Bitcoin's Hashrate as a Function of Mining Revenue and the Big Farms' Electricity Price: A Calibrated Model and Scenarios to 2036

Abstract. Mining calculators usually let the network's hashrate grow by a fixed rate, often 30 % a year. We test that against 154 months of data, January 2014 to October 2026: the hashrate, the miners' revenue in dollars, and the mining machines released since 2015. From 2019 to 2025 the hashrate grew by about 55 % a year, and in 2026 it stopped growing. Behind the 55 % are two things. The miners' revenue in dollars grew by 22 % a year, and the network's power use grew with it, by 22 %. The network needed 21 % less power per hash each year. A machine never gets better, so we count the machines: what the network needs beyond the machines installed is bought new, machines leave after five years, and the most efficient ones run. Counted this way, 29 to 49 % of the running hashrate was less than a year old in every year from 2019 to 2025, which is why the network's average followed the new machines, which got 22 % better a year. Cambridge's estimate of the network's power per hash gives the same 21 % a year for those years and stays within about 20 % of eight independent measurements. It fails where growth changes: it knows no quantities and weights machine models by age alone, so it shows 17 % less power per hash for the twelve months to October 2026, when no additional machine was needed; counted by machines it was 4 %. We describe the hashrate with one rule. The big farms count a mined coin at what it is worth a year ahead, the Power Law trend one year later. The network grows while the farms' running costs are below what mining pays at that value, the slower the closer the two come, and shrinks above that; in a year it closes 45 % of the distance. The running costs are power at USD 0.045 per kWh, the median of the miners surveyed by Cambridge, and USD 0.0105 per kWh for everything else a farm pays each month. Where growth ends is not fitted: it is the point at which mining stops paying. Only the pace is fitted. The rule fits only when the coins are counted at a slow, long-run value (r² 0.22 to 0.62); at the day's price it explains nothing. The trend a year ahead fits a little better than the trend of the same day, and it is the horizon at which a freely fitted end of growth falls where costs equal value; horizons from none to one and a half years forecast about equally well, four years worse. We also test whether farms sell extra coins at or around the top band to pay for new machines. It does not show in the hashrate: counting the coins at a price near the top band cannot be told apart from counting them a year ahead, the hashrate grew no faster after months at the top band than the rule gives, and the listed farms paid for their machines in 2021 and 2024 with shares and debt and sold coins in the weak years 2022 and 2026. The supply of machines was the limit twice, from late 2017 to mid-2018 and from late 2020 into 2022; it shows as a delay of one to two years, not as a ceiling near 30 % a year. In a forecast test from 2019 to 2025, a fixed 30 % was too low at every horizon, by a factor of 1.9 after four years; the model missed by a factor of 1.1 to 1.13 with its final pace and by 1.14 to 1.17 with the pace fitted only on earlier months. The model of version 3, which fitted the end of growth as well, missed by 1.3 to 2.0 when fitted on earlier months. For the next ten years the Power Law lets the miners' revenue in dollars grow by only about 5 % a year, because three halvings take back most of the price growth. Hashrate growth then depends almost entirely on new machines: about 34 % a year if they get 20 % better a year, 29 % at 15 % and 24 % at 10 %, on the floor path, fastest around 2030, when the machines bought from 2023 to 2025 are replaced. If the price stays where it is, the hashrate hardly grows at all. A reader's question led to one more test. At the halvings of 2016, 2020 and 2024 the hashrate lost 13 to 26 % and was back within 4 to 14 weeks; the model does not shrink its network at a halving either, and ten classes of power prices in place of the single price leave its forecasts unchanged. Read machine by machine, one price for every farm does not hold: a generation of machines paid its running costs at a median site for four to five and a half years, and up to two years longer where power is cheapest. The year 2026 speaks against the model: it called for growth, and the hashrate fell by 6 %. We state a test for December 2027.

Keywords: Bitcoin, mining, hashrate, difficulty, electricity price, hashprice, halving, miner revenue, mining machines, Power Law, out-of-sample validation.

Version 4, 14 November 2026. Data up to 31 October 2026. Version 1 let the whole network need 20 % less power per hash each year, the pace of Cambridge's estimate. Version 2 added what a new farm earns once personnel, rent, repairs and tax are counted (section 11). Version 3 counted the machines instead (section 4) and let the network grow until the farms' power bill took 90 % of what mining pays at the trend of the day. A reader of version 3 proposed three things: that the big farms count their coins at the fair value some time ahead, that the hashrate slows as it nears the point where mining stops paying, and that farms sell enough each month to pay their costs and sell extra at or around the top band to fund new machines. This version tests the three. The first two are now the model: a coin counts at the trend a year ahead, all running costs are counted, growth ends where the two are equal, and one number is fitted, the pace (sections 5 and 6). The third does not show in the hashrate (section 10); it is how the farm of the calculator's whale case now sells (section 11). Section 8 compares the new model with that of version 3. In the calculator the model switched a farm of older machines off at the next halving, and a reader asked whether it is broken, since the hashrate dipped only briefly after every halving so far. Section 13 tests this at the three halvings, by machine generation and with a spread of power prices in place of one price. The model of this paper runs in the mining calculator, where every assumption can be changed, and its formula is on the methodology page.

How to cite: Bitcoin Power Law Research (2026). Bitcoin's Hashrate as a Function of Mining Revenue and the Big Farms' Electricity Price: A Calibrated Model and Scenarios to 2036. Version 4, 14 November 2026. https://bitcoin-powerlaw.com/en/magazin/hashrate-forecast

1 Introduction

Whoever buys a mining device needs a forecast of the network's hashrate. A device's share of the blocks is its hashrate divided by the network's, so every percent the network grows is a percent less for the device. The usual forecast is a fixed growth rate. Our own mining calculator asked for one until November 2026, and 30 % a year is a common choice.

A fixed rate has no reason behind it. Mining is a business: a farm adds machines when they earn more than they cost and switches them off when they do not. Six questions follow, and this paper takes them in turn.

1. Is 30 % a year right? What did the hashrate do, and what would have been a good forecast?
2. Do the network's machines really get 20 % better a year? An old machine never gets better. How much of the network was new each year, and does Cambridge's estimate of the network's machines hold against what was measured?
3. Is the hashrate a function of the electricity price and of what Bitcoin pays? Most of the network runs in big farms. What do they pay for power and for everything else, at which price do they count their coins, and do the two explain the hashrate?
4. Or was the supply of machines the limit? If the price rose faster than the makers could build, the hashrate would have grown as fast as machines could be delivered, whatever mining paid.
5. Do farms sell extra coins near the top to pay for new machines? A farm that sells each month only what pays its costs, keeps the rest and sells it at or around the top band earns more per coin than the day's price suggests, and has the money for new machines just then. Does the hashrate show that?
6. Does the model hold at the halvings? After each halving the hashrate dipped for a few weeks and grew on. Does the model do the same, how long did a generation of machines stay in the network, and does it matter that farms pay different prices for power?

