The $93,700 Cost Table Compares the Wrong Two Numbers

A table with no inputs is not a model. It is an assertion wearing a decimal point. Vitalik Buterin published a mining cost breakdown that put Bitcoin production at $93,700 per coin. Direct electricity: $40,600. Other power-related line items: $2,400. Capital expenditure with financing costs baked in: roughly 40% of the total. Everything else: 14%. The number landed, the quote-tweets stacked up, and inside a single news cycle a cost table had become a thesis — AI is coming for the miners' power sockets, and Bitcoin's production layer is the casualty. I have spent enough time on the audit side of this industry to recognize the shape of that argument. In 2017 I ran a standardized checklist across 40-plus ICO whitepapers at the peak of the bubble, cross-referencing claimed tokenomics against historical market-cap data. Twelve failed on arithmetic alone. The lesson was never that the teams were dishonest. It was that a precise-looking number without disclosed inputs is a claim, not a measurement — and that markets, including smart ones, will price the claim anyway. The direction of Buterin's argument holds. The arithmetic does not survive contact with how miners actually make decisions. The gap between those two facts is where the trade sits. Here is what the table actually says, stripped of commentary. Electricity consumes about 45.9% of the per-coin cost — $43,000 of $93,700 once you add the $2,400 in auxiliary power. Capital expenditure, amortized and inclusive of financing, takes roughly 40%. The remaining 14% covers labor, ASIC maintenance, and everything downstream. The structural conclusion is clean: a miner's cost base is dominated by a single input whose price is set by someone else, in a market where the marginal bidder now includes hyperscale data centers. | Cost line | Share of $93,700 | Implication | |---|---|---| | Direct electricity | $40,600 | Price-taker on a single input | | Other power-related | $2,400 | Same exposure, smaller line | | Capex incl. financing | ~40% | Leverage is the real fragility | | Everything else | 14% | Labor, maintenance, overhead | Compare that against JPMorgan's $85,000 estimate. The 10.2% gap matters more than the headline. JPMorgan's figure is an industry-aggregate model. Buterin's number sits above it and above spot. If a cost figure exceeds the current price, then what the table describes is not the industry average — it is the marginal producer, the highest-cost cohort still hashing. That is a completely different object of analysis, and the distinction never gets made. This is the kind of methodological detail that most readers skip and that institutional clients pay for. In 2024 I reviewed settlement structures across five spot Bitcoin ETF issuers and found a 0.05% efficiency gap that nobody had bothered to model; that gap was the entire alpha. Same discipline applies here. Read the fine print before you trade the headline. The rest of the context is familiar to anyone who has read a miner's 10-K. Block subsidy plus fees, with fees historically under 5% of revenue outside inscription spikes. A hard cap of 21 million. A halving every 210,000 blocks, next one cutting the subsidy to 1.5625 BTC. Roughly 94% of supply already issued. What is new is not the supply model. What is new is that the power socket a miner sits on now has a second willing buyer with a deeper balance sheet and a higher tolerance for electricity prices. That buyer is also, in most jurisdictions, a policy priority. That is a real structural change. It is also routinely overstated, because the analysis stops at the electricity line and never asks what the marginal decision actually looks like on either side of the trade. Start with the accounting mismatch, because it is the flaw that propagates into everything else. Buterin's $93,700 is a fully loaded cost: capex amortized, financing costs included. That is correct accounting. It is the wrong number for competitive analysis. A miner deciding whether to run a rig tonight does not re-underwrite the ASIC it bought three years ago. The capex is sunk. The decision is marginal — does revenue per terahash exceed marginal electricity cost, yes or no. An AI operator signing a new power purchase agreement is making an incremental decision on fresh capital: new GPUs, new liquid cooling, new interconnection, new everything. I built my first automated liquidation engine on Aave V1 in 2020 and the single biggest lift came from standardizing exactly this distinction between sunk exposure and marginal exposure. Getting it wrong produced false positives at scale. Getting it right cut them by 15%. Compare a fully loaded miner cost against an incremental AI cost and you will systematically overstate how fast miners get pushed out. The miner with depreciated hardware and a legacy power contract can run profitably at a power price where a greenfield AI datacenter cannot clear its cost of capital. The correct comparison is marginal-to-marginal, and on that basis the miner's disadvantage is far narrower than a 46% electricity share implies. Now the physical constraint that the "miners already hold permitted grid capacity" argument glosses over. Converting a mining facility into a GPU training cluster is not a software upgrade. Rack density in a mining container runs 5 to 15 kW per cabinet. GPU training racks run 30 to 100-plus kW. You rebuild cooling, usually to liquid. You reinforce floor loading. You replace the network. Mining sites are provisioned for 1 to 10 Gbps; AI clusters want 400G InfiniBand. The interconnection right is genuinely valuable and genuinely hard for a new entrant to buy quickly. But the conversion cost per megawatt is large, the permitting is not automatically transferable, and the monetization window is narrower than the narrative assumes. Then there is the mechanism the analysis ignores entirely: difficulty adjustment. Every 2,016 blocks — roughly two weeks — the protocol reprices hashing. Miners exit, hash rate falls, difficulty drops, and the surviving cohort's cost per coin falls with it. This is not a footnote. It