An analysis of the big question: “Is immersion cooling worth it?” with long-term profitability projections for Antminer X19s in immersion with Braiins OS+.
The day is June 18th, 2021. China is extending its provincial cryptocurrency mining ban in Xinjiang and Inner Mongolia to be nationwide, just as the ultra-profitable rainy season in hydropower-dense Sichuan province is getting going. Network hashrate—once practically a sure bet to exceed 200 EH/s by mid-year—now drops back below 100 EH in July. The miners who have hash online in the summer of ‘21 see their revenue temporarily double in bitcoin terms, from under 500 satoshis/TH/day to nearly 1000 sats/TH/day.
Fast forward to late May, 2022. Resilient miners in China have gotten smarter about their network infrastructure and security so that they can continue hashing away, although not nearly at the same scale as before. While some employees of a leading Chinese hardware manufacturer have suggested in friendly conversations that as much as 30-40% of network hashrate remains in China today, this author (with almost no contacts in China) believes it’s more likely below 20%.
Meanwhile, total network hashrate is pushing to new all-time highs. The 30-day moving average for hashrate has been sitting above 200 EH/s for a couple of weeks now, and estimated real-time network hashrate is 208 EH/s at the time of writing. As for where the new hashrate is coming online, there is simply no debate. North America has picked up China’s slack and then some, with Texas becoming the clear global frontrunner in terms of hashrate/m2 (even though Texas is huuuuuge).
With this hashrate migration to North America and Texas in particular, we are in for an interesting summer. West Texas, the home of much intermittent solar and wind energy, is about to get hot. Like, swelteringly hot.
This presents a problem for bitcoin miners. When it’s really hot outside, mining hardware runs less efficiently and with a greater risk of failure. The consequences can cause downtime and loss of revenue due to machines automatically turning off once temperature readings reach dangerous levels. Additionally, it can lead to fried hashboards and other permanent hardware damage if the machines don’t shut down in time. (Braiins OS+ Dynamic Power Scaling feature fixes this.)
To make matters worse, the most popular ASIC hardware family today is the Antminer X19, which is extremely sensitive to heat relative to older miners like the S9 and Whatsminer M20S. As we’ve described in our research summary, the power consumption of Antminer X19 models can increase by 40%+ with higher temperatures even as frequencies are constant, meaning that the J/TH efficiency of the machines suffers significantly.
It’s no surprise then, that public miners such as Riot and Argo are building exclusively immersion cooling infrastructure for their new miners coming online in Texas. Immersion cooling alleviates the majority of the temperature impact on mining operations. Rather than worrying about your mining fleet’s uptime, power consumption, and lifespan, you can respond to temperature rises by increasing pump speeds and running dry coolers / cooling towers harder, keeping your total hashrate and power consumption much stabler and decreasing operational risks. For well-capitalized miners operating in hot climates, it just makes sense.
But what about for the pleb miners with a machine or two or five at home? What about the miners with 1-6 MW operations in Paraguay or Mexico? Hell, what about the miners in Wyoming, Montana, and the Dakotas where it’s very cold for a good portion of the year but those few summer months can see temperatures exceed 100F (38) on rare occasions?
Does it make sense for any of those miners to invest in immersion cooling? Well, let’s find out.
Before getting into the economic analysis, it’s important to understand the advantages of immersion compared to air cooling that help justify its higher up-front cost. To briefly summarize, immersion cooling offers the following benefits:
All of these benefits make immersion cooling superior to air cooling regardless of the climate that the miner is operating in. However, it comes at a much higher up-front cost, so there’s still a question of whether or not immersion cooling is worth it. While that will depend largely on the local climate of the operation, this article will outline the structure for making this determination with all of the not-so-small details to consider with it.
As is the case with all future profitability calculations we do for bitcoin miners, we’re going to have to make a lot of assumptions and generalizations here in order to get anywhere.