We answer with public data, report what speaks against our model as well as what speaks for it, and end with scenarios to 2036.

2 Data

Public series, averaged per calendar month, for the 154 months from January 2014 to October 2026:

• Hashrate: mempool.space, daily average of the network's hashrate, estimated from the blocks found and the difficulty [1].
• Miners' revenue: blockchain.com, new coins plus fees per day, in US dollars [2].
• Machines: the list of mining machines behind the Cambridge Bitcoin Electricity Consumption Index (CBECI): every model with its release date, hashrate and power draw [10]. From it we build the machines the network runs (section 4).
• Cambridge's power estimate: the CBECI "best guess" of the network's power draw in gigawatts [3]. Version 1 of this paper took the network's power from it; here it is the estimate we test.
• Measurements of the network's machines: eight published figures of the network's power per hash between 2018 and 2026, from the nonces in the blocks, from surveys of miners and from a count of machines [5][11][12][13][14][15], and the fleet figures listed miners report each quarter [16].
• Price: our daily closing prices, and from them the floor, trend and top band of our Power Law paper [4].

Power per unit of work is given in joules per terahash (J/TH). A machine's label states it for the machine alone; at the wall a farm needs about 10 % more for cooling and losses, the figure Cambridge uses [10]. Four measures are used throughout. The power bill is what the network's power costs a day; unless stated, power is priced at USD 0.045 per kWh (section 9). The power share is the power bill divided by what mining pays that day at the day's price. The running costs are the power bill plus everything else a farm pays each month, USD 0.0105 per kWh of power used (section 9). The cost share is the running costs divided by the value of the coins mined that day, a coin counted at the Power Law trend one year later. The trend is our measure of what a coin is worth in the long run; where the coins are valued in another way, the text says so.

3 What the hashrate did

Yearly averages; the shares are averaged on the log scale. Power and J/TH are those of the machines counted in section 4.

A fixed 30 % does not describe any stretch of this. Until 2018 the hashrate multiplied each year, as mining moved from the first chips to an industry. From 2019 to 2025 it grew by about 55 % a year, in single years between 22 % and 82 %. In 2026 it stopped: the monthly record is still the 1’086 EH/s of October 2025, and October 2026 stood 6 % below it.

The hashrate is power times machines. From 2019 to 2025:

hashrate            +55 % a year
= power             +22 % a year
× hash per watt     +27 % a year  (21 % less power per hash)

miners' revenue     +22 % a year  (price +55 % a year, two halvings)

The network's power use grew as fast as the miners' revenue in dollars. Everything beyond that came from a network that did more work with the same power. This is the core of the paper: the money mining pays sets how much power the network can use, and the machines set how much hashrate that power gives.

From 2019 on, the power took between 0.5 and 0.6 of the miners' revenue in every year but two. In 2020, the year of a halving, it took 0.82. In 2021 the price rose fourfold and half the network was switched off in China within weeks: the share fell to 0.24 and was back at 0.57 a year later.

4 The network's machines

Version 1 of this paper let the whole network need 20 % less power per hash every year, whatever else happened. That cannot be right as a rule. A machine never gets better: it needs the same power per hash on its last day as on its first. The network's average improves only when new machines are added and when old ones are switched off. From 2019 to 2025 the hashrate grew by 55 % a year, so a large part of the network was always new. Without growth, the average can improve only as fast as farms replace their machines. This section counts the machines, checks the count and Cambridge's estimate against what was measured, and says where Cambridge's estimate fails.

4.1 What Cambridge estimates

Cambridge does not measure the network's machines [10]. It lists the machine models of the three big makers, counts a model from two months after its release, leaves out the models that would lose money at USD 0.05 per kWh, and averages the rest. A model counts fully in its first year, with a weight of 0.8 in its second, then 0.6, 0.4 and 0.2, and not at all after five years. On top come 10 % for cooling and losses. Our replica of this method from the published list matches Cambridge's figures within 1 to 3 % since 2022.

Two things are missing in the method. It has no quantities: a model of which a million machines were sold counts as much as one that hardly sold. And the weights follow age alone: they are the same in a year in which the network doubles and in a year in which it shrinks. In October 2026, 37 % of the weight was on models less than a year old.

4.2 The network counted by machines

We count instead. Each month the network runs a hashrate. What it needs beyond the machines already installed is bought new, at the power per hash of the better half of the big three makers' models released in the 24 months before, two months after their release. A machine leaves after five years. Of the machines installed, the most efficient run; when the hashrate falls, the least efficient stand idle. At the wall they need 10 % more than on the label.

The five years rest on what was seen. The Antminer S9, released in 2016, made up to 78 % of the hashrate in 2018 [17], still 20 to 30 % from 2020 to 2022 and fell below detection by mid-2023 [12]. The S19 class, released in 2020, made about three quarters of the hashrate in 2023 and still about half in October 2025 [11][18]. The miners in Cambridge's survey say a machine's life "may extend beyond five years" [5]; listed miners write their machines off over three to four years [16].

From 2019 to 2025, 29 to 49 % of the running hashrate was less than a year old, and half to nearly all of it less than two years. In 2026 the count buys no machine at all, for the first time: 1’058 EH/s are installed, and the hashrate stayed below that all year.

The one check on these quantities we have is what the makers sold. Canaan sold machines for 7.2 EH/s in 2018 and 10.5 EH/s in 2019, and had 21.9 % of the hashrate sold in the first half of 2019 [27]: that makes about 48 EH/s for the whole market in 2019, where our count adds 44. In 2021 the count is too low: Canaan alone sold 22 EH/s, while our count adds 39 EH/s for the whole network, because it buys a machine only when the hashrate needs one, and that year machines stood in containers on their way out of China.

Yearly averages. From 2019 to 2025 the best machine on sale needed 21 % less power per hash each year, the new machines bought 22 % less, the network by our count 21 % less and by Cambridge's estimate 21 % less.

4.3 What was measured

The network's machines can be measured in three ways: from the nonces in the blocks, which carry the fingerprint of the machine model that found them; from surveys in which miners report their machines; and from counts of machines delivered. Eight such figures have been published. All are J/TH on the label, so we compare them with Cambridge's estimate and with our count without the 10 % at the wall:

On average Cambridge's estimate misses the measurements by a factor of 1.19 and our count by 1.17. Neither is exact, and the measurements differ among themselves by 10 % for the same month. Our count is too high for 2022, when the machines of 2017 and 2018 probably left sooner than after five years. Lives of four to eight years fit these figures about equally well.

