1 · How the model works
Most “cost of production” charts draw one line and imply it is a price floor. That is misleading on two counts: miners do not share a single cost, and the cost itself moves with the price. This model is built around both facts.
Energy per bitcoin
Start with one machine of efficiency E, measured in joules per terahash (J/TH). Whatever its share of the network, it spends the same energy to earn one bitcoin, because both its power draw and its slice of the block reward scale with its hashrate. That share cancels:
where Hnetwork is total network hashrate (TH/s) and btc_per_day is daily issuance (the block subsidy, optionally plus transaction fees). A higher total hashrate means more energy per bitcoin for everyone. That is the difficulty channel, and it falls out of the arithmetic for free.
From energy to cost
Energy becomes money through the electricity price, then overheads and hardware are layered on:
all_in_cost = elec_cost × opex_mult + capex_per_btc (economic)
PUE is the datacenter cooling and power-delivery overhead; opex_mult covers non-energy operating costs (labour, maintenance, pool fees). Hardware capex is a timeline, not a constant: a terahash of gear cost roughly $20,000 in 2013 and about $15 in 2025, so the model carries a per-era rig price and amortises it over the machine’s life. Older cohorts carry less remaining book value, so fully-written-off machines are not charged like new rigs.
The fleet is a distribution
There is no single miner. Each day the model builds a two-dimensional spread of cohorts: over hardware efficiency E, and over the electricity price P that different operators pay. It reports the average all-in cost, the average electricity cost, and the 10th-to-90th percentile band across that spread. That band is what the chart reports, and what is worth reading.
The shutdown frontier
Production cost is reflexive. When the price falls, every operator whose electricity cost per coin exceeds the coin’s value switches off, because each coin mined loses money. The expensive tail drops out, so the average cost of the miners still running falls with the price. The model reproduces this explicitly: each day it removes cohorts whose electricity cost exceeds the price and renormalises over the survivors. The shutdown test is operational: a miner keeps running while power costs less than revenue, even with capex underwater, because capex is already sunk.
The marginal (price-setting) miner hugs the price by construction, because the worst surviving miner breaks even at the price. Do not read that line as “cost predicts price”. The causality runs both ways. The informative outputs are the average all-in cost and the band.
Two efficiency backbones
The fleet’s efficiency is the other main assumption. The model offers two sources. The hardware model takes the best-available efficiency from a hand-maintained timeline of ASIC launch figures and spreads the fleet above it. The measured (CBECI) mode instead derives the realised average efficiency from Cambridge’s network-power estimate divided by hashrate, and shapes the cohorts around that measured mean. Measured here means external and independent, not metered: Cambridge’s figure is itself an estimate.
2 · The electricity-price assumption
The default spread of power prices across the fleet is 3 to 7 US cents per kWh, triangular, weighted toward the cheap end. This is the single largest lever of subjectivity in the model, so it deserves evidence rather than assertion. In March 2026 we ran a structured research pass over primary sources to test it (see section 4).
In brief: the centre of the band is well supported, the tails are too tight after 2022, and nothing before 2021 could be verified.
The centre is right
Every verified fleet-average anchor lands between 4 and 5 cents, close to the band’s cheap-weighted mean of about 4.3 cents.
| Anchor | Value | Period | Source |
|---|---|---|---|
| CleanSpark, all-in power (owned sites) | 4.3¢ | Q1 2024 | SEC 8-K |
| TeraWulf, cost-to-mine calculator default | 4.5¢ | Q4 2024–Q1 2025 | Investor relations |
| Cambridge CBECI, global average assumption | 5.0¢ | constant | CBECI methodology |
| Hashrate Index, industry average estimate | ~4¢ → 5–6¢ | early 2022 → Jan 2023 | Luxor research |
The cheap tail runs below 3 cents after 2022
Cipher’s Odessa site buys power under a fixed-price contract at about 2.8 cents. Riot reported a fleet-wide all-in power cost of 2.8 cents in July 2025, net of the credits it earns by reselling contracted power back to the Texas grid during scarcity; its audited full-year 2024 figure was 3.4 cents. These sub-3-cent numbers are seasonal or monthly extremes, and demand-response credits do not scale to the whole fleet, but the cheap end is verifiably below the model’s 3-cent floor.
The expensive tail ran above 7 cents in 2022–2023
Before 2022, a typical North American hosting contract offered all-in power at 5 to 6 cents. Through the 2022 energy-price inflation, 8 to 9 cents became common and anything below 7.5 cents was called a bargain; Luxor’s hosting index averaged about 8.2 cents in the fourth quarter of 2022. The Antminer S19 Pro broke even near 9.2 cents at end-2022 economics, and that tail carried real hashrate. Even a well-run institutional miner, TeraWulf, averaged 7.8 cents in December 2024 during a winter spike. A fixed 7-cent ceiling clips a real segment of the fleet in those years.