Honing in on a narrow price range for building mining infrastructure just isn’t possible. It depends on all sorts of factors that change over time and across different locations, as well as with the size of the operation being built. For example, labor in Paraguay and Mexico is relatively cheap compared to the US and Canada. On the other hand, many of the parts required for infrastructure would need to be shipped internationally, which may also include extra import costs. And modular, mobile infrastructure (i.e. mining containers) will generally cost more per MegaWatt than large static facilities that can host 10+ MW worth of miners.
So, let’s simplify matters and just lay out some assumptions. Building air cooled infrastructure can cost anywhere from $150-400k/MW, depending on everything described above. Some portion of that cost is for cooling, including wet curtains and industrial-sized intake and exhaust fans, as well as insulative material to separate hot and cold aisles. All together, those components will rarely account for more than 10% of total infrastructure CapEx, while the more common case is probably at or below 5%. Meanwhile, the lion’s share of the costs will be from labor, materials, and electrical equipment and wiring.
Immersion mining infrastructure will still have nearly all the same costs as air cooled, minus about 5% ($7.5k-20k/MW) for the cooling components that are no longer needed. However, immersion will add all sorts of new and costly components to the infrastructure:
After analyzing dozens of different immersion systems varying in size from small DIY tanks with 2-4 mining machines to industrial-scale facilities, we’ve found that the extra CapEx for building immersion infrastructure is almost identical to the original air cooling CapEx range, $150k-350k per MegaWatt, not including shipping costs. This means that the total cost for immersion infrastructure should be somewhere in the wide range of $280k-730k/MW, although the upper portions of both the air cooled and immersion ranges are generally for modular containers which wouldn’t be combined with each other. More realistically, then, let’s say that the total immersion cost should be somewhere in the $280-600k range.
Now, all that’s left to do is answer the question of whether that extra CapEx is worth it for all the benefits immersion provides.
Note before we begin: our mining profitability calculator is undergoing its biggest-ever update at the time of writing. The update will include more helpful profitability calculations and input options, so this article will be republished in the future to use the improved version of the calculator and get even more Mining Insights.
Finally, we’ve made it to the fun part. Let’s lay out the rest of our assumptions and crunch the numbers.
First, some arbitrary inputs to narrow down the scope:
One interesting way to put the infrastructure CapEx into perspective is by comparing it to the hardware CapEx. If we assume stock power consumption of 3068W per S19j Pro and a 5% buffer in total power consumption, that means that we will have 325 S19j Pro’s per MW. At a unit cost of $8500, the hardware CapEx is $2.76MM/MW—approximately 10x the air cooling infra cost and 5.5x the immersion cooling infra cost. Just food for thought.
With these assumptions, we’ll analyze the bullish and bearish cases for each of 5 different scenarios:
A side note: the bullish case will result in less BTC mined over a long (2+ year) time period because large BTC price increases will incentivize more hashrate to come online, causing difficulty to increase faster.
To fill 9.5 MW at the stock power consumption of 3068 W, we can purchase 3,096 S19j Pro’s for a total hardware CapEx of $26.32MM. This brings our total CapEx for the 10 MW air-cooled operation to $28.82MM (720.5 BTC). Our total hashrate is 322 PH/s and our pool fee is 1.2%.
Both the hypothetical bullish and bearish cases for an air cooled operation look good in fiat terms. Even the moderately bearish case with difficulty increasing faster than price over 4 years still has an annual IRR of 210%. This will be our baseline for comparing the projections with different immersion solutions.
Let’s not move on too fast though. This is an economic analysis of immersion cooling for BITCOIN miners, not FIAT miners. We’re working on a BTC-centric version of our profitability calculator, but in the meantime we’ll make do with some excel charts to see the mining profitability in bitcoin terms.
