Ethereum co-founder Vitalik Buterin has compared Bitcoin mining with AI inference, sharing cost estimates that put electricity at 45.9% of mining expenses and computing capital expenditure at approximately 76% of inference costs.
Summary
- Electricity represents 45.9% of Bitcoin mining costs in the comparison Buterin shared.
- Computing capital expenditure accounts for approximately 76% of the AI inference estimate.
- Both tables include cooling, power equipment, and financing within computing capital expenditure.
- US operators have pursued AI facilities through customer leases, borrowing and crypto treasury sales.
In an Oct. 17 X post, Buterin described Bitcoin mining costs as split roughly evenly between computing investment and electricity, while calling AI inference almost completely dominated by computing capital expenditure.
Fun fact of the day: Bitcoin mining cost is split about 50/50 between compute capex and electricity. Meanwhile, AI inference is almost completely dominated by compute capex. pic.twitter.com/PCFKZSuNRw
— vitalik.eth (@VitalikButerin) October 17, 2026
Bitcoin electricity costs approach half of the estimated total
Buterin’s post draws a distinction between the spending required to purchase computing infrastructure and the recurring expense of powering it.
“Bitcoin mining cost is split about 50/50 between compute capex and electricity,” he wrote.
The mining table attached to his post provides a more detailed breakdown. Electricity supplied directly to computers costs an estimated $40,600 per BTC, accounting for 43.3% of the total. Other electricity expenses contribute another $2,400, or 2.6%.
Together, the two electricity categories amount to $43,000 per BTC and 45.9% of the listed cost.
Computing capital expenditure contributes approximately $37,600 per BTC, representing around 40%. The category covers hardware, cooling, and power equipment, including financing, so its scope extends beyond the purchase price of mining machines.
The remaining categories include labor at $3,200 per BTC, facility rent at approximately $2,500, and other expenses at roughly $7,400. Maintenance, spare parts, facility operations, insurance, taxes, and bandwidth fall into the final category.
Adding the listed dollar amounts produces an estimated total of $93,700 per BTC. The calculation describes the table Buterin shared and does not establish a uniform production cost for every Bitcoin miner.
AI inference spending centers on computing infrastructure
The AI table uses a different unit, dollars per GPU-year, and assigns a much larger share of expenses to computing capital expenditure.
Its estimate places that category at approximately $19,900 per GPU-year, or 76% of the total. As with the mining table, the category includes hardware, cooling, power equipment, and financing.
Electricity supplied directly to computers accounts for $1,050 per GPU-year, representing 4%. Other electricity contributes $210, or 0.8%, bringing the combined electricity share to 4.8%.
Labor adds $1,400 per GPU-year, while facility rent contributes approximately $900. The remaining maintenance, operating, insurance, tax, and bandwidth expenses total roughly $2,800.
The listed amounts sum to approximately $26,260 per GPU-year. Because one table measures costs per Bitcoin and the other per GPU-year, their dollar totals do not provide a direct comparison of the price of performing the same task.
Buterin summarized the difference by writing that AI inference is “almost completely dominated by compute capex.” The accompanying breakdown assigns approximately three-quarters of total inference costs to that category, with electricity, labor, rent and other operating expenses making up the remainder.
TeraWulf separates AI hosting from processor ownership
US data center agreements illustrate how companies divide responsibility for facilities and computing equipment.
As crypto.news previously reported, TeraWulf was considering a potential $3.5 billion financing in July to support an AI campus in Kentucky. The financing plans, attributed to Bloomberg, included possible leveraged loans and high-yield bonds.
The planned borrowing followed TeraWulf’s announced 20-year lease with Anthropic at its Justified Data campus in Hawesville. The company said the facility would support approximately 401 megawatts of critical computing capacity and estimated $19 billion in revenue over the initial contract term.
Initial capacity was expected in the second half of 2027, with full deployment targeted for early 2028. The revenue estimate covered two decades of the agreement and was not an upfront payment.
Chief Financial Officer Patrick Fleury described an operating model in which customers remain responsible for servers, processors and technology upgrades, while TeraWulf supplies power and physical infrastructure.
Earlier financing supported its New York operations. In October 2025, a TeraWulf subsidiary priced $3.2 billion of senior secured notes carrying a 7.75% annual interest rate and maturing in 2030.
Crypto sales and power agreements support AI development
Other companies have used digital assets to help finance their AI data center projects. A July 31 account documented transactions announced the previous day by Quantum Solutions and US-listed Hyperscale Data.
Quantum’s subsidiary GPT Pals Studio sold 1,000 ETH for $1.903 million after transaction fees to support Japanese AI data center expansion plans. Hyperscale monetized approximately 100 BTC and established a Bitcoin-backed credit facility for construction in Michigan.
Hyperscale’s initial customer deployment covered approximately 20 megawatts under a ten-year agreement with two optional five-year extensions. The company estimated more than $1.2 billion in revenue over the maximum term, subject to those extensions.
MARA’s July announcement focused on securing land and power capacity for future development in Texas. The company disclosed a definitive 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 was expected to provide access to an initial gigawatt of grid capacity by October 2027, with total available capacity projected to reach two gigawatts by April 2028.
MARA planned to develop the property with Starwood Digital Ventures, with phased construction subject to regulatory approvals. Seller HIF would retain a minority ownership interest once an HPC lease was executed.