Recommended bounds by era
If the model moves to era-varying defaults, the table below is the recommendation. The distribution stays triangular and cheap-weighted throughout; only the endpoints move.
| Era | Period | Low | Mode | High | Confidence |
|---|---|---|---|---|---|
| Hobbyist | 2013–2014 | 3 | 5 | 7 | low (no verified evidence) |
| China / Sichuan | 2015 – mid-2021 | 3 | 5 | 7 | low (unverified) |
| Post-ban migration | mid-2021 – early 2022 | 3.0 | 5.0 | 7.0 | high |
| Energy inflation | 2022–2023 | 2.5 | 5.5 | 9.0 | high |
| Institutional, post-halving | 2024–2025 | 2.5 | 4.5 | 8.0 | high |
| Current | 2026 | 2.5 | 4.5 | 7.0 | medium |
All values in US cents per kWh, all-in.
What is not established
No claim about the hobbyist era (2013–2014) or the China/Sichuan era (2015–2021) survived verification. Widely-cited figures such as Sichuan wet-season hydro at 1 to 2 cents circulate everywhere, but none came from a source strong enough to pass an adversarial check. The 3-to-7 cent band applied to those years rests on assumption. The share of hashrate in each price tier is unknown too, so the triangular shape is guesswork.
One modelling subtlety keeps the input band wide rather than pre-truncated: the shutdown frontier already removes miners who cannot cover their power bill. The input bounds should describe prices offered to miners (contracts and hosting rates) and let the shutdown logic decide who survives. Raising the high bound does not claim those miners were profitable; the model switches them off when the price falls.
3 · Where the data comes from
Three on-chain series are fetched daily, each from a chain of independent free providers; the first that answers wins.
- Bitcoin price
- blockchain.info, then CoinMetrics, then mempool.space (daily USD).
- Network hashrate
- blockchain.info, then CoinMetrics, then mempool.space (TH/s).
- Transaction fees
- blockchain.info, then CoinMetrics (BTC per day, optional).
- Block subsidy
- the deterministic halving schedule, computed directly.
- Fleet efficiency and capex
- a hand-maintained CSV of per-era ASIC efficiency and rig prices, or the Cambridge CBECI power estimate in measured mode.
- Reported-cost calibration
- cost-per-BTC that public miners (Riot, MARA, CleanSpark) report in quarterly filings, plus industry aggregates, overlaid as scatter points.
4 · How the power-price research was done
The electricity-price bounds were tested with a structured research pass in March 2026. The question was split into five angles: primary miner filings, methodology benchmarks, the China/Sichuan era, the marginal hosting tail, and net-of-credits economics. Each angle was searched independently; 22 sources were retrieved and 107 factual claims extracted.
The 25 most load-bearing claims were then verified adversarially: three
independent checks per claim, each trying to refute it, with a majority
needed to kill. Twenty survived; five were refuted and discarded. Primary
sources (SEC filings, company releases, the Cambridge methodology) were
favoured over commentary. The full write-up, including the
refuted claims that should not be repeated, lives in the repository at
docs/electricity-price-bounds.md.
Three claims that did not survive, and should not be cited: that Cambridge’s 5-cent default was calibrated from miner interviews (it is a stated convention); that industry research requires sub-5-cent or sub-3-cent power for profitability in 2025–2026; and a specific 3-cent hydro figure for one operator’s Paraguay site.
5 · Limits
This is a model, not a measurement. Nobody knows every miner’s power bill, so the electricity spread and the fleet efficiency are assumptions, exposed as sliders on the interactive version precisely because they are the main levers of subjectivity. The production cost is not a price floor: the price can fall below it, and it has. The marginal line follows the price by construction and should not be read as a forecast. Nothing here is investment advice.
The verified evidence also skews toward US-listed miners, the cheap and well-capitalised slice of the global fleet. Private, hosted-retail and off-grid miners carry a large share of hashrate and appear only indirectly, through hosting rates.
6 · Sources
- Cambridge CBECI methodology
- CoinShares mining reports, Q4 2025 and Q1 2026
- Riot Platforms, August 2023 production update (SEC 8-K)
- Riot Platforms, July 2025 production update
- CleanSpark, Q2 FY2024 results (SEC 8-K)
- TeraWulf, December 2024 update and cost-to-mine calculator
- Hashrate Index (Luxor), hosting-rate research 2022–2023
Research pass and figures current as of the CoinShares Q1 2026 report (March 2026).