4.4 Where Cambridge's estimate holds and where it fails

The level holds. Cambridge's estimate stays within about 20 % of what was measured, and so does our count. For the years 2019 to 2025 both give the same pace, 21 % less power per hash each year. Version 1 of this paper was right about those years.

It held for a reason that does not carry forward. The 21 % a year were no property of the machines in use. They came about because a third to a half of the running hashrate was new each year, and the new machines were 22 % better a year. Cambridge's age weights describe such a network well, by coincidence: a network that grows by half each year has about the age mix the weights assume.

It fails where growth changes. Three places show it.

• 2018. The network tripled within a year and was newer than any age weights allow: measured 89 J/TH, Cambridge 126.
• Since 2024. Growth slowed, the network grew older, and the estimate ran ahead of it: 21 % too low against the nonces in May 2024, 21 % too low against Luxor's count in July 2026. From mid-2023 to mid-2026 the measured figures fell by 15 to 18 % a year, Cambridge's estimate by 22 %, our count by 20 %. The listed miners, who buy more new machines than anyone, report fleets that have hardly changed since December 2024: CleanSpark 16 to 17 J/TH, MARA 17 to 19, Riot 20 to 21 [16].
• 2026. From October 2025 to October 2026 the hashrate fell by 6 %, and no additional machine was needed. Cambridge's estimate still shows the network using 17 % less power per hash (22.5 to 18.7 J/TH), because newer models keep entering its list and gaining weight. By our count it was 4 % (23.1 to 22.1 J/TH), from old machines standing idle. Luxor counted machines for about 1’150 EH/s in August 2026, of which about 915 EH/s were running [15]; our count has 1’058 EH/s installed.

What follows for the forecast. Cambridge's estimate is a fair picture of the past and no basis for the future. Carrying its 20 % a year forward for the whole network, as version 1 did, assumes that the network stays as young as it was. Over a long stretch the average does improve as fast as the new machines, growing or not; what growth decides is how far the average lags behind them. With machines that get 20 % better a year and live five years, a network growing by 55 % a year needs 1.36 times the power per hash of a new machine, and a network that does not grow 1.64 times. In 2019 the factor was 1.35; in November 2026 it is 1.53. When growth stops, the average loses a fifth against the new machines, and until old machines are replaced it hardly improves at all. That is what 2026 shows, and what the model now contains.

5 The model

In plain words: the big farms count a mined coin at what it is worth a year ahead. The network grows while their running costs are below what mining pays at that value, and shrinks above that. It does not jump there. Machines have to be ordered, delivered and connected, and old ones go off one by one, so each year the network covers only part of the distance, and the closer costs and value come, the slower it grows. How much power the network needs follows from its machines, counted as in section 4.

machines installed        = what was bought in the last five years, with the J per TH of its time
bought new                = hashrate − machines installed, if the hashrate is larger
J per TH of the network   = average of the most efficient machines installed, up to the hashrate, × 1.1
running costs (USD a day) = hashrate (TH/s) × J per TH × 24 h ÷ 1000
                            × (power price + other running costs, USD per kWh)
value (USD a day)         = (new bitcoin + fees per block) × 144 × price a year ahead
cost share                = running costs ÷ value
hashrate a month later    = hashrate ÷ cost share^(pull ÷ 12)

The model has one fitted number. pull = 0.6 is how fast the network moves: in a year it closes 1 − e^−0.6 = 45 % of the distance to the point where costs equal value, half of it in about 14 months. That point itself is not fitted. It follows from the question a farm asks: do the coins of a month, at the price I count on, pay what the month costs? Version 3 fitted this point as well, as a power share of 0.9, and section 8 shows what that cost.

Four inputs are assumptions that the reader can change in the calculator: the farms' power price (USD 0.045 per kWh), their other running costs (USD 0.0105 per kWh, section 9), how much better new machines get each year (20 %, from 12.4 J/TH on the label at the end of 2026) and the price path of Bitcoin. The price a year ahead is read from that path; in the fits it is the Power Law trend one year later, which today is 29 % above the trend of the day and in 2036 will be 20 % above it. The new bitcoin per block halve at every halving.

The model starts from the machines installed at the end of October 2026:

Of these, the most efficient 902 EH/s run in November 2026 and need 20.9 J/TH at the wall. A machine bought today needs 13.7.

Where the network stops growing, the running costs equal the value, so

hashrate at rest = value ÷ (J per TH × 24 h ÷ 1000 × (power price + other running costs))

Its growth follows from that line:

hashrate growth ≈ growth of the miners' revenue in dollars
                + gain in hash per watt of the machines running
                − growth of the farms' running costs per kWh

Two things about this point. It is the average machine that just pays its running costs there; a new machine, which needs a third less power, still earns money, and an old one has long been switched off. And the costs of the machines themselves are not in it: a farm that buys machines needs more than its running costs back. That is one reason why the network slows long before it gets there, and the pull is the measure of it.

6 Fitting the model

6.1 The fit

For every month we take the cost share and the growth of the hashrate over the following 6 or 12 months, as a yearly rate, and fit a straight line on the log scale. In the model the line runs through zero growth at a cost share of 1, and its slope is the pull. As a check we also fit the line freely and see where it crosses zero growth.

The windows overlap, so the number of months overstates the evidence. Eleven windows that do not overlap, from each January to the next, give the same picture: pull 0.59 and r² 0.53; fitted freely, growth ends at a cost share of 1.08.

The lower the cost share, the faster the hashrate grew in the year that followed. Sorted by the cost share, all months from 2019 on:

Growth slows as the costs approach the value. That is what the rule says, and what the months show: the hashrate doubled within a year when the costs took a third of the value, and grew by a fifth when they took two thirds. The months between 0.55 and 0.65 fall below the rule; half of them are the months of late 2020, which ran into the mining ban in China, and of summer 2025, which ran into 2026 (section 12).

We use pull 0.6, a round number between the fits since 2019 (0.5 to 0.6) and those from 2015 (0.66). Fitted freely, the line puts the end of growth between 0.64 and 1.11, and in the two windows with the most months at 0.89 and 0.92; without the months around the mining ban in China it is 0.98 and 1.01. So the data do not speak against an end of growth where the costs equal the value, and they cannot place it exactly either: since 2019 the cost share has stayed between 0.26 and 0.71. The network has not been observed at the point where the model lets growth end, and 2026 (section 12) shows that growth can end before it.

