Vitalik Buterin’s cost tables put electricity at 45.9% of listed Bitcoin mining costs, while computing infrastructure makes up about 76% of estimated AI inference costs. The contrast is useful, but the tables measure different things, and neither figure is a universal benchmark.
- Bitcoin estimate: $43, 000 in electricity costs per BTC.
- AI inference estimate: about $19, 900 in computing-related costs per GPU-year.
- The units and key assumptions are not fully specified.
- Crypto-linked companies are taking different approaches to AI infrastructure and financing.
What Buterin’s Tables Estimate
In an Oct. 17 post on X, Ethereum co-founder Vitalik Buterin said Bitcoin mining costs are split “about 50/50 between compute capex and electricity.” The year of the post and the methodology behind the tables are not specified in the available information.
The Bitcoin table puts total costs at $93, 700 per BTC. Electricity accounts for $43, 000, or 45.9%. That includes $40, 600 per BTC for power supplied directly to computers and $2, 400 for other electricity expenses.
The table assigns approximately $37, 600 per BTC, around 40%, to computing capital expenditure. The label needs some context: this category includes hardware, cooling, power equipment and financing. It is not just the purchase price of mining machines, and “capital expenditure” covers more here than it usually does in accounting.
The other listed Bitcoin costs are $3, 200 for labor, approximately $2, 500 for facility rent and roughly $7, 400 for other expenses. Those include maintenance, spare parts, facility operations, insurance, taxes and bandwidth.
Buterin’s AI estimate covers inference, meaning the running of AI computations rather than model training. It lists approximately $19, 900 per GPU-year for computing capital expenditure, about 76% of the $26, 260 per GPU-year total. As in the mining table, this category includes hardware, cooling, power equipment and financing.
The AI estimate assigns $1, 260 per GPU-year to electricity: $1, 050 for power supplied directly to computers and $210 for other electricity expenses. Together, those costs make up 4.8% of the total. The table also lists $1, 400 for labor, about $900 for facility rent and roughly $2, 800 for maintenance, operating expenses, insurance, taxes and bandwidth.
Why the Figures Are Not a Direct Comparison
The tables point to different cost profiles. Electricity makes up a substantial share of listed mining costs, while computing-related costs dominate the AI inference estimate. But $93, 700 per BTC and $26, 260 per GPU-year are not measures of equivalent work. One is calculated per bitcoin produced, the other per GPU over a year.
The Bitcoin figure also does not establish a standard cost to mine one BTC. Miners face different electricity prices, equipment and financing costs, as well as different operating conditions. The AI table leaves key assumptions unclear, too, including GPU models and utilization. Without those details, the figures should be treated as estimates, not reproducible benchmarks for every miner or AI operator.
The category labels matter as well. Both tables group financing with hardware and facility equipment under computing capital expenditure. So the percentages reflect the tables’ own cost categories, rather than a straightforward accounting comparison between electricity bills and hardware purchases.
Different Routes to Building AI Infrastructure
Several companies with crypto ties have announced plans involving AI data centers, computing capacity or related financing. Their plans show a range of approaches. They do not establish an industry standard or prove that Buterin’s cost estimates influenced the companies’ decisions.
TeraWulf said its Justified Data campus in Hawesville, Kentucky, would serve Anthropic under a 20-year lease arrangement. The planned facility is expected to provide approximately 401 megawatts of critical computing capacity. TeraWulf estimated $19 billion in revenue over the initial contract term. That projection spans two decades and is not an upfront payment. Initial capacity was expected in the second half of 2027, with full deployment targeted for early 2028.
TeraWulf CFO Patrick Fleury described a setup in which customers provide servers and processors and handle technology upgrades, while TeraWulf supplies power and physical infrastructure. The available details do not spell out all the lease’s financial terms or assign every project risk.
Bloomberg reported in July that TeraWulf was considering potential $3.5 billion financing for a Kentucky AI campus, including leveraged loans and high-yield bonds. Separately, in October 2025, a TeraWulf subsidiary priced $3.2 billion of senior secured notes with a 7.75% annual interest rate and a 2030 maturity date. The available information does not explain how the notes relate to the potential financing or detail their collateral and covenants.
Other announced plans have involved digital assets. A July 31 report described transactions announced the previous day: Quantum Solutions subsidiary GPT Pals Studio sold 1, 000 ETH for $1.903 million after transaction fees to support plans to expand Japanese AI data centers. Hyperscale Data monetized approximately 100 BTC and established a Bitcoin-backed credit facility to fund 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, which depends on those extensions. The conditions for exercising them and the credit facility’s detailed terms have not been specified.
MARA announced plans to acquire a 1, 200-acre Texas property for high-performance computing, flexible computing services and Bitcoin mining. It projected access to an initial gigawatt of grid capacity by October 2027, with total available capacity reaching two gigawatts by April 2028.
MARA planned to develop the site in phases with Starwood Digital Ventures, subject to regulatory approvals. The seller, HIF, would retain a minority ownership interest once an HPC lease was executed. The acquisition, development and grid-capacity figures are plans and projections, not completed milestones.
Key Questions and Answers
-
What share of Bitcoin mining costs is electricity in Buterin’s estimate?
The table assigns 45.9% of listed costs to electricity: 43.3% for power supplied directly to computers and 2.6% for other electricity expenses.
-
What dominates the AI inference estimate?
Computing capital expenditure, at approximately $19, 900 per GPU-year, or about 76% of the listed total.
-
Can the Bitcoin and AI totals be compared directly?
No. The Bitcoin estimate is per BTC, while the AI estimate is per GPU-year. They do not measure the same output or equivalent work.
-
How does TeraWulf’s described operating model divide responsibilities?
According to CFO Patrick Fleury, customers provide servers and processors and handle technology upgrades, while TeraWulf supplies power and physical infrastructure.
-
How are the companies supporting AI infrastructure plans?
The examples include customer leases, potential borrowing, ETH and BTC sales, a Bitcoin-backed credit facility, and plans to secure land and grid capacity. These are different approaches, not one industry-wide financing model.
What the Estimates Leave Open
Buterin’s tables show that computation costs involve more than electricity. In these estimates, power is a major part of Bitcoin mining costs, while computing-related costs dominate AI inference. The missing assumptions and different units limit what the comparison can tell us.
The company plans raise a separate practical question: can financing, power access, construction and customer commitments line up on schedule? Projected capacity and revenue are not delivered capacity or guaranteed returns. Execution, not a cost table or an ambitious press release, will determine what gets built.
