Most people read the headline as a verdict. They should read it as a ledger entry, because that is all it is: a row of numbers somebody wrote down, with the methodology left off the page.
On a Friday afternoon in October, Vitalik Buterin posted a cost breakdown for Bitcoin mining. The figure he printed was $93,700 per coin. The breakdown split roughly into electricity at $40,600, other power-related costs at $2,400, capital expenditure at around 40% of the total, and the rest absorbed by operations and financing. JPMorgan, working from its own models, had landed closer to $85,000. The difference between the two numbers is 10.2%, which is not a rounding error; it is the gap between a marginal producer and an industry average. And in that gap sits the entire argument, unexamined.
I spent the first half of my career auditing Solidity before it was fashionable to do so. In 2017, in Istanbul, I sat in a rented office above a currency exchange and reviewed forty thousand lines of contract code across three token projects during the ICO frenzy. I found three reentrancy vulnerabilities and five integer overflows. The team leads wanted to ship. I would not sign. Two of those projects eventually raised anyway; one of them lost its treasury within a year to a flaw I had flagged as medium severity. What that period taught me was not that auditors are heroes. It taught me that a number without a stated methodology is not evidence. It is a claim wearing a lab coat.
So let us treat this claim properly.
The structural logic holds. Electricity is roughly 46% of the cost basis in the table Buterin published, counting the $40,600 in direct power and the $2,400 in adjacent power costs. A miner whose cost basis is nearly half electricity is far more sensitive to a one-cent move in kilowatt-hour pricing than a business whose electricity is a rounding error on a capital-intensive balance sheet. That elasticity difference is real, and it is the honest core of the argument. When a grid operator raises rates or a utility renegotiates a power purchase agreement, the miner feels it in the same quarter. An AI inference operator amortizing a GPU cluster over three years feels it as a line item they can hedge, defer, or pass through to a customer paying per token.
But here is where the argument becomes arithmetic theatre. The $93,700 figure includes financing costs on capital expenditure. That is correct accounting. It is also the wrong lens for competitive analysis. When a miner bids for a power contract, they are making a marginal decision: the ASIC is already bought, the lease is already signed, the sunk cost is sunk. The question is whether the next kilowatt-hour produces more revenue than it costs. When an AI operator bids for the same site, they are making an incremental decision on a new build with new financing. Comparing a fully loaded cost against an incremental cost and concluding that one party is being priced out is not analysis. It is a category error that flatters the conclusion.
A fully loaded cost basis and a marginal cost basis are two different instruments, and Buterin's table measures only the first while the competitive argument requires the second.
The same asymmetry runs through the other side of the comparison. The article that carried this discussion described AI's electricity share as "much smaller" without a single number attached. No cost per dollar of inference, no kilowatt-hour per thousand tokens, no breakdown of training versus serving. Precision on one side, adjectives on the other. The persuasive force of any argument is bounded by its weakest evidence, and here the weakest evidence is doing structural load-bearing work.
There is a physical constraint that gets skipped entirely, and it is the one I would flag first if this were an audit. Converting a mining facility into an AI data centre is not a software upgrade. Rack density tells the story: a mining cabinet draws somewhere between five and fifteen kilowatts, while a GPU training cluster sits at thirty to over a hundred. That difference propagates through everything. Substation capacity, cooling architecture (air to liquid, which means new plumbing and new floor loading), structural reinforcement, and network fabric. Mining sites typically run one to ten gigabits of connectivity. AI clusters want four hundred gigabit InfiniBand. "The miners hold permitted grid capacity" is true, and the article is right to raise it, but the window to monetize that capacity is far narrower than the framing implies.
Yet here is the contrarian turn, and it cuts against the thesis rather than for it. Mine the permits more carefully. Grid interconnection rights are the scarce asset, not the racks. Interconnection queues in most developed markets now run three to seven years. Environmental review, water rights, noise ordinances, community hearings: an AI data centre faces a materially stricter approval path than a mining operation, because a mining operation is an interruptible load that a grid operator can curtail during peak demand and compensate, while an inference cluster serving paying customers is not. In the eyes of a grid operator, a curtailable miner is a high-quality asset and a latency-sensitive AI tenant is a difficult one. That is an implicit moat, and the article's framing inverts it into a weakness.