6.2 Which price do the farms look at?

The same fit with the coins counted in different ways. The first two columns give the r² of the model's line, which ends growth where the running costs equal the value; a negative r² means that the line is worse than assuming the average growth for every month. The last column fits the line freely and gives the cost share at which growth then ends.

The day's price explains nothing. When the price is high, mining pays well that day, but the hashrate of the following year does not grow faster for it. What explains the growth is the cost share measured with a slow, long-run value of the coins. That fits how a farm decides: a machine runs for years, a site takes a year or more to build, and both are paid for by what the coins will bring over that time, not by today's price. The day's price matters for the opposite decision, switching an old machine off, and that is what 2026 shows.

Farms look ahead, by about a year. Two things in the table say so. Counted at the trend half a year to two years ahead, the coins explain the hashrate better than counted at the trend of the day, most clearly in the last years: for the months from 2021 the r² over 12 months is 0.13 at the trend of the day and 0.43 a year ahead. And the freely fitted line tells how far. At the trend of the day it lets the network grow until the running costs are a quarter above the value of the coins, which no farm would do. At the trend four years ahead it ends growth when the costs take half the value, so farms do not count on the full four years either. Between half a year and one year ahead the line ends where the costs equal the value. We take one year.

How firm is the year? Not very. The forecast test of section 8 cannot separate horizons from none to one and a half years; only three and four years forecast clearly worse. Each of these values is the trend times a factor that shrinks slowly over the years: the trend a year ahead was 2.2 times the trend of the day in 2015, 1.65 times in 2019 and 1.29 times in 2026. In a fit such a factor is hard to tell apart from the calendar, and twelve years with three halvings are few. What speaks for the year is that it needs no fitted end of growth.

The floor is not what farms count on. Counted at the floor, today or a year ahead, growth would have to go on until the costs are two to three times the value. Farms have built as if their coins were worth clearly more than the floor.

The top band. The last three rows count the coins at or near the top band, as a farm would that sells its coins there. The middle between trend and top band fits as well as the trend a year ahead. Section 10 takes this up.

7 Was the supply of machines the limit?

There is another way to read the 55 % a year: mining paid so well that farms would have bought any number of machines, and the hashrate grew as fast as the makers could build them. Then the growth of the past would be the makers' capacity, perhaps 30 % a year, and not a reaction to what mining pays. The market for machines shows when that was the case.

Machines were the limit twice, from late 2017 to mid-2018 and from late 2020 into 2022. The year 2021 is the clearest case: the price rose fourfold, machines were sold out for seven to twelve months, and the hashrate grew by 22 %; by our count the network added machines for 29 % of its hashrate. The ban in China falls into the same year, so the two cannot be separated cleanly.

A ceiling near 30 % a year does not exist. In six of the seven years from 2019 to 2025 the hashrate grew faster than 30 %. In 2023, with machines in oversupply, the makers delivered machines for 115 % of the network of the year before, and in 2018 for almost three times the network. If we give the model a ceiling, the fit places it at +101 % a year since 2019, where it changes nothing (r² 0.52 with and without it), and at +263 % a year from 2015: the makers could always build far more than the network took in an average year.

The limit shows as a delay. The machines ordered in 2021 arrived in 2022 and 2023, and the hashrate grew by 51 % and 73 % while the price fell. Since 2019 the average price of the past two years explains the growth of the next twelve months almost as well as the trend a year ahead does (r² 0.50 against 0.54, fitted freely, section 6.2); from 2015 only the trend explains it (0.63 against 0.05). Both say the same about the last years: the hashrate follows a slow measure of what the coins are worth, with a delay of one to two years, and it does not follow the price of the day. The pull of the model is that delay. It is also the model's blind spot: in a year like 2021 the hashrate cannot grow as fast as the model wants, and in the years after it grows faster.

8 A forecast test

A fit to the past is not a forecast. For every quarter from January 2019 to October 2025 (28 starting points, 16 of them old enough for four years) we forecast the hashrate one to four years ahead and compare it with what happened. The table gives the average miss as a factor: 1.27 means the forecast was off by 27 % on average, up or down.

The model rows start from the machines installed at each starting point and value the coins with today's trend formula. Neither that formula nor the pace of new machines was known in 2019, so even the "earlier months only" rows have some hindsight in them. In the rows from "this model" on, growth ends where the running costs equal the value of the coins; only the pull is fitted. The rows of versions 1 and 3 are calculated again here, so they differ a little from the table printed in version 3.

Five readings. A fixed 30 % was the worst simple forecast for long horizons: after four years the real hashrate was almost twice the forecast. With their final numbers, version 3 and this model forecast equally well, by a factor of about 1.1; both were chosen with all data in view. In real time this model is far better. Version 3 fitted two numbers, where growth ends and how fast, and fitted on what was known at each starting point they came out too hopeful: a miss by a factor of 1.3 to 2.0. Here the end of growth is given by reason and only the pace is fitted, which came out between 0.67 and 0.71 at every starting point: a miss of 1.14 to 1.17, slightly on the high side, and better than any fixed rate. From the 2021 starting points on it misses by 1.11 to 1.14, against 1.20 to 1.39 for version 3 and 1.23 to 1.92 for the fixed 30 %. Counting all running costs matters: with power alone the model is too high at every horizon. The horizon is not settled by this test: a coin at the trend of the day, a year ahead or at the middle between trend and top band forecast about equally well, and only four years ahead is clearly worse. The pace of new machines hardly matters over four years either: most of what the network gains in that time comes from replacing machines that already exist. The model's use is not that it beat every constant. It says where the 55 % came from, and therefore why it does not continue.

One run over eight years. Started in January 2019 with the machines of that day and never corrected, on the coins really paid out, valued at the trend a year ahead, and with new machines as they came, the model stays within a factor of 1.12 of the real hashrate on average over 94 months; version 3 stays within 1.14:

It follows the hashrate through three price cycles and two halvings, and it misses 2026, where it is 25 % too high.

What each assumption rests on.

9 What the big farms pay for power

Cambridge's Digital Mining Industry Report surveyed 49 mining firms that together make up about 48 % of the network. Their median electricity price was USD 45 per megawatt hour, USD 0.045 per kWh, and electricity was more than 80 % of their cash operating costs; all running costs together came to USD 55.50 per megawatt hour of power used [5]. We take the median as the big farms' power price and the difference, USD 0.0105 per kWh, as their other running costs: personnel, rent, repairs and the like. The machines themselves are not in it. For comparison: a household in Germany, Austria or Switzerland pays six to nine times as much, which is why mining at home in these countries rarely covers its power.

