In 2021 there were 6.6 million electric vehicles sold worldwide and, this year, that number is expected to rise to 10.5 million. Global laptop sales are projected to fall slightly, from 277 million last year to 272 million in 2025. Figures like these are the result of in-depth research and, in many cases, are based on actual data reported directly by the sectors' major players.
Yet we have no clear industry consensus of how many bitcoin mining machines will be sold this year - not even how many were sold last year. The bitcoin mining hardware sector is new on the radar of mainstream financial analysts and the major players do not always openly report sales volumes. A more nuanced approach is needed to estimate the size of the market. Rather than taking a bottom-up approach - trying to stitch together as much of the opaque sales information as we can from what is provided by manufacturers - we can instead resort to the most transparent information source of all: the bitcoin blockchain.
Starting from the most general and narrowing in, let's take a top-down approach to reasoning. Using bitcoin's network hashrate as a proxy measure for the number of machines on the network, accompanied by data-derived assumptions about the characteristics of those machines, we can calculate an approximation for the size of the bitcoin mining hardware market. What we learn in the process will allow us to generate projections about its future.
What's the combined hashrate of all the machines in circulation?
Let's start with what we do know. Every 2016 blocks, bitcoin mining difficulty is automatically adjusted so that the next 2016 blocks are found, on average, every 10 minutes. For any given historical period and prevailing difficulty level, we can infer the expected network hashrate over that period by comparing the actual rate at which blocks are found with the target rate. Essentially, we guess the network hashrate by comparing how quickly blocks are found with the rate at which the network was expecting them to be found. It is this estimate that commentators refer to when they discuss network hashrate. At the time of writing, the 30-day average network hashrate stood at 190,000,000 TH/s or 190 EH/s.
This figure represents the hashrate of only those bitcoin mining machines hashing to the network at any given time. It doesn't account for all the hashrate that could be provided by machines that are currently not hashing to the network. The sum of these two distinct quantities reflects the maximum potential network hashrate. It is this quantity that most closely reflects the number of machines in circulation though, unfortunately, it has no readily available measure. We need to adjust what we know already, the hashrate of the machines currently online, by making a few additional assumptions.
Say over the course of a machine's lifetime it spends, on average, 2 - 3 months not dead but not hashing and that it has a 4 - 5-year average life expectancy. We, therefore, expect the machine to be offline approximately 5% of the time. Conversely, at any given moment approximately 95% of the miners in circulation are online. So, for a network hashrate of 190 EH/s, we estimate that there are approximately (190 / 95%) EH/s = 200 EH/s worth of mining hardware in circulation.
How many machines are in circulation?
To understand how many machines 200 EH/s represents, we first need to know the average hashrate per machine in circulation.
Four of bitcoin's major mining pools, Binance Pool, Poolin, Slush Pool, and ViaBTC, openly report both their aggregate hashrate and the number of machines hashing to them. Leveraging this transparency allows us to compute the average hashrate per machine across each of these pools.
Interestingly, there is a significant divergence amongst them: the average machine hashing to Slush Pool and Poolin outputs 58 TH/s and 55 TH/s respectively. For Binance Pool it's 38 TH/s and is as low as 23 TH/s for ViaBTC. This divergence implies that none of these pools represent an unbiased sample of the bitcoin network: different miners, each with different quality hardware, favour different pools. Perhaps hobbyist operators with old generation hardware are more likely to mine to ViaBTC, whereas industrial-scale miners, likely boasting the latest generation hardware, are more likely to favour Slush Pool or Poolin.
Let's instead revert back to the network. Data provider CoinMetrics provides insights on the proportional split of the different models on the network. Bitmain Antminer S7s and Antminer S9s both have a signature nonce distribution that allows us to estimate their contribution to the network hashrate. Despite both models having been long eclipsed in terms of performance, these old generation machines still contribute a significant proportion of the network's hashrate. CoinMetrics' latest calculations put the current network hashrate split at 5% for the S7s, 25% for the S9s, and the remaining 70% among all other models.
The S7s' 5% hashrate contribution represents around 10 EH/s which, for a model with around 5 TH/s output, corresponds to approximately 2,000,000 machines. The 25% attributed to the S9s constitutes around 50 TH/s or approximately 4,000,000 machines. Using an average hashrate approximation of 70 - 90 TH/s for all other machines, they make up 140 EH/s and around 1,500,000 - 2,000,000 machines. Therefore, adding these all together, we estimate that there are around 7,500,0000 - 8,000,000 machines in circulation. Of these machines, at least three-quarters are old generation models.
