Ethereum co-founder Vitalik Buterin recently drew parallels between Bitcoin mining and AI inference operations. According to his shared financial estimates, electricity accounts for 45.9% of total mining overhead, whereas hardware capital expenditure is estimated at roughly 76% of overall AI inference costs.
In an Oct. 17 post on X, Buterin outlined Bitcoin mining costs as a near-even split between hardware investment and power consumption, though AI inference was described as an operation almost entirely driven by computing capital expenditure.
Bitcoin Electricity Costs Near Half of the Estimated Total
A distinction between upfront computing infrastructure investments and ongoing electricity operational expenses is highlighted in Buterin’s analysis.
“Bitcoin mining cost is split about 50/50 between compute capex and electricity,”
he wrote.
The attached breakdown table offers deeper financial details. Power supplied directly to rigs runs about $40,600 per BTC, taking up 43.3% of expenses, while secondary energy costs are calculated at an additional $2,400, or 2.6%.
Combined, both power categories reach $43,000 per BTC, accounting for 45.9% of the overall expense tally recorded in the post.
Computing capital expenditure accounts for roughly $37,600 per BTC, forming about 40% of the total. Hardware, cooling, and power equipment are included within this category alongside financing, extending its scope far beyond the initial purchase price of mining rigs.
The remaining figures list labor costs at $3,200 per BTC, location leases around $2,500, and miscellaneous expenses at roughly $7,400. Maintenance, spare parts, site management, insurance, taxes, and internet bandwidth are categorized under this final expenditure group.
Summing the itemized figures yields a projected cost of $93,700 per BTC. This tally reflects the specific chart Buterin shared and is not intended to establish a universal production benchmark across all Bitcoin miners.
AI Inference Spending Focuses on Computing Infrastructure
The AI chart relies on a distinct metric—dollars per GPU-year—and is structured to assign a substantially heavier proportion of expenses to hardware capital outlays.
The projection pegs that category at roughly $19,900 per GPU-year, making up 76% of overall expenses. Similar to the mining breakdown, hardware, cooling infrastructure, energy equipment, and financing costs are grouped into this single classification.
Direct computing power accounts for $1,050 per GPU-year, making up 4% of the total. Secondary energy usage adds $210, or 0.8%, which is combined to bring the overall electricity share to 4.8%.
Labor adds $1,400 per GPU-year, while facility leases contribute roughly $900. The remaining maintenance, operational, insurance, tax, and network bandwidth expenses are calculated at approximately $2,800.
The itemized figures total roughly $26,260 per GPU-year. Since one framework evaluates expenses per Bitcoin and the other per GPU-year, their overall dollar amounts are not designed to offer a direct head-to-head comparison for executing equivalent workloads.
Buterin highlighted the contrast by noting that AI inference focuses almost entirely on hardware capex. His accompanying figures assign roughly three-fourths of overall inference expenditure to that primary category, while electricity, labor, rent, and miscellaneous operating costs are grouped into the remaining portion.
TeraWulf Splits AI Hosting Operations From Processor Ownership
American data center contracts demonstrate how tech firms divide operational duties between facility infrastructure and computing hardware, a balance that is highlighted across recent industry agreements.
As previously reported, TeraWulf weighed a potential $3.5 billion funding round in July to back a Kentucky AI campus. According to Bloomberg coverage, the proposed financing structure was expected to feature leveraged loans alongside high-yield bonds.
The proposed debt raising followed TeraWulf’s announced 20-year lease agreement with Anthropic at its Hawesville Justified Data facility. Company executives noted the site would supply roughly 401 megawatts of critical computing load and was projected to yield $19 billion in revenue across the initial contract duration.
Initial operations were slated for late 2027, aiming for full operational status by early 2028. The projected revenue figure spanned two full decades of the agreement, meaning the money was not received as a single upfront cash lump sum.
Chief Financial Officer Patrick Fleury outlined an operational framework where clients handle servers, processors, and system upgrades themselves, while physical infrastructure and power delivery are provided by TeraWulf.
Prior funding secured earlier supported its expansion across New York state. In October 2025, a TeraWulf subsidiary issued $3.2 billion in senior secured notes bearing a 7.75% annual coupon, which are scheduled to mature in 2030.
Crypto Sales and Power Deals Help Fund AI Development
Several other firms leveraged digital assets to fund their AI data center expansion plans. A July 31 report detailed deal structures announced the preceding day by Quantum Solutions and US-traded Hyperscale Data, which were designed to raise fresh development capital.
GPT Pals Studio, a subsidiary of Quantum, sold 1,000 ETH for $1.903 million after fees to fund AI data center projects in Japan. Meanwhile, Hyperscale liquidated roughly 100 BTC and set up a crypto-backed credit line, through which funds were secured for ongoing facility construction in Michigan.
Hyperscale’s initial client launch spanned roughly 20 megawatts under a ten-year deal featuring two optional five-year renewal periods. Company leaders estimated total revenue exceeding $1.2 billion across the full potential timeframe, provided that those extensions are exercised by the client.
MARA’s July announcement focused on securing acreage and utility connections for future expansion across Texas. The crypto firm finalized a binding contract to purchase a 1,200-acre site, where high-performance computing, flexible tech services, and Bitcoin mining operations are planned to coexist.
Based on MARA’s public filing, the new site will grant access to an initial gigawatt of power grid capacity by October 2027, while total available output is projected to hit two gigawatts by April 2028.
MARA mapped out development plans for the property alongside Starwood Digital Ventures, staging construction in phases dependent on regulatory approvals. Seller HIF maintains a minority stake, which will be retained once an HPC lease agreement reaches final execution.
