I drew a chart juxtaposing the Bitcoin hash rate with the market availability of mining ASICs and their energy efficiency. This allows calculating with certainty the lower and upper bounds for the global electricity consumption of miners. I decided to do this research after seeing that so many other analyses were flawed. See for example the flagrant errors in Digiconomist’s Bitcoin Energy Consumption Index. I believe that my market-based and technical approach is superior and more accurate.
[Edit: this post was also published in Bitcoin Magazine.]
I split the timeline in 10 phases representing the releases and discontinuances of mining ASICs. See the references and a commentary on the data behind this chart:
I reached out to some Bitcoin ASIC manufacturers when doing this market research. Canaan was very open and transparent (thank you!) and gave me one additional extremely useful data point: they manufactured a total of 191 PH/s of A3218 ASICs.1
Determining the upper bound for the electricity consumption is then easily done by making two worst-case assumptions. Firstly we assume that 100% of the mining power added during each phase came from the least efficient hardware available at that time that is still mining profitably.2 Furthermore, despite A3218 being the least efficient in phases 5-8 we can only assume 191 PH/s of it were deployed, and the rest of the hash rate came from the second least efficient ASIC:
- Phase 0: 290 PH/s @ 0.51 J/GH (BM1384)3
- Phases 1-3: 150 PH/s @ 0.51 J/GH (BM1384)
- Phase 4: 40 PH/s @ 0.25 J/GH (BM1385)
- Phase 5: 191 PH/s @ 0.29 J/GH (A3218) +
159 PH/s @ 0.25 J/GH (BM1385)
- Phase 6: 670 PH/s @ 0.25 J/GH (BM1385)
- Phase 7: 350 PH/s @ 0.20 J/GH (Bitfury 28nm)
- Phase 8: 150 PH/s @ 0.13 J/GH (BF8162C16)
- Phase 9: 1250 PH/s @ 0.15 J/GH (A3212)
- Average weighted by PH/s: 0.238 J/GH
Secondly we assume none of this mining power, some of it being barely profitable, was ever upgraded to more efficient hardware.
Therefore the upper bound electricity consumption of the network at 3250 PH/s assuming the worst-case scenario of miners deploying the least efficient hardware of their time (0.238 J/GH in average) is 774 MW or 6.78 TWh/year.
Now, what about a lower bound estimate? We start with a few observations about the latest 4 most efficient ASICs:
- Bitfury BF8162C16’s efficiency can be as low as 0.06 J/GH. But the clock and voltage configuration can be set to favor speed over energy efficiency. All known third party BF8162C16-based miner designs favor speed at 0.13 J/GH (1, 2). Bitfury’s own private data centers also favor speed with their immersion cooling technology (1, 2, 3). The company once advertised the BlockBox container achieved 0.13 J/GH (2 MW for 16 PH/s), presumably close to the efficiency achieved by their data centers. But we want to calculate a lower bound, so let’s assume the average BF8162C16 deployed in the wild operates at 0.10 J/GH.
- KnCMiner Solar is exclusively deployed in their private data centers and achieves an efficiency of 0.07 J/GH.
- Bitmain BM1387’s efficiency is 0.10 J/GH.
- Canaan A3212’s efficiency is 0.15 J/GH.
Therefore the lower bound electricity consumption of the network at 3250 PH/s assuming the best-case scenario of 100% of miners currently running one of the latest 3 most efficient ASICs (at best 0.10 J/GH) is 325 MW or 2.85 TWh/year.
Can we do better than merely calculating lower and upper bounds? I think so, but with the exception of Canaan,1 other mining hardware manufacturers tend to be secretive about their market share, so anything below are just educated guesses…
Virtually all of the 1750 PH/s added after June 2016 came from BF8162C16, BM1387, and A3212, with the latter having the smallest market share. So the average efficiency of this added hash rate is likely around 0.11-0.13 J/GH. This represents 190-230 MW.
I would further venture that out of the 1500 PH/s existing as of June 2016, perhaps half was upgraded to BF8162C16/BM1387/A3212, while the other half remains a mixture of BM1385 and A3218. This represent 750 PH/s at 0.11-0.13 J/GH, and 750 PH/s at 0.26-0.28 J/GH, or a total of 280-310 MW.
I believe an insignificant proportion of the hash rate (less than 5%?) comes from all other generations of ASICs. Bitfury BF864C55 and 28nm deployments were upgraded to BF8162C16. KnCMiner/GoGreenLight represents 0.3%. BM1384 is close to being unprofitable. RockerBox, A3222, Neptune have long been unprofitable.
Therefore my best educated guess for the electricity consumption of the network at 3250 PH/s adds up to 470-540 MW or 4.12-4.73 TWh/year.