It is worth noting, however, that these estimates carry a lot of uncertainty and some complications may obscure their accuracy. One particular risk is the potential discrepancy between machines' actual outputs and the output specified by the manufacturer, which often occurs when miners overclock their machines. Overclocking refers to the process of manually raising a microchip's clock speed, the rate at which it performs computational operations, to boost its hashing power. This requires drawing additional power and typically comes at a detriment to a machine's power efficiency. The additional power draw results in extra heat, requiring more effective cooling, and as a result overclocking is popular with miners operating immersion-cooled set-ups. Whilst this technology is becoming ever more popular - Riot Blockchain last year announced the construction of a 200 MW immersion-cooled facility - overclocked miners are unlikely to have a significant effect on the mean hashrate of miners on the network.
How do different models measure up?
Bitmain's Antminer S7 was released in 2015 and hashed at a, then, impressive 4.7 TH/s. Following its success, Bitmain's Antminer S9 range launched in 2016 with the 12.5 TH/s S9, culminating in the 18 TH/s S9 Hydro in 2018, before 2019 saw the final 13.5 TH/s S9k edition. These two models were, without doubt, the most influential of their respective eras.
Since their release, prevalent models have had outputs varying from 50 - 110 TH/s, although two new Bitmain models announced for 2022, the S19 XP and S19 Pro Hydro, are set to up the ante as they hash at 140 TH/s and 198 TH/s respectively (the latter facilitated by water-cooling technology). This wide range makes estimating the hashrate of the network's average non-S7 or S9 machine a difficult task.
How many miners were sold last year?
When we plot network hashrate over the past 4 years we see that, except for two significant event-driven departures, it has a consistent upwards trajectory. The slope of this trend is the rate of net hashrate addition to the network and, over the past 4 years, it has represented an increase of around 50 - 60 EH/s per year.
But, from a mining hardware perspective, what we're most interested in is the gross hashrate addition to the network. It is this quantity that represents the full contribution of all new machines and it serves as a proxy for manufacturer sales. As in other contexts, net means gross addition subtract gross loss. We need, as a result, to assume the rate of gross hashrate loss from the network. We need to account for the death of existing machines.
So, how do we estimate the quantity of hashrate lost from the network? By making some distributional assumptions about the lifetimes of machines, we can begin to estimate how many break irreparably each year and, thus, the gross hashrate loss from the network. Let's assume that the average S7 or S9 machine on the network has a remaining use of 1 - 2 years and that all other models have an average remaining use of 3 - 4 years. If we model machine lifetimes as being exponentially distributed, we arrive at an estimated gross hashrate loss of 25 - 30% per year.
Throughout 2021, this corresponded to around 35 - 50 EH/s worth of hardware breaking irreparably. In 2022, with a year-opening total hashrate in circulation of 200 EH/s, 50 - 70 EH/s will be lost.
That leaves us with gross hashrate added to the network. Subtracting the rate of gross hashrate loss, the figure derived above, from the rate of net hashrate change, the slope of the trend line, gives us a gross hashrate addition of 90 - 110 EH/s in 2021. This figure represents new miners added to the network. It is a proxy for manufacturer sales and, when translated into machines, it's 1,000,000 - 1,200,000 equivalent of 90 TH/s units sold last year. This is, approximately, the current size of the bitcoin mining hardware market measured in units sold - as illuminated by a top-down, network-centric analysis.
Departures from the trend
There are two notable departures from the past 4 years' trend. The first, occurring mostly over 2018 Q4, is a result of a year-long slide in the price of bitcoin. Bitcoin hit a high of just under $20k in the months leading up to 2018 but, thereafter, tumbled down to $4k as the year progressed. With it came hash price as mining rewards per TH per day dropped as low as $0.15, at a time when lower power efficiency meant the hash cost was around 2x what it is now. This prompted many miners to scale back loss-making operations and network hashrate decreased. Although since shutdowns were only temporary, hashrate was able to recover quickly as hash price began to pick back up again from 2019 Q2.
The second departure was less organic. In the summer of 2021, the Chinese Communist Party enacted a comprehensive ban on crypto mining activities within their borders. At a time when Chinese miners provided a significant portion of the network's hashrate, 44% in May 20216, this wiped off nearly 100 EH/s from the network in a matter of weeks. Only months after, at the start of 2022, had hashrate recovered enough to match pre-crackdown levels. However, due to the not insignificant cost of internationally resettling miners, many old-gen models were sent for early retirement. It remains to be seen how this event will impact the network's long-run trend.
How many bitcoin mining ASIC microchips were fabricated last year?
We can use the above estimate, the gross hashrate added to the network per year, to understand the industry's current production dynamics.