Uh-oh. While we do end up mining more BTC than our initial investment (CapEx), after subtracting the OpEx each month we never break even with our mining profit. This means that the only way this operation could out-perform a simple Buy & Hold BTC Strategy is if it’s financed (e.g. with BTC-collateralized loans) such that the miner doesn’t need to sell BTC to cover the OpEx or if the BTC-denominated resale value of the hardware at the end of the analyzed period can make up the difference.
We add $2MM to the total CapEx, giving us $30.82MM (770.5 BTC). With the fans removed, there’s about 5% more power to go towards actually hashing, making the hashrate per machine approximately 109 TH/s. This brings the total hashrate to 338.1 PH/s.
Our first two cases had similar performances in terms of fiat CapEx break even periods, with the immersion cooled operation able to mine about 5% more BTC thanks to the efficiency improvements of running without fans. With the 100% HODL Ratio on BTC profits, the immersion miner with more BTC ends up being more profitable in the end and the extra up-front investment pays off in fiat terms, but they break even on the CapEx in bitcoin terms more slowly and ultimately have smaller margins between their BTC-denominated CapEx and Total BTC Mined. In other words, the operation with air cooling is doing better in BTC terms but worse in fiat terms relative to the miner with immersion cooling when both are running stock firmware at the stock power consumption.
Next, we’ll analyze the same exact case as above but switching stock firmware for Braiins OS+ running at the stock power consumption level, 3068 W. The average hashrate at this power input for an S19j Pro running Braiins OS+ in immersion is approximately 114.5 TH/s, bringing our total hashrate for 3,096 devices to 354.5 PH/s. For all remaining analysis, there is an additional 2.5% dev fee added to Other Fees, but the 1.2% pool fee is changed to 0% which is the pool fee on Braiins Pool (formerly Slush Pool) for Braiins OS+ users.
The biggest thing to note is that running Braiins OS+ instead of the stock firmware results in a significantly lower marginal cost to mine 1 BTC and significantly greater BTC production.
The immersion miner running stock firmware is not taking full advantage of the infrastructure that they spent $450k/MW to build, as they are not using the immersion systems to overclock their miners at all. Let’s see what happens when they run Braiins OS+ on their S19j Pro’s at 4000 W with autotuning. If the PSU’s are immersed along with the miners, the stock PSU’s can safely run in this range.
Based on our data, a typical S19j Pro in immersion can hash at around 136 TH/s with a power consumption target of 4000 W, although this is fairly conservative and we’ve seen some hashing above 140 TH/s in ideal conditions. (Disclaimer: results vary for each hardware device and depending on the temperatures maintained by your immersion system. Always measure power consumption at the wall and hashrate on your pool account.)
Since we are planning to consume 4000 W per machine, we won’t buy as many miners to fill the available 9.5 MW of capacity. Rather than 3,096, we will only need to buy 2,375 S19j Pro’s. This brings our hardware CapEx to $20.19MM and our total CapEx to $24.69MM (617.25 BTC). Total CapEx is actually far lower than an equivalently sized air cooled operation because, even though the immersion infrastructure costs more, it enables us to reduce the more significant $/MW cost, our hardware. At 136 TH/s per machine, our total hashrate is 323 PH/s.
Note: we are adding a 2.5% fee on Other Fees for the Braiins OS+ dev fee and a 0% pool fee assuming the miner is using Braiins Pool, although Braiins OS+ works with any pool.
We don’t need to look beyond the quick eye test to see that this is by far the best performing mining operation so far, both in the bearish and bullish cases. They are breaking even sooner in fiat terms and BTC terms. And despite running 750 W higher than the machines in immersion with stock firmware, the average cost to mine 1 BTC is nearly the same. (Note: this is the average marginal cost to mine, not including hardware depreciation.)
Running Braiins OS+ at the stock power consumption level with more mining machines in the same power capacity still wins in terms of the total production. But if the goal is to maximize returns on the initial CapEx, then overclocking at 4000 W is the obvious winner in fiat and BTC terms.