Economics of mining
Given the apparent high energy-efficiency, hence relatively small percentage of mining income that one needs to spend on electricity to cover the operating costs of an ASIC miner, it may seem that mining is an extremely profitable risk-free venture, right?
Not necessarily. Though mining can be quite profitable, in reality it depends mostly on (1) luck about when BTC gains in value and (2) timing of how early a given model of mining machine is put online (compared to other competing miners deploying the same machines.) I say this as founder of mining ASIC integrator TAV, as an investor who deployed over time $250k+ of GPUs, FPGAs, and ASICs, and as someone who once drove 2000+ miles to transport his GPU farm to East Wenatchee, Washington State in 2011 in order to exploit the nation’s cheapest electricity at $0.021/kWh—yes it was worth it!
To demonstrate the real-world profitability of a miner, I modeled the income and costs generated by an Antminer S5 batch 1 ($418, 590 W, 1155 GH/s, released on 27 December 2014) assuming mined bitcoins are sold on a daily basis at the Coindesk BPI, and assuming $0.05/kWh. See income-antminer-s5.csv:
- On 27 December 2014 (day 1) mining starts; 15% of the first day’s mining revenues ($0.71 out of $4.64) is spent on electricity, generating an income of $3.93.
- On 15 January 2016 (day 385) $854 of income has been generated (84% of the lifetime income); 39% of the daily mining revenues ($0.71 out of $1.84) are being spent on electricity.
- On 15 July 2016 (day 567) $1013 of income has been generated (99% of the lifetime income); 78% of the daily mining revenues ($0.71 out of $0.90) are being spent on electricity.
- On 8 October 2016 (day 652) $1021 of income has been generated (lifetime income); the daily mining revenues become insufficient to cover electricity costs. This day is economically and rationally the end-of-life of the S5. Past this point, mining is unwise or at best futile: either very small amounts of money will be made, or large amounts will be lost, depending on day-to-day BTC price fluctuations.
- On 15 March 2017 (day 810) if the miner had continued mining, the total income actually manages to reach $1030. This represents $0.06 gained every day beyond day 652. Futile indeed.
The lifetime electricity costs are $0.71 × 652 = $463. The lifetime mining revenues are $463 + $1021 = $1484. A miner who had invested $418 into purchasing an S5 would have, after deducting electricity costs, turned it into $1021, a 2.4× gain. So mining was quite profitable!4
An S5 mined $1484 over its lifetime:
- 31% ($463) recoups electrical costs
- 28% ($418) recoups the hardware cost
- 41% ($603) are potential profits but a part of them are needed to recoup other expenditures: data center, labor, etc
I also modeled other miners:
Antminer S7 batch 1 ($1823, 1210 W, 4860 GH/s, released on 13 October 2015) in income-antminer-s7.csv. As of 15 May 2017 the S7 continues to be profitable, has operated for 581 days, and has mined $3323 ($842 electricity + $2481 income). Out of $3323 mined by the S7:
- 25% ($842) recoups electrical costs
- 55% ($1823) recoups the hardware cost
- 20% ($658) are potential profits
Antminer S9 batch 1 ($2100, 1375 W, 14.0 TH/s, released on 12 June 2016) in income-antminer-s9.csv. As of 15 May 2017 the S9 continues to be profitable, has operated for 338 days, and has mined $3531 ($558 electricity + $2973 income). Out of $3531 mined by the S9:
- 16% ($558) recoups electrical costs
- 59% ($2100) recoups the hardware cost
- 25% ($873) are potential profits
Canaan Avalon 4 ($?, 680 W, 1.0 TH/s, released on 12 September 2014) in income-avalon4.csv. Its last profitable day was 8 July 2016 (666 days of operation). It spent 31% of its total mining revenues to recoup electricity.
Canaan Avalon 6 ($1100, 1000 W, 3.5 TH/s, released on 21 November 2015) in income-avalon6.csv. As of 15 May 2017 it continues to be profitable, and has operated for 542 days. So far it has spent 32% of its total mining revenues to recoup electricity.
Canaan Avalon 721 ($888, 900 W, 6.0 TH/s, released on 21 November 2016) in income-avalon721.csv. As of 15 May 2017 it continues to be profitable, and has operated for 176 days. So far it has spent 28% of its total mining revenues to recoup electricity.
Bitfury BF864C55 (comes in different configurations, model assumes a 0.50 J/GH, 5000 W, 10.0 TH/s machine in operation since 3 March 2014) in income-bitfury55.csv. As of 15 May 2017 it continues to be profitable, and has operated for 1170 days. So far it has spent 9% of its total mining revenues to recoup electricity.
