Ethereum co-founder Vitalik Buterin compared Bitcoin mining to AI inference and shared cost estimates: electricity accounts for 45.9%, while computational capital expenditure makes up approximately 76% of inference costs.
- In Buterin's comparison, electricity accounts for 45.9% of Bitcoin mining costs.
- Capital expenditure accounts for approximately 76% of the estimated cost of AI inference.
- Both tables include refrigeration, power supply equipment, and financing in the calculation of capital expenditures.
- U.S. operators are advancing AI infrastructure through customer leases, loans, and the sale of crypto assets.
In an October 17 X post, Buterin stated that Bitcoin mining costs are roughly split equally between computational investment and electricity, whereas AI inference is almost entirely dominated by computational capital expenditure.
The electricity cost of Bitcoin approaches nearly half of the total cost.
Buterin's post distinguishes between the upfront costs of purchasing computing infrastructure and the ongoing expenses required to power it.
He wrote: "The cost of Bitcoin mining is approximately split 50% between capital expenditure on computing hardware and electricity."
The mining table attached to his post provides a more detailed breakdown. The estimated electricity cost directly consumed by the computer is $40,600 per BTC, accounting for 43.3% of the total cost. Additional electricity expenses add another $2,400, or 2.6%.
The combined cost of the two types of electricity amounts to $43,000 per BTC, accounting for 45.9% of the stated costs.
The capital expenditure is calculated at approximately $37,600 per BTC, accounting for about 40%. This category includes hardware, cooling, and power supply equipment, including financing, and thus extends beyond just the purchase price of mining rigs.
Other categories include $3,200 per BTC in labor costs, approximately $2,500 in facility rent, and around $7,400 in other expenses. Maintenance, spare parts, facility operations, insurance, taxes, and bandwidth are all included in this final category.
After summing the listed amounts, the estimated total cost is approximately $93,700 per BTC. This calculation is based on the table shared by Buterin and does not imply that all Bitcoin miners have the same production cost.
AI inference spending is concentrated on computational infrastructure.
The AI table uses different units—dollars per GPU year—and allocates a larger proportion of costs to computational capital expenditures.
This estimate assigns this category a cost of approximately $19,900 per GPU per year, accounting for 76% of total costs. Similar to the mining table, this category includes hardware, cooling, power supply equipment, and financing.
Direct power supply to computers amounts to $1,050 per GPU annually, accounting for 4%. Other power expenses amount to $210, or 0.8%, bringing the total power cost percentage to 4.8%.
Labor costs increased by $1,400 per GPU per year, and facility rent is approximately $900. The remaining expenses for maintenance, operations, insurance, taxes, and bandwidth total approximately $2,800.
The total amount listed is approximately $26,260 per GPU per year. Since one table is billed per bitcoin and the other per GPU per year, the total dollar amounts cannot be directly compared to determine the cost of performing the same task.
Buterin summarized this difference by writing that AI inference is "almost entirely dominated by computational capital expenditures." The accompanying breakdown shows that approximately three-quarters of the total inference cost falls into this category, with electricity, labor, rent, and other operational expenses making up the remainder.
TeraWulf separates AI hosting from processor ownership.
The U.S. data center agreement outlines how the company allocates responsibility for facilities and computing equipment.
According to previous reports from crypto.news, TeraWulf considered a potential $3.5 billion financing in July to support its AI campus in Kentucky. Bloomberg reported that these financing plans included possible leveraged loans and high-yield bonds.
Prior to the financing plan, TeraWulf announced a 20-year lease agreement with Anthropic at its Justified Data campus in Hoswell. The company stated that the facility will support approximately 401 megawatts of critical computing capacity and is expected to generate $19 billion in revenue over the initial term of the contract.
Initial capacity is expected to come online in the second half of 2027, with full deployment targeted for early 2028. Revenue estimates cover the 20-year term of this agreement, not upfront payments.
Chief Financial Officer Patrick Fleury stated that the operating model has customers responsible for servers, processors, and technology upgrades, while TeraWulf provides power and physical infrastructure.
Previous financing has supported its New York operations. In October 2025, a subsidiary of TeraWulf issued $3.2 billion in senior secured notes with a coupon rate of 7.75%, maturing in 2030.
Sale of crypto assets and power supply agreements support AI development
Other companies are also leveraging digital assets to finance their AI data center projects. A report on July 31 documented the transaction announced the previous day by Quantum Solutions and Hyperscale Data, a U.S.-listed company.
Quantum’s subsidiary, GPT Pals Studio, sold 1,000 ETH, receiving $19.03 million after transaction fees, to support the expansion of its AI data center in Japan. Hyperscale liquidated approximately 100 BTC and established a bitcoin-collateralized credit facility for construction in Michigan.
Hyperscale's initial customer deployment covers approximately 20 megawatts over a 10-year term, with optional extensions for two additional five-year periods. The company estimates that revenue could exceed $1.2 billion over the maximum term, assuming these extension options are exercised.
MARA’s July announcement focused on securing land and power capacity for future development in Texas. The company disclosed a final agreement to acquire a 1,200-acre property intended to support high-performance computing, flexible computing services, and Bitcoin mining.
According to MARA's announcement, the site is expected to have an initial 1 GW grid capacity available by October 2027, with total available capacity projected to reach 2 GW by April 2028.
MARA plans to collaborate with Starwood Digital Ventures to develop the property in phases, subject to regulatory approvals. The seller, HIF, will retain a minority stake after signing the HPC lease.