When you can safely overclock to much higher power limits via immersion cooling, you’ll find that there is A LOT of hidden potential in Antminer X19’s. In fact, they can run at 6000, 7000, or even 8000 W per unit if you want to really push them. However, the stock PSU’s that come with these miners cannot run nearly that high. To push the limits of the ASICs, you’ll need PDU’s and PSU’s that can supply more than 4000 W per machine.
Custom PSU’s can vary a lot in price and are subject to some of the same supply chain issues as pretty much everything else in mining these days. That said, a typical price for custom 6 kW PSU’s in recent weeks is about $500/each, so we can just add this to our S19j Pro cost and bring it to $9000/machine.
Running at 5500 W each, we need 1727 S19j Pro’s with custom PSU’s to fill our 9.5 MW available capacity. This gives us a hardware CapEx of $15.54MM and a total CapEx of $20.04MM (501 BTC). If we assume a hashrate of 160 TH/s at 5500 W (34.4 J/TH), this gives us a total hashrate of 276.3 PH/s. (Note: it is possible to get higher hashrate than 160 TH/s at 5500 W with Braiins OS+, but we are being conservative in these estimates as was the case with the 4000 W projections.)
Note that due to the worse efficiency of overclocking at this higher power limit, mining becomes unprofitable after Month 42, so this analysis stops then. In reality, the miners in this case would simply change to more efficient settings after the halving, such as running at 4000 W, in order to continue mining profitably for the entire 48 months.
Although this is our lowest CapEx option, it does not outperform the scenario where we simply use the stock PSU and run Braiins OS+ at 4000 W after subtracting out operating expenses. The takeaway is that, although overclocking these machines to 160+ TH/s seems great, the sacrifices in efficiency may not be worth it.
In the summary table above, you can see that the scenario which performs very well across all the BTC and USD profitability metrics is Immersion with Braiins OS+ at 4000 W consumption. While the Immersion with Braiins OS+ at 3068 scenario is the most efficient, producing the most BTC per Watt, it has a significantly higher CapEx bill. As for the Immersion with Custom PSU’s and Braiins OS+ at 5500 W scenario, it does well in the BTC Mined - BTC CapEx metric, but it performs the worst in everything else due to the fact that the highly overclocked machines are running several J/TH worse than the stock efficiency levels.
Before dismissing the custom PSU’s and high overclocking idea altogether, one more detail worth noting is that overclocking to this high level does maximize the BTC mined per machine, although it’s mining the least BTC per Watt consumed. In the unlikely event that we see another huge difficulty drop in the future like what occurred with China’s mining ban in 2021, having the option to overclock those machines and take advantage of the much lower difficulty during that time would be extremely valuable. Likewise, if there is a crunch for bitcoin mining hardware in the future where miners have extra power capacity and are waiting on their machines to ship (the opposite of the current situation where the bottleneck is on rack / tank space), using up that extra available power by overclocking the machines you already have would also make sense to maximize the BTC mined in that time.
It would also be interesting to compare this low CapEx and relatively bad efficiency option with an even lower CapEx and worse efficiency option: overclocking Antminer X17’s in immersion. The X17 generation of hardware is terrible, but most of its problems are with heat dissipation. Those problems can be solved with immersion, making it possible to then overclock X17’s to X19 level hashrate and keep them under 50 J/TH. This article is already too long, but if you’re building immersion and can’t easily purchase a bunch of new-gen hardware, check out this option.
You probably noticed that none of these hypothetical mining operations have a final cumulative bitcoin profit after 4 years which is greater than their CapEx in bitcoin terms. In other words, even though they perform excellently in fiat terms, they do not outperform the alternative of using the initial CapEx amount and the monthly OpEx amounts to simply buy and hold bitcoin.
Some thoughts on this:
There are some nuances that would be too complex to include in the mining profitability projections, but which miners should be aware of if they’re considering immersion.