The median stands for a wide spread. The miners in the survey reported power prices from USD 0.020 to 0.065 per kWh, half of them between 0.0385 and 0.055, and all running costs from USD 0.025 to 0.070 per kWh [5]. The listed farms' own reports show the same range: after its power credits Riot paid 2.6 cents per kWh in August and 3.9 cents in December 2025, IREN paid 3.0 cents at its largest site in June 2025, and MARA and Hut 8 paid 4 cents in the second quarter of 2025 [35]. A machine placed in someone else's farm in the United States cost 7.8 cents per kWh in 2023 [36]. For the size and the growth of the network the middle of this spread counts. For the question which machines still run, the spread itself counts (section 13).

In the model the power price works in two ways.

It sets the size of the network, not its long-run growth. In the formula of section 5 the hashrate at rest is the value of the coins divided by the running costs per kWh. Farms that pay 3.5 instead of 4.5 cents carry a network that is 22 % larger; at 6 cents it is 21 % smaller. Once that level is reached, growth is the same.

A power price that rises takes growth away. If the farms' power gets 3 % more expensive every year, the hashrate grows about 3 points less per year (31 % instead of 34 % on the floor path of section 11).

What farms pay is not the only price that matters. Since 2025 the same power can be rented to AI data centres, and a farm then compares mining with what the power would earn there. Section 12 comes back to this.

10 Do farms sell at the top to pay for new machines?

The third proposal describes how a farm handles its coins. Each month it sells as many as pay its running costs and keeps the rest. When the price stands at or around the top band, it sells extra, and that money buys new machines. If the big farms work this way, the money for machines comes in bursts, at the tops, and the hashrate should grow faster after them than the cost share alone gives. We take "at or around the top band" to mean a monthly price above the middle between trend and top band. Since 2015 that was the case in 26 months: November 2017 to March 2018, eleven months of 2021, and ten months between November 2024 and October 2025.

What the valuation says. A farm that sells its spare coins near the top band counts them at a price near the top band. Counted at the middle between trend and top band, the coins explain the hashrate as well as counted at the trend a year ahead (section 6.2), and the forecast test gives the same miss (section 8). Today the two are almost the same number: the middle between trend and top band is 1.27 times the trend, the trend a year ahead 1.29 times. In 2019 they were 2.0 and 1.65 times; in 2036 they will be 1.08 and 1.20 times. So the past cannot tell a farm that looks a year ahead from one that counts on selling near the top. The two part ways in the 2030s, when the top band will have come close to the trend: a network of farms that count on the top band would then grow less.

What the hashrate did after months at the top band. The months sorted by where the price stood, with the growth that followed and the growth the rule of section 5 gives for the same months:

After months at the top band the hashrate grew slower than the rule gives, not faster. Put into the fit as a number of its own, the part of the past 12 or 24 months that the price spent at the top band changes its sign with the window: fewer machines after tops in the fits from 2015 and 2017, more in those from 2019 and 2021. In the forecast test it brings no gain.

Where the money was. What the whole network had left after its running costs, and the machines it added (section 4.2):

The money was there: the top of 2021 left the network more than the four years before it together. And machines followed, 95 EH/s in 2022 and 285 EH/s in 2023. But they came one to two years later, when the price had fallen by two thirds, and they had been ordered in 2021, when machines were sold out (section 7). The model has this as its delay, the pull. After the months at the top band in 2025, with machines on offer everywhere, the network added nothing.

What the farms did. The listed farms publish what they do with their coins.

• In 2021, the year at the top band, they did not sell. In October 2021 the listed miners of North America held 20’459 BTC, more than USD 1 billion, and had "borrowed funds or issued equity for expenses and expansion" [29]. K33 writes about the same year that they "were able to raise equity or debt to pay for these expenses" [6].
• In 2022, at the low, they sold. From January to April it was 20 to 40 % of what they mined. In June, as the price fell to about $20’000, they sold about 14’600 BTC, almost four times what they mined that month, and their reserves fell by 23 % [6]. From January to November 2022, ten large listed miners sold about 40’300 of the 40’700 BTC they mined, 99 % [7].
• In late 2024, with the price at the top band again, they borrowed. Seven listed miners raised USD 5.2 billion with convertible notes from June to early December 2024, 3.6 billion of it in the last four weeks, and MARA used its part to pay running costs and to buy more bitcoin [30].
• In 2026 they sell again. In the fourth quarter of 2025 mining a coin cost the listed miners about $80’000 in cash on average, and the price then fell below that. Their reserves dropped by more than 15’000 BTC from the peak. Core Scientific sold about 1’900 BTC in January and planned to sell nearly all of the rest, Riot sold 1’818 BTC in December 2025, and MARA's loan against its bitcoin stood at 87 % of their value [8].

What a top can still bring. The top band keeps coming closer to the trend: it is 1.62 times the trend today, 1.40 times in 2029, 1.24 times in 2033 and 1.17 times in 2036. A farm that sold all its spare coins exactly there would get that much more for them than at the trend, and less with every cycle.

The answer. The proposal is right about two things. Farms do sell for their running costs: in weak years nearly everything they mine. And the money for new machines is raised when the price is high. But the big farms raised it with shares and loans and kept their coins, and they sold coins when money was short, at the lows. For the hashrate both ways come to the same: money raised at a top becomes machines a year or two later, and the model's delay holds that. A sale at the top band as a number of its own adds nothing that the hashrate shows, so it is not in the network model. It belongs to the single farm that cannot issue shares and has to pay for its next machines out of its coins. The farm in the calculator now does this (section 11).

11 Scenarios to 2036

Start: 902 EH/s on 14 November 2026 with the machines of section 5, new machines at 12.4 J/TH on the label, power at USD 0.045 and other running costs at USD 0.0105 per kWh, fees of 0.03 BTC per block, halvings in 2028, 2032 and 2036. Average hashrate growth per year until November 2036, with the hashrate reached in brackets:

For comparison, a fixed 30 % a year would give 12’400 EH/s. On every path the farms count a coin at the price the path has a year later. The fit of section 6 belongs to the trend path; the floor path is the cautious case, and the one the calculator starts with. Today the farms' running costs take 0.39 of the value of the coins on the trend path, 0.60 on the floor path and 0.84 if the price stays where it is.

Against version 3 the ten-year figures have hardly moved: +34 % a year on the floor path and +36 % on the trend path then as now. The new model starts faster on the floor path (+35 % in the first year, where version 3 gave +29 %) and slower on the trend path (+65 % against +71 %); by 2036 the two are within 5 % of each other on every path with a rising price. At a price that stays where it is today the new model ends lower, at 950 against 1’100 EH/s with new machines 15 % better a year.

