Vitalik's $93,700 Table Prices Miner Risk, Not Bitcoin Risk

A single table, posted without a methodology note, generated more commentary this week than capital. Vitalik Buterin put the all-in cost of producing one bitcoin at $93,700: roughly $40,600 in direct electricity, $2,400 in ancillary power, 40% capex inclusive of financing costs, 14% everything else. JPMorgan's desk, working from a traceable model, had landed near $85,000. The 10.2% gap between those two figures is the entire story, and almost nobody read it. Every chart is a frozen moment of human emotion. This one freezes the quiet anxiety of a producer watching its input costs bid up by a competitor with a cheaper cost of capital. Bitcoin mining has always been a marginal-cost business wearing the costume of a commodity business. The protocol does not care what your ASIC cost, what lease you signed, or how you financed the rigs. It pays 3.125 BTC every ten minutes to whoever finds the block, and it adjusts difficulty every 2,016 blocks so the network keeps producing regardless of who leaves. That adjustment mechanism is the most important sentence in this entire debate, and the original commentary never mentions it. The structural pressure is older than AI. Subsidies halve on a four-year cadence, scheduled to reach 1.5625 BTC in 2028, while transaction fees have historically covered under 5% of miner revenue outside anomalous inscription manias. Miners are long a depreciating capital asset and short an issuance schedule they cannot influence. History repeats, but the narrative layer shifts — in 2018 the villain was a bear market, in 2022 it was leverage, in 2026 it is a rival bidder for the same substation. What AI changed is not the mechanism but the outside option. A miner holding a permitted grid interconnect now has a buyer for that capacity who is not another miner. That is genuinely new, and it is the only part of the story that deserves a price. Here is where the arithmetic becomes fragile. The $93,700 figure is a full cost — it amortizes capex including financing charges. That is correct accounting and wrong competitive analysis. When a miner bids for a power contract, it makes a marginal decision: the ASICs are already sunk, the lease is already signed, and the only question is whether revenue per megawatt-hour clears the electricity bill. An AI operator makes an incremental decision, sizing a new build against returns on the entire capital stack. Comparing full cost to marginal cost overstates how fast miners will actually surrender their sites. The evidence is asymmetric. We are given a table precise to the hundred dollars for mining and a single qualitative phrase — that AI's electricity share is "much smaller" — for the other side. No cost-per-inference breakdown, no utilization assumptions, no power usage effectiveness. When I built a miner cost model for an institutional brief in 2024, the number that moved the conclusion was never the electricity line; it was the assumed utilization rate on the capex. That input is missing here entirely. Then there is physics, which no spreadsheet amortizes. A bitcoin rack draws perhaps 5 to 15 kilowatts. A GPU training cluster draws 30 to 100 kilowatts or more per rack. Retrofitting means rebuilding cooling to liquid, reinforcing floor loading, and replacing 1-10 Gbps site networking with 400G InfiniBand fabric. The permitted interconnect transfers; the building rarely does. Two more omissions matter. First, difficulty is a negative-feedback loop: miners exit, difficulty falls, survivors' margins improve, equilibrium returns. Exit is gradual and self-correcting, not avalanche. Second, the "production cost floor" that market participants keep citing as support has failed repeatedly — in 2022 miners bled for months while BTC printed $16,000. Production cost is a supply-side constraint, not demand-side support. Confusing the two is a directional error, not a nuance. And a counterintuitive one: miners are persistent sellers. If marginal miners are squeezed out, forced selling declines. Miner capitulation has historically clustered near cycle bottoms, not tops. The bear-market instinct to read this as bearish may be exactly inverted. So the framing of miners being pushed out is probably backwards. The code is permanent; the meaning is fluid. What is actually happening is reclassification — operators choosing higher-margin hosting over hash production, because the same megawatt is worth three to five times more sold as AI capacity than sold as hashrate. Core Scientific's contracts with CoreWeave, Hut 8's hosting buildout, TeraWulf's conversion: not evictions, but voluntary upgrades. The drama lives in the passive verb; the economics live in the active one. There is also a moat nobody priced. Miners run interruptible load. In ERCOT they are paid to curtail when the grid tightens. Inference workloads cannot be curtailed on fifteen minutes' notice without breaching service levels. To a grid operator, the miner is a shock absorber and the AI campus is a liability. Add multi-year interconnect queues across Texas, the Nordics, and the Gulf, and permitted capacity looks less like a liability and more like an option on the AI buildout. The genuine fragility sits on the balance sheet, not the hashrate. Capex that includes financing costs means leverage, and leverage — not difficulty — is what broke Core Scientific in 2022. Clarity emerges only after the noise subsides, and the noise here is the headline. The tradable variable was never whether AI replaces mining. It is the relative price of cheap power against cheap capital, and which operator can finance hardware fastest. The question worth carrying into next quarter is narrower and harder: when the interconnect queue is the scarce asset and the ASIC is the depreciating one, how long before the market stops valuing miners as producers and starts valuing them as landlords?