Bitfury 28nm (comes in different configurations, model assumes a 0.20 J/GH, 2000 W, 10.0 TH/s machine in operation since 28 February 2015) in income-bitfury28.csv. As of 15 May 2017 it continues to be profitable, and has operated for 808 days. So far it has spent 16% of its total mining revenues to recoup electricity.
Bitfury BF8162C16 (comes in different configurations, model assumes a 0.13 J/GH, 1300 W, 10.0 TH/s machine in operation since 12 October 2016) in income-bitfury16.csv. As of 15 May 2017 it continues to be profitable, and has operated for 216 days. So far it has spent 24% of its total mining revenues to recoup electricity.
KnCMiner Solar (comes in different configurations, model assumes a 0.07 J/GH, 700 W, 10.0 TH/s machine in operation since 4 June 2015) in income-solar.csv. As of 15 May 2017 it continues to be profitable, and has operated for 712 days. So far it has spent 6% of its total mining revenues to recoup electricity.
We can calculate the upper bound for the global electricity consumption of Bitcoin miners by assuming they deploy the least efficient hardware of their time and never upgrade it. As to the lower bound it can be calculated by assuming everyone has upgraded to the most efficient hardware. The table below summarizes the model’s estimates as of 26 February 2017:
|Lower bound||Best guess||Upper bound|
|Power consumption (MW)||325||470-540||774|
|Energy consumption (TWh/yr)||2.85||4.12-4.73||6.78|
|Energy consumption (Mtoe/yr)||0.245||0.354-0.407||0.583|
|Energy consumption (quad Btu/yr)||0.010||0.014-0.016||0.023|
|Percentage of world’s energy consumption5||0.00260%||0.00376-0.00432%||0.00619%|
|Electricity cost (million USD/yr)6||$142||$206-$237||$339|
|Energy efficiency (J/GH)||0.100||0.145-0.165||0.238|
I kept the above table for reference as of the time this post was written, and have also produced new estimates as of 28 July 2017 (10-day avg hash rate of 6398 PH/s):
|Lower bound||Best guess||Upper bound|
|Power consumption (MW)||640||816-944||1248|
|Energy consumption (TWh/yr)||5.61||7.15-8.27||10.93|
|Energy consumption (Mtoe/yr)||0.482||0.615-0.711||0.940|
|Energy consumption (quad Btu/yr)||0.019||0.024-0.028||0.037|
|Percentage of world’s energy consumption5||0.00512%||0.00652-0.00755%||0.00998%|
|Electricity cost (million USD/yr)6||$280||$357-$413||$547|
|Energy efficiency (J/GH)||0.100||0.128-0.148||0.195|
This may sound like a lot of electricity but when considering the big picture I believe Bitcoin mining is not wasteful. Also an interesting comparison to make is that according to a 2008 study from the United States Energy Department’s Energy Information Administration (EIA) these figures are comparable to or less than the annual electricity consumption of decorative Christmas lights in the country (6.63 TWh/year.)
Lastly, when modeling the costs and revenues of a miner over its entire life such as the Antminer S5, we find out that the hardware cost is as high as, if not higher than its lifetime electricity cost. Therefore a miner’s business plan should not look at the electricity costs alone, and cannot trivialize hardware costs when calculating expected profitability.
On 11 March 2017 I removed the assumption that sales of A3218 dwindled down to practically zero post-June 2016, because although sales volume did decrease I do not have precise metrics to justify it.7
On 13 March 2017 I made the calculation of the upper bound for the electricity consumption more accurate (was 861 MW, now 774 MW), thanks to A3218 production volume provided by Canaan.
On 16 March 2017 I added the section Economics of mining.
On 30 March 2017 I added the comparison to the electricity consumption of decorative Christmas lights.
On 16 May 2017 I reworked the section Economics of mining to add more miners such as S7, S9.
On 4 June 2017 I added all miners released in the last 2.5 years to the section Economics of mining.
On 28 July 2017 I produced updated estimates in the conclusion.
References and commentary
The chart covers the period 15 December 2014 to 26 February 2017. Starting as early as December 2014 is sufficient for accurate modeling because only one ASIC released in phase 0 is still profitable: Bitfury BF864C55. All others are no longer profitable.
The daily hash rate data was obtained from Quandl; the curve was smoothed out by calculating each day as the average of this day and the 9 previous ones.
The cost of electricity is assumed to be $0.05/kWh which is half the worldwide average. It is logical to assume miners seek geographical locations with the cheapest electricity.
All energy efficiency values given in joule per gigahash are reported at the wall, taking into account the power supply’s efficiency.