First of all, the analysis assumes constant hashrate and power consumption year-round. This may roughly be the case in places like Iceland and Siberia that never get too hot, but it’s not realistic for places like Texas and Paraguay where heat in the summer would result in increasing power consumption and worse efficiency for air-cooled miners. Since immersion would enable more stable operating conditions year-round, regardless of weather conditions, immersion should compare even more favorably than what we’ve shown in the projections above for miners who operate in hotter climates.
For both air cooled and immersion operations, miners can use saved autotuning profiles on Braiins OS+ and sophisticated power management strategies to optimize their operations on a much more granular basis. For example, they can optimize for efficiency during the hot hours of the summer, optimize for maximizing hashrate during the coldest hours of the winter, and so on. And as difficulty continues to go up over time and their profit margins shrink, they can respond by shifting from a general strategy of maximizing hashrate to a more efficiency-focused strategy.
Also worth mentioning is the impact of immersion on hardware resale value. Once hardware has been running in immersion, it probably shouldn’t ever be run in air. For one thing, the thermal paste used in hardware including Antminer X19’s will dissolve in most dielectric coolants, resulting in poorer thermal conductivity which would make the machine dangerous to operate in air later on.
As a result, the future of the secondary hardware market will likely be segregated between machines that have been used in immersion and those that have not (but still can be in the future). It’s hard to say whether the immersion machines will be worth more or less relative to the non-immersion machines. On the one hand, immersion should improve the lifespan and decrease wear and tear from months or years of operating, meaning that the ASICs are still in great condition when they hit the market. On the other hand, the buyers for those machines can only be miners who have available immersion capacity, and it’s questionable whether or not it will make sense to put older hardware into immersion considering the higher infrastructure CapEx. Therefore, it seems possible that hardware running in immersion does have a longer operating life, but is difficult to sell on secondary markets when it's 4+ years into that lifespan and effectively “old-gen” hardware.
Another out-of-scope factor to consider is that immersion infrastructure won’t become useless when the current new-gen machines are no longer profitable or stop working. With proper maintenance, it should be the case that immersion systems can be used for multiple generations of hardware across 10+ years. Perhaps 3-5 years from now, you’ll swap out your S19’s or M30S’s for the latest gen hardware and you’ll be able to operate it more efficiently and safely overclock it from the beginning because you already have the immersion systems ready to go. If the economics on immersion look good with a single generation of hardware, it’s up only for multiple generations.
Lastly, immersion systems are better at efficiently (and quietly) capturing the heat output of the miners and transporting it elsewhere. The majority of the time, that 'elsewhere' is a cooling tower or dry cooler, but sometimes it might be a swimming pool, green house, or some other place that can use the heat from miners. If miners can find a customer for the heat (even if it’s themselves), they are effectively dropping their electricity price by making the electricity count twice. Immersion helps enable more use cases for doing just that.
Long-term, the clear winning strategy based on this introductory analysis is moderate overclocking near the limits of the stock PSU in immersion cooling with autotuning firmware like Braiins OS+, which enables miners to still achieve the stock efficiency while producing much higher hashrate. This provides the best balance of total production (i.e. mining as much BTC as possible), low CapEx, and high efficiency (i.e. stacking the most sats per Watt of energy consumed).
Although the price tag of immersion seems extremely high for many miners, it is relatively insignificant compared to the cost of new-gen ASIC hardware. If you don’t have easy access to capital, or you are building a smaller-scale operation and using older, less efficient mining machines, then immersion likely won’t make sense. But for well-capitalized miners building industrial-scale operations, particularly in places with hot summers like Texas and the rest of the southern US, then immersion is probably worth it. And of course, whether you overclock or not, you’re missing out on extra BTC if you don’t use Braiins OS+ autotuning firmware to maximize ASIC efficiency at any power consumption level.
Bitcoin mining software company: Braiins Pool, Braiins OS+ & Stratum V2.
By miners, for miners.