is the single most important economic feature of Bitcoin's production layer, and it makes miner exit a gradual, self-correcting process rather than a cliff. AI competition is a second variable stacked on top of a mechanism that was already designed to absorb exits. The framing of "AI pushes miners out" implies a linear collapse. The protocol says otherwise. Code executes what words promise. And the evidence is asymmetric in a way that should force any careful reader to discount the conclusion. We have a line-item miner cost table down to the dollar. We have, on the AI side, the phrase "electricity is a much smaller share." No number. No inference-cost breakdown. No split between training — capex-dense, power-dense, genuinely electricity-sensitive — and inference, which is a continuous operating load with a different cost profile. Buterin's claim was about inference. The part of AI that competes hardest for power is training. Swap the object of comparison and the argument weakens materially. Confidence on the mechanics: high. Confidence on the numbers: medium. A cost table with no disclosed kWh, no power price, no ASIC efficiency assumption is not reproducible and therefore not auditable. It is an estimate that looks like a model, and the difference between those two things is the difference between a research note and a marketing asset. The blind spot is that this is being framed as displacement when the data describes reclassification. Look at what has already happened. Core Scientific contracted capacity to CoreWeave. Hut 8 built an AI hosting business. TeraWulf repositioned. In every case the miner is selling megawatts and shell space, not hashrate. That is not a miner being pushed out. That is a miner choosing a higher-margin business with the same physical asset. The framing converts a voluntary margin decision into a forced eviction because the forced version travels further on social media. The valuation consequence is where the actual inefficiency sits. A megawatt packaged as AI hosting commands a materially higher multiple than the same megawatt packaged as Bitcoin mining — the spread has run in the three-to-five-times range depending on contract duration and counterparty. Same substation. Same interconnection queue position. Same transformers. Two very different multiples. Participants have been slow to do sum-of-the-parts work on miners holding both an asset book and a mining cash flow stream. Structure precedes profit; chaos demands a fee. The second blind spot is the moat nobody is pricing: interruptibility. Texas miners participate in demand response. They shut down when ERCOT tightens and get paid for it. That makes them a flexible, dispatchable load — a premium asset from the grid operator's perspective. AI inference load is not interruptible in the same way. You cannot tell a customer mid-request that the cluster is going dark for ninety minutes. The curtailment revenue stream is a genuine competitive advantage that appears in no cost-per-coin table, and it argues the opposite of the headline: the miner's position in the power market may be more defensible than the AI operator's, not less. The third blind spot is regulatory. Grid interconnection is not a data center permit. Water rights, noise ordinances, environmental review, and community hearings all bite harder on a hyperscale facility than on a container yard. The permitted capacity miners hold is scarce precisely because the permitting path for everyone else got longer. Watch the interconnection queues, not the tweets. And the historical record should temper the whole discussion. The "production cost is a price floor" narrative has failed repeatedly. In 2022, with miners deeply underwater and BTC printing $16,000, the cost floor did nothing at all. Production cost is a supply-side constraint. It is not a demand-side support. Anyone treating $85,000 or $93,700 as a level to buy is repeating a mistake that already cost people real money, and the loss was entirely predictable from the mechanism. Which brings the framing back where it belongs. Miners being squeezed is a business-cycle problem for a business. It is not a token-economics problem for the asset. Miners are service providers selling hashrate for subsidy. If some exit, sell pressure from a constant cohort of sellers actually declines. Survival is a function of liquidity, not optimism — and BTC's liquidity does not depend on any single miner's balance sheet. One more caution on the solutions being floated. Arthur Hayes' proposed Bitcoin-style AI mining network is tokenized GPU capacity wearing a familiar wrapper. The structural problem in that sector was never technical; it was proving that the GPUs are real and that demand is organic rather than circular. Most of these networks historically collapsed into wash-mining against their own incentive schedules. I run AI in my own stack, but as transparent rule-based decision trees trained on ten years of my own P&L, not as a black box. Technology accelerates a logic; it does not replace one. Tokenizing stranded compute does not solve a power-price competition, it just moves the risk from the utility bill to the token curve. Two things are being priced at once and the market is collapsing them into one. Bitcoin the asset is close to indifferent to which cohort of miners survives. Miner equities are not — they are about to split into two baskets: capacity operators that get infrastructure multiples, and pure hashing operations that get commodity-producer multiples, with a persistent spread between them that will not close on earnings. The real variable is not whether AI displaces mining. It is the relative price of cheap power versus cheap capital, and which of the two is the binding constraint at any given moment. Right now capital is. That is a financing question, not an electricity question, and it changes which miners win. So ask the question the table never answers: if the binding constraint has shifted from cost per kilowatt-hour to the speed at which you can finance hardware, what exactly is $93,700 measuring — and who is still using it to size a position?