Mining hardware manufacturers only sell one generation of miners at any given time. Usually it is a result of producing and selling small batches one by one, as Bitmain and Canaan have done. But it is also a result of aggressive competition: when a company launches a new ASIC significantly outperforming the efficiency of the competition, their sales come to a stop until a more efficient successor is available, as Canaan CEO N.G. Zhang recounted. Therefore my model actually errs toward overestimating electricity consumption by assuming that the previous ASIC generation is being sold/deployed at the same rate until the very day preceding the introduction of the next generation, which we know is not true in some cases.7
- Neptune launched in June 2014 and achieves 0.70 J/GH.
- Solar launched in June 2015 and achieves 0.07 J/GH. The company declared bankruptcy in May 2016, however they certainly stopped deploying mining capacity months earlier. This chart assumes they stopped in January 2016. Later, KnCMiner was bought by GoGreenLight. So far they have not added new hash power, but merely reactivated the hardware they acquired.
- RockerBox was included in the SP20/SP30/SP31/SP35 Yukon product series; it launched in May 2014 and achieves 0.66 J/GH. The company failed to launch its successors—PickAxe, RockerBox II—and declared bankruptcy in May 2016, however they certainly stopped selling products months earlier as they were far behind competition in terms of energy efficiency. This chart assumes sales stopped in January 2016.8
- BF864C55 launched in March 2014 and achieves 0.50 J/GH.
- Their 40nm ASIC never entered full-scale production, hence its absence from the chart.
- Their 28nm ASIC launched in February 2015 and achieves 0.20 J/GH.
- BF8162C16 launched in October 20169 and achieves 0.06 to 0.13 J/GH. This wide efficiency range is due to the ASIC being operated in a variety of configurations—sometimes manufactured by third parties—from air cooling (1, 2) to immersion cooling (1, 2, 3) where voltages and clocks are pushed to their limits.
- BM1384 was included in the AntMiner S5 series; it launched in December 2014 and achieves 0.51 J/GH.
- BM1385 was included in the AntMiner S7 series; it launched in October 2015 and achieves 0.25 J/GH.
- BM1387 was included in the AntMiner S9/T9/R4 series; it launched in June 2016 and achieves 0.10 J/GH.
- A3222 was included in the Avalon 4/5 series; it launched in September 2014 and achieves 0.68 J/GH.
- A3218 was included in the Avalon 6 series; it launched in November 2015 and achieves 0.29 J/GH.
- A3212 was included in the Avalon 721/741 series; it launched in November 2016 and achieves 0.15 J/GH.
In an email exchange with Canaan staff (VP of Engineering Xiangfu Liu, Jon Phillips, and Wolfgang Spraul) they gave me the following historic metrics:
- number of wafers made: ca. 950
- number of chips per wafer: ca. 3330
- performance per chip: 25 GH/s
- total: 79 PH/s
- number of wafers made: ca. 1100
- number of chips per wafer: ca. 3650
- performance per chip: 47.5 GH/s
- total: 191 PH/s
Hardware that is no longer profitable has obviously been retired. As of 26 February 2017 (difficulty = 441e9, 1 BTC = 1180 USD, and assuming $0.05/kWh) an ASIC is profitable if its efficiency is better than 0.56 J/GH:
1e9*3600 (hashes per hour of 1 GH/s) / (2^32 * 441e9 (difficulty)) * 12.5 (BTC reward per block) * 1180 (USD per BTC) / 0.05 ($/kWh) * 1000 (Wh/kWh) = 0.56 J/GH
So 3 ASICs in the chart are no longer profitable: Neptune, RockerBox, and A3222. ↩
Most of the hardware deployed during phase 0—CPUs, GPUs, FPGAs, first-generation ASICs—has not been profitable for a long time, so we make the assumption these miners upgraded to the least efficient ASIC available at the end of phase 0 that is still profitable: BM1384. ↩
However another investor who on 27 December 2014 bought $418 worth of bitcoins would be worth $818 on 8 October 2016, a 2.0× gain. It could be argued that a large reason why mining was profitable came simply from BTC gaining value. ↩
In an email thread with Canaan staff, although they were unable to give metrics, they confirmed that sales of the A3218 slowed down after June 2016 when competitor Bitmain launched BM1387 with a 3× better efficiency. ↩ ↩2
I reached out to Spondoolies CEO Guy Corem to get official confirmation of when their sales stopped, but have not received a reply so far. ↩
In a 2015 post on HackerNews I previously incorrectly assumed hardware based on this ASIC launched in December 2015 but large-scale deployment only started in October 2016, as shown by this tweet from Bitfury Executive Vice Chairman George Kikvadze. ↩