Where the growth comes from. On the floor path the price is 12.5 times as high in 2036 as today, which is 29 % a year. Three halvings cut the new coins per block to an eighth. Together the miners' revenue in dollars grows by a factor of 1.7, about 5 % a year. The network's power use can grow no faster than that in the long run, plus the room it has today because the running costs are below the value of the coins: from 19 GW now to between 37 and 45 GW in 2036. Everything else is machines. If new machines never got better than today's, the hashrate would still grow by 16 % a year, because the network's average needs 20.9 J/TH today and a new machine 13.7.

How long can new machines get 20 % better a year? The best machines on sale went from about 98 J/TH (Antminer S9, 2016) over 29.5 (S19 Pro, 2020) and 21.5 (S19 XP, 2022) to 9.5 J/TH (S23 Hydro, 2026), about 17 to 21 % less each year. At 20 % a year for ten more years a new machine needs 1.3 J/TH in 2036, a seventh of today's best, and the network's average 2.3 at the wall. At 10 % a year it is 4.3 for a new machine and 5.9 for the network. We consider 10 to 15 % a year the more likely pace for the 2030s.

The growth slows within the decade. On the floor path with new machines 20 % better a year, the hashrate grows by 38 % a year until 2031 and 30 % a year after that; with 10 %, by 31 % and then 17 %. The reason is the Power Law itself: the floor's growth falls from 36 % a year today to 22 % a year in 2036, while every halving still takes half.

It is uneven. Year by year, on the floor path with 20 %: +35, +21, +28, +49, +56, +23, +20, +34, +46, +25 %. The weak years are those after a halving. The strong years around 2030 are a replacement wave. Nine tenths of the machines installed today were bought from 2023 to 2025. They reach five years from 2028 to 2030 and are replaced by machines that need half the power or less: the network's average falls from 19.1 J/TH in November 2027 to 7.6 in November 2030, by 26 % a year. In the first year it falls by 9 %. In reality the wave will be flatter, because machines do not all leave on their fifth birthday.

So is 30 % a year right? As a law, no. It was too low for 2019 to 2025 and far too high for 2026. For the next ten years it is roughly what comes out if Bitcoin follows its Power Law and new machines keep getting 15 to 20 % better a year, but that is a result of those two assumptions and not a constant. With new machines 10 % better a year, 20 to 25 % is the better number, and less in the 2030s. And if the price stays where it is, the hashrate does not grow.

What a new farm earns under the model. If the network grows until the big farms' running costs take most of what mining pays, a new machine cannot earn much more than buying bitcoin would. We check this with the mining calculator's whale case: 167 Antminer S23 XP Hydro (100 PH/s, 8.9 J/TH, USD 2.55 million), power at USD 0.045 per kWh, personnel and rent of USD 15 each per machine and month, repairs of 2 % of the hardware's price a year (a failed hashboard costs about USD 150 to repair), 21 % tax on the profit. Besides power these come to about a third of the power bill, so power is three quarters of the farm's running costs, close to the more than 80 % Cambridge finds. The farm has no money besides its hardware, and it sells the way section 10 describes: each month as many of the mined bitcoin as pay its running costs and tax at that month's price; the rest it keeps, and in months with the price at or around the top band it sells what it kept as well, as the money for new machines. It switches the machines off once their bitcoin no longer pay the running costs. Started on 14 November 2026, with fees of 0.03 BTC per block and new machines 20 % better a year:

Buying pays the fair value plus 10 % for fees. On every path the machines run for five years and five months and go off at the halving of April 2032. The first halving, in April 2028, halves what they mine, but an S23 needs less power than the machines the network buys today (8.9 against 12.4 J/TH) and keeps paying its running costs. By 2032 the network has grown about sixfold on new machines that need less than half an S23's power, and after the second halving what an S23 mines no longer pays its own. That is how long machines have run so far (section 13). On the floor and the trend path the price never comes near the top band, the farm sells nothing extra, and it ends with a quarter to a third fewer bitcoin than buying. On the top band path it sells everything it does not need for its costs, 13.8 BTC for USD 3.00 million: the price of 196 such machines today, a little more than the farm it started with. Without personnel, rent, repairs and tax the farm would keep 21.8 BTC at the floor and 20.0 BTC at the trend, about as much as buying, and sell 18.1 BTC for USD 4.02 million at the top band. A new farm bought at today's prices for the best machine ends up behind buying, and a higher price helps less than it seems, because the network grows with it and the machine mines fewer bitcoin.

With the Antminer S21 (17.5 J/TH) at the same 4.5 cents the machines go off at the first halving, in April 2028, after 17 months. Section 13 shows why, and what a cheaper site changes.

This also explains a result that looks too good. If the farm pays its running costs from money outside and keeps all its bitcoin, it ends with 36.6 BTC on the floor path, worth USD 30.9 million at the floor of November 2036, for USD 4.7 million spent on hardware and power. But the same USD 4.7 million spent on bitcoin as the bills fall due would buy 30.6 BTC, worth USD 25.8 million. Most of the more than sixfold is the rise of the Power Law floor, not mining. With personnel, rent, repairs and tax the farm spends USD 5.4 million and holds 36.6 BTC against 34.3 BTC bought, 7 % more. So the model and the calculator agree: once all costs are counted, a new farm does about as well as buying bitcoin, a little better if it can pay its bills from other money.

The mining calculator runs this model with today's hashrate and price. In the Whales scenario the farms' power price, personnel, rent, repairs and tax are under "More settings", the price path is under the chart:

12 The year 2026: evidence against the model

In October 2025 the hashrate stood at a record 1’086 EH/s, and the cost share was 0.66. The model forecast +33 % for the following twelve months; with the pull fitted on earlier months only, +36 %. Version 3, calculated again as in section 8, forecast +30 %, and +44 % fitted on earlier months. The hashrate fell by 6 %, to 1’024 EH/s in October 2026, with a low of 900 EH/s in July. Because the hashrate fell while the trend rose, the cost share dropped to 0.42, a level at which the hashrate grew by about 60 % a year in the past.

What happened is no secret. The price fell from an average of $114’000 in October 2025 to $63’000 in June 2026, close to the floor, and the miners' revenue from $54 million to $30 million a day. For the listed miners, mining a coin cost about $80’000 in cash at the end of 2025 and about $75’500 in the second quarter of 2026 [8][9], more than a coin brought for most of the year. Old machines were switched off: CoinShares puts the capacity that went offline at about 250 EH/s [19], and Luxor counted machines for about 235 EH/s standing idle in August 2026 [15]. At the same time AI data centres pay more for the same power: by CoinShares' estimate about $1.5 million a year per megawatt against $0.5 million for mining [9]. At least 35 EH/s is leaving the listed miners for good; one of them stopped mining in June 2026, and CoinShares expects two more to have left by the end of 2027 [9].

For the model this means four things.

