AI 수요가 비트코인 채굴 기업의 전력을 원하지만, 실제 수익을 가져가는 쪽은 계약 내용을 보면 드러난다.

Bitcoin miners spent years securing cheap electricity, building substations, and negotiating the grid connections needed to power their operations. Now AI companies want those connections so badly that they're willing to treat aging mines as valuable real estate.

The ASICs that mine Bitcoin are practically useless for training AI models, but the land, substations, and power agreements around them could spare data-center developers years of waiting.

SWI Group took control of Bitcoin miner Genesis Digital Assets earlier this year and announced an €800 million capital increase on Oct. 12 to expand its AI infrastructure business. Two weeks later, it launched SWI Compute, disclosing roughly $1 billion in NVIDIA equipment orders and its first customers, with some contracts still being finalized.

As part of the fundraising, Genesis is becoming part of a much bigger bet on selling computing power rather than mining BTC.

On paper, it's an irresistible trade: take an industrial property built to consume enormous amounts of electricity and supply that electricity to customers willing to pay considerably more for it.

Unfortunately, you can't turn a Bitcoin mining facility into an AI data center by plugging different computers into the wall. The conversion requires extensive retrofits, cooling that can handle much denser hardware, sophisticated networking, construction financing, and customers willing to commit to the finished product. Until the facility is operational, the promised revenue is just that: a promise.

Core Scientific's agreements with AI cloud provider CoreWeave show where the money goes. CoreWeave helps fund construction but recovers some of that contribution through rent credits, while Core Scientific borrows billions, builds the facilities, and risks penalties for late delivery.

That's how a mining property can become much more valuable without its owner receiving anything close to that value. Electricity draws the customers in, but the contracts determine who gets paid.

## Bitcoin mining equipment is losing value

Bitcoin mining is an unforgiving business. Operators run specialized computers known as ASICs, competing for block rewards whose value depends on Bitcoin's price and on how efficient their machines are relative to the rest of the network.

These machines require a lot of power regardless of market conditions, which means companies face the same bills in bull and bear markets. The first wave of public miners was seen as leveraged Bitcoin bets that institutions could easily make when access to BTC was hard or impossible.

But the market is a harsh mistress, and these companies now look nothing like leveraged Bitcoin bets. Increased competition and insatiable demand for AI have pushed many of them toward becoming data center landlords.

Core Scientific's second-quarter filing puts a price on the old business's declining appeal. During the first quarter, the company wrote down $266.5 million of mining equipment and infrastructure as Bitcoin prices, mining revenue per unit of computing power, and secondhand equipment values deteriorated. Of that charge, $151.6 million related to mining machines and $114.9 million to mining infrastructure. Meanwhile, its newer AI colocation business was already producing revenue. The company's mining operation has been losing money at the gross-profit level, while AI hosting earned a positive gross margin.

Mining hardware loses value as it ages, but a working grid connection doesn't. In fact, it can become even more desirable because developers can't simply go out and buy one. Utilities still need to assess transmission capacity, identify upgrades, and determine who pays for them. Having enough cash for GPUs doesn't make those steps go away.

Electricity access has become the miners' real prize, but the backlog of proposed projects also makes that access extremely difficult to value. In Texas, regulators approved ERCOT's Batch Zero review process in June, while the state ordered an audit of approximately 474 GW of proposed large-load connections.

Electricity already flowing to a facility differs from capacity approved for future use, and both differ from a place in the connection queue. Grouping them under a single headline megawatt figure can make a portfolio appear much closer to generating AI revenue than it is.

SWI's June announcement about Genesis described land with more than 1.3 GW of energized and approved connections across 15 facilities. By September, its half-year results described approximately 1.2 GW of secured grid connections. The disclosures don't reconcile the two figures or provide a site-by-site breakdown separating currently usable capacity from power that requires additional work.

SWI also didn't simply pay a fixed price for those megawatts. Its initial $500 million investment bought about 38.3% of Genesis's equity and rights to 77.2% of the value of certain preferred-share liquidation preferences, which govern claims on proceeds rather than ordinary ownership. SWI subsequently took control, but dividing that first payment by 1.3 GW would tell us little about what it paid for deliverable power.

Even an agreed valuation of the whole company wouldn't reveal what an operating AI data center costs. Genesis came with a mining business, equipment, land, existing electrical assets, and various rights to future power. However, those are just ingredients, and they require a lot of money before they can become a profitable AI business.

## Turning a Bitcoin mine into an AI data center is an expensive conversion

To the untrained eye, Bitcoin mining facilities and data centers might look similar, but they are two very different businesses. Bitcoin mines are built to run thousands of independent machines that can be switched off when necessary. AI data centers need sophisticated cooling, high-speed networking, redundant power systems, and much stricter uptime guarantees to keep densely packed GPU clusters operating reliably. Acquiring an existing facility is therefore just a head start.

Data centers consume electricity beyond what their servers use, particularly for cooling and power distribution. At a power usage effectiveness ratio of 1.25, a site drawing 100 MW from its electrical supply would deliver about 80 MW to computing equipment and use the other 20 MW to support the facility. Comparing that site's 100 MW grid connection with a competitor's 100 MW of usable IT capacity would make two very different properties seem equivalent.

Nvidia's GB300 NVL72 reference architecture describes liquid-cooled racks with 72 GPUs that each draw up to 142 kilowatts. Delivering the electricity is only one part of supporting them, since cooling, networking, and building design must also meet standards most Bitcoin mines weren't built for.

CoinShares estimated in its September mining report that converting mining infrastructure to AI-grade facilities costs roughly $8 million to $15 million per MW, compared with $700,000 to $1 million per MW for conventional mining infrastructure. That difference in order of magnitude explains why a powered mining site and a completed AI facility command such different valuations.

A review of more than $100 billion in AI deals found that only around 550 MW was billing out of more than 4 GW under contract. HIVE's $350 million AI agreement required an estimated $185 million hardware buildout before most of that revenue could arrive.

Under colocation arrangements, the operator generally provides the buildings, electricity infrastructure, cooling, and other services, while the customer supplies its own computing equipment. Selling AI computing capacity directly means buying GPUs as well, then finding enough customers to keep them occupied while newer, more efficient hardware makes existing systems less competitive.

Core Scientific earns fees for hosting other companies' machines, while SWI intends to sell computing itself. Both stand to benefit from former mining infrastructure, but their capital requirements and exposure to hardware costs differ greatly.

## Core Scientific's contracts show how the AI revenue gets divided

We can get a good sense of the math behind these deals from Core Scientific, because some of its converted facilities are already operating and its agreements with CoreWeave spell out where the money goes.

The company reported $136.7 million of colocation revenue in the second quarter, with 437 MW billing customers by mid-July. Its quarterly filing attributes $98.8 million to license fees and $35.1 million to power charges passed through to customers, with the balance coming from other services. The electricity reimbursement also appears as an equivalent expense, so it adds to reported revenue without creating gross profit.

More than a quarter of the colocation revenue therefore represents a utility bill being passed along, rather than money Core Scientific earns for providing the facility.