cango is betting on distributed AI inference through its subsidiary EcoHash, launched in April – running finished AI models rather than training them. The logic: Cango’s more than 30 sites worldwide, at 10 to 50 megawatts each, are too small for hyperscalers hunting for 100-megawatt campuses, but well suited for low-latency inference close to customers. Add to that a remarkable market figure: more than 70 percent of the mining sector’s power capacity sits with small, independent operators – only 30 percent with the publicly listed companies. Cango wants to bring these small sites on board as partners, supplying the technology, the customers, and the financing. Early customers include GPU marketplaces like Runpod and Vast.ai as well as AI startups for whom hyperscaler pricing is out of reach.

The transformation was expensive here too: Cango sold 6,451 Bitcoin for around $442 million and cut its long-term debt by 94.5 percent to $30.6 million within a single quarter – but was left with only $7.2 million in cash afterward. Mining continues as the cash machine: 31.7 exahash generated $98.4 million in mining revenue in the first quarter.

Verdict: A Lifeline With Question Marks

So does the pivot pay off? The honest answer: for many miners, it is less a growth story than a survival strategy – and whether it pays off depends heavily on the starting position.

For companies with large, grid-connected sites and deep-pocketed tenants willing to shoulder part of the construction costs, the business is attractive: predictable, long-running revenues that beat any mining margin in the current Bitcoin bear market. For the broader field, however, the sober Bernstein math applies: single-digit returns on capital, billions in upfront investment, a sector-wide funding gap of around $50 billion – all of it built on GPU hardware that ages quickly and with concentration risks tied to individual anchor tenants.

The mining stock rally is currently pricing in, above all, the expectation that the contracted billions will actually turn into built, leased, and profitable data centers. That is precisely the open question for the second half of 2026: the market has already repriced the miners as AI infrastructure companies – now they have to prove it. Should the Bitcoin price recover toward $100,000, some conversions may look poorly timed in hindsight. If it stays in the basement, the pivot was, for many, the only option.