• The day's price does matter on the way down. A farm looks a year ahead when it buys machines, and at today's price when it decides whether an old machine stays on. The model has only the first of the two. Mining a coin cost the network about USD 70’000 in running costs in October 2025 and USD 58’000 in October 2026, by our count. At the day's price that was 0.61 of a coin in October 2025, 0.71 in October 2026 and 0.84 in mid-November 2026: for the average machine, and the older half of the network is well above it.
• Looking ahead did not help in 2026. In October 2025 the trend a year ahead stood at USD 107’000. A year later the price was USD 82’000 and falling. A farm that had counted on the fair value of the next year was wrong by a quarter, and farms short of money sold coins at the low (section 10).
• What a farm pays for power is no longer what its power is worth. A farm that can rent its site to an AI customer compares mining with that rent. With a farm power price of USD 0.097 instead of 0.045 the model shows no growth for the coming year on the floor path, and with 0.15 on the trend path. That is one way to read 2026: the network behaves as if its power cost 10 cents or more.
• Without growth the network's machines hardly improve. Version 1 of this paper let the network need 20 % less power per hash in 2026 as in every year, and so does Cambridge's estimate. Counted by machines it was 4 % (section 4.4). That alone takes a good part of the recovery out of the first year: on the floor path the model now shows +35 % for the first year, where version 1 showed +44 %.

We have not changed the model because of one year. But the first years of every scenario in section 11 start from a network that is, by the model's own measure, too small for what mining pays at the floor or above. Whether it recovers is the test.

13 Halvings, machine generations and a spread of power prices

Run through the mining calculator, the model produced a result that a reader questioned. A new farm of Antminer S21 at 4.5 cents per kWh switched its machines off at the halving of April 2028, a year and a half after it started. But after the halvings of 2016, 2020 and 2024 the hashrate dipped only briefly. Is the model broken? The suspicion falls on one assumption: in the model every farm pays the same USD 0.045 per kWh, so all machines of one kind stop paying in the same month. In reality power prices are spread, and old machines keep running where power is cheap. We test this in three steps: what the hashrate did at the three halvings and what the model gives, which machines paid their running costs at one price and at a spread of prices, and how long a generation of machines lasted.

The halvings. The daily hashrate, averaged over seven days and compared with its highest value in the 30 days before the halving, next to the monthly averages the model works with:

Others measured the same: in 2020 about a fifth of the hashrate went off within four days [31], in 2024 the 7-day average fell by 11 % within three weeks [32]. The model, started three months before each halving, shows no dip. It moves in monthly steps, and after the three halvings the farms' running costs took 0.22, 0.66 and 0.70 of the value of the coins a year ahead, so it let the network grow on, slower than before. Over the following twelve months it was too low each time. At the level of the whole network the model held: a halving does not shrink its network either. What it does not show is the dip of a few weeks.

Which machines paid. Now machine by machine. For each month we take the machines that run by our count (section 4.2) and ask which of them mined less than their running costs. For the spread we use Cambridge's survey: the miners reported power prices from USD 0.020 to 0.065 per kWh, half of them between 0.0385 and 0.055 [5]. We split the network into ten equal parts along that range, at 2.4, 3.1, 3.9, 4.1, 4.4, 4.7, 5.1, 5.5, 5.9 and 6.3 cents. The highest month of the eight after each halving:

A rule that keeps a machine running only while its coins pay its costs at the price of the day fails in 2020. By that rule four fifths of the network would have gone off in June and July 2020, and half or more would have stayed off until October. It lost a quarter at its lowest point and was back after eight weeks. Our count has too many old machines for 2020 (82 J/TH on the label where Cambridge estimates 58), but by Cambridge's figure the average machine only just paid, so about half of the network did not. The spread of power prices does not repair the rule. Its middle is the same 4.5 cents, so the cheap sites save about as many machines as the dear sites lose, and even with the oldest machines at the cheapest sites the summer of 2020 does not come out. What does fit is the way this paper counts the coins. At the trend a year ahead, between none and 13 % of the running hashrate did not pay after the three halvings, which is the size of the dips that were seen.

The Antminer S9 shows it. Four years old, it still mined about 32 % of the blocks in May 2020, down from 78 % two years before [33]. At the hashprice of June 2020 it paid for its power only below 3.3 cents per kWh, and for all running costs only below 2.2 cents. The Block calculated a loss of 56 % of revenue for it at USD 0.05 per kWh and still a loss of 17 % at USD 0.032, its price for Sichuan's rainy season [34]. A third of the network ran machines that paid their running costs only at the very cheapest sites. Their owners either had power that was nearly free for those months or ran them through a loss, counting on a higher price. A year later the price was almost four times as high.

How long a generation lasted. The power price up to which a machine paid all its running costs at the price of the day, as the average of each year, in cents per kWh:

At the median site, 4.5 cents, the S9 paid until February 2020, just under four years after its release. The price of 2021 brought it back, and from May 2022 it was done there. At 2.4 cents, the cheapest tenth of the survey, it paid through most of 2022 and hardly at all from 2023. That matches what was measured: 20 to 30 % of the hashrate from 2020 to 2022 and gone from the blocks by mid-2023 (section 4.2). The S17 paid at the median site, with a few months' break, until the halving of 2024, five years after its release. The S19 fell below 4.5 cents for the first time in August 2024 and for good in November 2025, after five and a half years, and it still pays below about 3 cents. CoinShares wrote in March 2026 that every machine below an S19 XP loses money at 6 cents or more, and put those machines at 15 to 20 % of the network [19].

So a generation pays at a median site for four to five and a half years, until the first or second halving after its release, and for one and a half to two years longer at the cheapest sites. The five years of our count stand. The spread of power prices makes the tail: it is why the S9 was seen in the blocks for seven years and not for five.

The model's future, read the same way. On the floor path of section 11 an S21 pays the running costs of section 9 up to a power price of 6.3 cents per kWh until March 2028 and up to 2.6 cents after the halving of April 2028. An S23 pays up to 13.4 cents today, 6.1 cents after that halving, 4.7 cents in early 2032 and 1.8 cents after the halving of 2032. For the single farm of the calculator, which also pays personnel, rent, repairs and tax (section 11), this gives the month in which the machines go off, again on the floor path:

The farm that raised the question is the S9 of 2020. An S21 is one generation behind: it needs 17.5 J/TH where the machines bought today need 12.4. Bought in its third year and run at a median site, it meets its first halving and stops, as the S9 did at 4.5 cents. At 2.5 cents the same machines run until 2031, and the S23 runs to the halving of 2032 at every price up to 4.5 cents. In the model's network the same halving looks like 2020. On the floor path 58 % of the running hashrate does not pay at the price of the day in April 2028, and 42 % at the price a year ahead; on the trend path it is 45 and 26 %. The network adds 4 % in the seven months that follow, and by the end of 2029 the share is below 20 %. A dip of some weeks, as after every halving so far, is to be expected on top, and the model will not show it.

