Vitalik Spills the Tea: Bitcoin Miners Spend More on Power, AI Bros Bleed Cash on GPUs ⚡

Ethereum co-founder Vitalik Buterin took to X with a "fun fact of the day," posting two cost tables that lay bare the structural difference between Bitcoin mining economics and AI inference. The breakdown shows that producing a single BTC costs roughly $93,700, split almost evenly between compute capital expenditure and electricity, with electricity accounting for approximately 46% of the total. By contrast, AI inference is dominated by hardware capex at around 76% of costs, with power contributing less than 5%.

The numbers highlight a flexibility advantage held by power-rich mining operators that pure AI data centers cannot easily replicate. Because electricity is such a large share of Bitcoin miners' operating costs, facilities with curtailment-friendly power contracts can switch capacity off when hashprice falls, a relief valve unavailable to operators whose outlays are sunk into GPUs and accelerators. That asymmetry has fueled a wave of miner-to-AI pivots across the industry.

Core Scientific announced a deal with AMD to expand into AI compute infrastructure, while MARA Holdings disclosed a $168 million agreement with EDF to secure power capacity that can be flexibly allocated. TeraWulf struck a 20-year lease with Anthropic, locking in long-term AI compute revenue on infrastructure originally built for Bitcoin mining. Each transaction reflects the same underlying logic Buterin's tables illustrate: the cost mix of mining lends itself to dual-use, where the cost mix of AI does not.

Buterin's post did not include market forecasts or investment guidance, and he framed the figures as an observation rather than a recommendation. The tables list approximate totals per unit of work for Bitcoin mining on one side and AI inference on the other, with capex and opex components itemized separately. According to the breakdown, miners with negotiable power agreements retain the ability to throttle output during low-margin periods, a margin management lever that AI operators, bound by fixed hardware amortization schedules, structurally lack.