Bitcoin mining hardware employs specialist Application Specific Integrated Circuit microchips (ASICs) to power its computational work, with new generation machines typically incorporating around 300 of these ASICs. The level of production of the bitcoin mining hardware sector is best measured in the volume of these microchips fabricated.
There are currently only three semiconductor companies fabricating high-quality bitcoin mining ASICs: Taiwan Semiconductor Manufacturing Company (TSMC), Samsung, and Semiconductor Manufacturing International Corporation (SMIC). Of these firms, TSMC is currently dominant: Bitmain, who best guesses have as accounting for up to 75% of the mining hardware market, currently rely on TSMC as their exclusive ASIC provider.
Most high-quality contemporary bitcoin mining hardware uses ASICs fabricated with a 7nm process. Computer processing microchips, like ASICs, are made up of billions of transistors; 7nm refers to the width of these transistors. Generally speaking, the smaller the process width, the more powerful the microchip.
Let's assume, for simplicity, that all new hashrate added to the network in 2021 was generated by 7nm ASICs. The Antminer S19's 7nm TSMC N7 chips hash at 0.32 TH/s; the MinerVa MV7's 7nm SMIC-made chips hash at around 0.29 TH/s. A gross hashrate increase of 90 - 110 EH/s represents 300 - 370 M equivalent of such chips. This figure is an approximation for the volume of ASIC production across all foundries globally in 2021.
What does this mean for the bitcoin network in 2022?
Let's apply what we've learnt to project network hashrate. Given our estimate of ASIC production volumes in 2021, can we forecast production volumes for 2022? And what does this mean for network hashrate?
To understand the wider context, we must first examine the state of the semiconductor industry. 2021 saw a global microchip shortage as supply, hampered by COVID-19's impact on supply chains, struggled to meet rising demand. This led to a significant price increase for those able to source microchips and has curtailed the manufacturing of those not. Whilst each of the three major ASIC-manufacturing semiconductor companies has expansion plans, which for TSMC and Samsung include building new fabrication plants in the USA, it seems unlikely that their production capacities will change drastically over the next 12 months.
To semiconductor companies, bitcoin hardware manufacturers are lower priority customers than some of the other technological giants they're partnered with — the likes of Apple, Tesla and Nvidia — and ASIC supply is therefore unlikely to drastically increase in the foreseeable future. This leaves annual ASIC production volume where it is now, 300 - 370 M microchips per year.
If 7nm ASICs remain bitcoin mining's prominent microchip for 2022, assuming such chips output around 0.3 TH/s, we would expect to see a year-end network hashrate of 220-260 EH/s.
If, however, the industry begins transitioning to a narrower ASIC-fabrication process then this figure will be higher. This eventuality may soon be upon us. Bitmain's Antminer S19 XP, set to debut in July 2022, makes use of TSMC's newly-developed 5nm N5 ASIC. Whilst the exact performance specifications are currently unknown, this chip will likely hash at a rate 20-30% higher than its predecessor. If Bitmain can orchestrate the widespread manufacture and distribution of their newest model, we could see a year-end network hashrate of up to 290 EH/s. This scenario is perhaps more plausible.
All this assumes that the incumbent semiconductor manufacturers continue to dominate production. However, in January, Intel announced their first foray into the industry and is expected to unveil their new “Bonanza Mine” ASIC at the ISSCC, set to take place at the end of February. The ultra-low-voltage chip is expected to hash at 0.137 TH/s with a prospective industry-leading 18 J/TH power efficiency. It's not clear at the moment what level of supply Intel will be able to provide the market, though the semiconductor company has agreed on a minimum 4-year partnership with $3.3bn-valued cryptomining start-up GRIID, with the option for the US-based cryptomining start-up to secure up to 25% of Intel's entire production until 2025. It is also unclear when we can expect these ASICs to start contributing. Whilst it seems unlikely that supply will give a significant boost to network hashrate in 2022, if Intel can manufacture to the order of 100 M chips then expect to see an additional 10 - 15 EH/s.
Bitcoin hardware manufacturing is an often secretive but fascinating industry. Using the bitcoin blockchain itself, along with a few other facts and assumptions, we have been able to bypass some of the opacity and arrive at a top-down estimate for the number of machines in circulation, the current annual capacity for microchip production across the major ASIC foundries, and scenario-based forecasts for year-end hashrate. By taking things down a level and examining the production of mining hardware's constituent components, the microchips themselves, we've been able to bootstrap a comprehensive picture of the industry. Each of these findings helps shed an invaluable bit of light on a murky, expanding sector.