Ethereum co-founder Vitalik Buterin turned to X on October 17, 2026, with a side-by-side comparison of electricity economics, contrasting how money is spent in Bitcoin mining versus AI inference. His core observation is straightforward: Bitcoin mining costs are divided between hardware and electricity, while AI inference spending is concentrated almost entirely in hardware.

The Numbers Behind the Split

Estimates cited in the comparison place the cost of mining one Bitcoin at roughly $93,700. Of that total, approximately $37,600, or 40%, goes to compute capital expenditure — the upfront cost of acquiring mining machines. Electricity accounts for about $40,600, or 43.3%. Put another way, the bill for the rigs and the bill for the power that runs them are nearly even.

AI inference presents a markedly different picture. Inference is the stage at which a trained model answers questions, generates images, or writes code for users. The estimated cost there runs about $26,260 per GPU-year, with compute capex making up roughly $19,900, or 76% of the total. Electricity falls between $1,050 and $1,260, or just 4% to 4.8%. For an AI operator, the power bill is closer to a rounding error than a strategic concern.

The gap widens further at the facility level. Building and operating a Bitcoin mining site is estimated at $700,000 to $1 million per megawatt, while AI infrastructure comes in at approximately $20 million per megawatt.

Why Cost Structure Matters More Than the Total

Buterin's analysis centers on how costs are composed, not on which sector spends more or consumes more energy overall. It does not attempt to equate absolute spending or total energy use between the two industries.

That composition also determines which levers each industry can pull: miners can pursue cheaper power, while AI operators' main efficiencies come from extracting more output from hardware they have already paid for.

Bitcoin miners remain highly sensitive to ongoing power costs, since electricity represents such a large share of what it takes to produce each coin. That sensitivity has long shaped where mining operations set up, with sites gravitating toward low-cost energy sources. AI operators' economics instead hinge on hardware efficiency and utilization — how much useful work each expensive GPU performs over its life.

Background: Two Industries Chasing the Same Power

Bitcoin mining and AI data centers increasingly compete for the same resources: grid connections, cheap power, and physical space.

Mining is built around a proof-of-work system in which machines race to solve puzzles, and the winner earns newly issued Bitcoin. Electricity is a structural input to that process, which helps explain why it accounts for such a large slice of the cost estimate.

Ethereum, the network Buterin co-founded, moved away from proof-of-work years ago, switching to proof-of-stake in September 2022. When he weighs in on Bitcoin's energy economics, he is speaking as someone whose own chain no longer pays that particular bill.

AI inference, by contrast, depends on specialized chips. Those chips are costly, and the data centers that house them are far more expensive to build per megawatt, as the facility-level estimates show.

Implications for Miners and Operators

With electricity representing more than 40% of the cost to mine a coin, miners with access to cheap, stable energy hold a structural edge, while those to volatile power markets carry real margin risk.

Some miners could explore repurposing their infrastructure for AI applications, where capital efficiency may be greater. The estimates, however, underscore the scale of any such shift: mining sites cost an estimated $700,000 to $1 million per megawatt, while AI infrastructure runs around $20 million per megawatt. A miner moving toward AI would take on a radically different capital commitment, swapping electricity risk for hardware and utilization risk.

On the AI side, with electricity at roughly 4% to 4.8% of inference costs, the pressure falls on keeping GPUs busy and efficient, since idle hardware carrying $19,900 a year in capex is an expensive asset to leave sitting unused. For anyone tracking both sectors, power prices and GPU utilization rates are the figures that follow most directly from this cost split.