The farm in the calculator switches off at the price of the day, although the network does not. It has no money besides its coins, so a month that does not pay has to be paid from the coins it kept. We tried the other rule, to run on while the coin pays at the price a year ahead. The S21 farm at 4.5 cents then runs until December 2029 on the trend path and ends with 5.8 instead of 7.2 BTC; on the floor path nothing changes. On a price path without a bull market the months of loss are not earned back, so the farm keeps its rule.

Did the model hold? In three parts.

• For the network it held. At none of the three halvings did the hashrate fall for longer than a few weeks, and the model does not let it fall either. With ten classes of power prices in place of one, each growing by its own cost share, the forecast test of section 8 gives the same miss (a factor of 1.13, 1.14, 1.11 and 1.11 after one to four years, against 1.13, 1.13, 1.10 and 1.10) and the same growth after the halvings. The rule "the slower, the closer" already is what a spread looks like from a distance: machines leave one site after the other and not all in one month.
• One power price for every machine did not hold. Read machine by machine, the single price of 4.5 cents switches a whole generation off in one month. That did not happen, and it is not what the model's network does. The spread of section 9 belongs to every statement about single machines and single farms, and a farm's result depends on its own power price more than on anything else (the table above).
• Two things the model does not have. It shows no dip in the weeks after a halving, and it was too slow in the year after each of the three halvings. And on the floor path it enters the halving of 2028 with running costs at 0.96 of the value of the coins. The past never went above 0.71, so what the network does there has not been observed (section 15).

14 A test of the model

On the floor path and on the trend path of section 11 the model puts the hashrate above its old record by December 2027, whatever the pace of new machines: at about 1’270 EH/s on the floor path and about 1’580 EH/s on the trend path. If the price stays where it is today, below the floor, the model shows about 975 EH/s and no record.

If Bitcoin's price averages at or above the floor in 2027 and the monthly average hashrate of December 2027 is not above 1’086 EH/s, the record of October 2025, the model is wrong in this form. Then the farms stop growing long before their running costs reach the value of the coins, unlike from 2015 to 2025, most likely because their power has a better use, and the costs in the model have to go up. If the hashrate is above the record, 2026 was a dip in a rule that still holds.

15 Limitations

• The machines are counted, not measured. Nobody knows how many machines of which model run. Our count assumes that every machine leaves after five years and that farms buy the better half of the recent models. Against eight measurements it misses by a factor of 1.17 on average, and lives of four to eight years fit about equally well. The fixed five years also make the replacement of 2028 to 2030 sharper than it will be.
• The machine list. It covers the three big makers and has no quantities. Machines of other makers and machines running on modified firmware are missing.
• The makers' capacity. The model lets farms buy any number of machines at once. Twice in twelve years that was not possible (section 7).
• One power price. The network model gives every farm the median of one survey of 2025, in every year since 2015. The miners surveyed paid between 2.0 and 6.5 cents. For the network's total a spread changes nothing that we can measure, but which machines run, and for how long, depends on it (section 13).
• No dip at a halving. The model moves in monthly steps. The 13 to 26 % of the hashrate that went off for some weeks after each halving are not in it, and in the year after each halving it was too slow (section 13).
• 2026. The model missed the direction of the last twelve months (section 12).
• Hindsight. The trend that values the coins is today's formula. The pull was chosen with all data in view. The real-time rows of section 8 are the fairer evidence, and there the model was a little too high.
• The end of growth is reasoned, not observed. Since 2019 the running costs have taken at most 0.71 of the value of the coins a year ahead. That growth ends at 1 follows from the model's idea and from a line through lower values (section 6.1).
• The other running costs. The USD 0.0105 per kWh come from the same one survey of 2025 as the power price. The costs of the machines themselves and of money are not counted.
• China 2021. The mining ban removed half the hashrate for reasons outside the model. Without the months around it the freely fitted end of growth from 2015 is at a cost share of 0.98 to 1.01 instead of 0.89 to 0.92.
• Few independent years. Twelve years, three halvings and four price cycles carry the fit. Overlapping windows make the r² look firmer than it is, and neither the fits nor the forecast test can separate a look ahead of one year from none, from one and a half years or from a price near the top band (sections 6.2 and 10).
• New machines. How much better new machines get is an assumption, and over ten years it is the input the result depends on most (section 11).
• Fees. The scenarios keep fees at 0.03 BTC per block. With 0.3 BTC per block the hashrate in 2036 would be 43 % higher, because fees are not halved.
• The price path. Every scenario rests on a price path, three of them on our Power Law model, which has its own limits [4].
• No financial advice. The paper describes what the hashrate did and what follows from stated assumptions. It does not say that a mining device will pay.

16 Data and reproducibility

Every series is public and free: the hashrate from mempool.space (/api/v1/mining/hashrate/all), the miners' revenue from blockchain.com (chart miners-revenue), the machine list and the power estimate from Cambridge (CBECI, version 1.4.0: the SHA-256 mining equipment list and the best guess at USD 0.05 per kWh), and the daily closes with trend and floor from our own file /power-law/btc-daily.csv. All series are averaged per calendar month.

The machines are counted month by month from January 2015. The power per hash of a new machine is the average of the better half, by J/TH, of the models of Bitmain, MicroBT and Canaan released in the 24 months before, each counted from two months after its release. What the monthly hashrate needs beyond the machines installed is bought at that figure; machines leave 60 months after they were bought; the most efficient machines installed run, and their average times 1.1 is the network's J/TH at the wall. The power share is the hashrate × J/TH × 24 h × USD 0.045, divided by the miners' revenue of the day. The cost share is the hashrate × J/TH × 24 h × USD 0.0555, divided by the coins paid out that day × the trend 365 days later. The pull is the slope of a line through zero, of the yearly growth rate on the log scale against minus the log of the cost share. The forecast test starts each quarter from the machines counted up to then, lets the model run on the coins really paid out, and compares on the log scale. For section 13 the daily hashrate is averaged over seven days. A machine pays its running costs in a month when the miners' revenue per TH/s and day, or the coins paid out counted at the trend a year ahead, is above its J/TH × 1.1 × 24 h × (power price + USD 0.0105 per kWh). The ten classes of power prices follow the survey's lowest value, quartiles and highest value in straight lines. The model's lines are in section 5, and the methodology page lists the values the calculator uses today. You are welcome to use, reproduce and extend this work. Please cite it as shown at the top.