The second structure the discussion omits is the one that governs everything: difficulty adjustment. Every 2,016 blocks, roughly two weeks, the protocol retunes. Miners exit, hashrate falls, difficulty drops, the remaining miners earn more per unit of work, and the system finds a new equilibrium. This is not a footnote. It is the core adaptive mechanism of a proof-of-work network, and any argument about miners being "pushed out" that does not account for it is describing a cliff where a gradient exists. The exit is gradual and self-correcting. What the AI competition does is accelerate a clearing process that the protocol was designed to absorb.
That distinction matters enormously for how one reads the consequences. A miner leaving does not damage Bitcoin. The miner is a service provider selling hashrate for block rewards, not a value capturer. Hashrate declines, difficulty drops, the protocol persists. The damage, if there is any, lands on the equity of leveraged mining companies whose capital structures assumed a coin price above their fully loaded cost. The 2022 template is unambiguous: Core Scientific did not fail because hashrate left the network. It failed because the debt was denominated in dollars and the collateral was denominated in a coin trading at sixteen thousand. Reading this discussion as a Bitcoin thesis is a category error of a different kind.
And on the price itself: the production cost floor narrative has failed repeatedly. Anyone who treated eighty-five to ninety-three thousand as support in a prior cycle watched it evaporate. Production cost is a supply-side constraint, not a demand-side floor, and conflating the two has cost more capital than any single protocol exploit I have audited.
The most honest sentence in the entire discussion arrives at the end, and it quietly demolishes the headline above it. Cheap power still favours the miner; cheap capital favours the AI operator; the next round is decided not by who mines more cheaply but by who can finance hardware fastest. If the binding constraint is capital, then the electricity story is a component, not the mechanism. The author knew the thesis was conditional. The headline maximized distribution anyway. That is the media business, and one discounts for it the way one discounts for any known bias in a data source.
The energy arbitrage thesis itself has an earlier lineage than most readers realize. During DeFi Summer in 2020, I ran a team analysing fifteen liquidity pools to model impermanent loss under volatility. We built a static hedging algorithm that cut user slippage by twelve percent during peak hours, and I refused to deploy it for three weeks while we backtested against 2017 data. The lesson was not that hedging works. It was that mechanisms which appear to offer free efficiency are almost always repricing risk somewhere invisible. AI operators bidding up power contracts are not creating efficiency. They are repricing grid access, and the repricing is visible only to whoever holds the interconnection agreement.
That is where I would look rather than at the coin. Mining equities are about to split into two valuation regimes: power-asset operators, who will be valued on data-centre multiples if they sign the right hosting contracts, and pure hashrate producers, who will be valued as commodity manufacturers. The spread between those two regimes is three to five times on the same underlying asset, and the market has not yet done the sum-of-parts work to separate them. Core Scientific with CoreWeave, Hut 8's hosting line, TeraWulf's pivot: those are not rescue operations, they are reclassifications. Wrapping the same electricity asset as AI hosting instead of mining is worth multiples, and the incentive to reclassify is overwhelming regardless of whether any miner is ever "forced out."
Which raises the question I cannot answer from a tweet and neither can anyone reading it. Arthur Hayes appearing in the discussion with a tokenized GPU mining network is not a neutral data point; it is product placement inside an analysis. The structural history of that model, from io.net to Render to Akash, is littered with projects that solved for token emission and never solved for verified compute supply. A GPU network is a bridge between the two sides of this arbitrage, and bridges collect tolls whether or not the river floods.
So the question is not whether AI displaces mining. The question is who holds the interconnection agreement when the grid operator decides it needs a curtailable load, and whether that holder is capitalized in dollars or in hashrate. Track the capital structure, not the coin price. The metal is in the permits, and the permits do not appear on any cost table I have seen published.