It’s been a year since the Ethereum merge dramatically changed the landscape of GPU mining. Unfortunately, our hopes for a last-minute rescue were dashed, and the merge went through as planned. There’s really no need for us to tell you how bad the situation is since you’ve all lived through it and the results speak for themselves. But we believe many miners will find value in examining the post-merge landscape of mining profitability and GPU values and exploring the possibilities moving forward.
Current GPU mining profitability
In our previous article, we outlined the range of possibilities for GPU mining profitability post-merge. The potential outcomes ranged from terrible to utterly apocalyptic. So how did it turn out? Superficially, not as bad as pre-merge mathematics suggested – but the devil is in the details.
The base case suggested that immediately post-merge, the average GPU (roughly equivalent to an RTX 3080) would be pulling in $0.07/day. Even though revenue never really tanked to the single digits, the reason for that is not because our math was incorrect, but because most miners simply never bothered to mine anything else and just threw in the towel immediately. Post-merge it took mere minutes for revenue per GPU on other coins to adjust to the low to mid double-digit cents and there hasn’t been much action in either direction since. We don’t think it’s instructive to dive deep into the specific returns of any coin post-merge. We’ve just been cycling through the flavor of the day with no real progress on overall returns. Profit-switching services such as NiceHash ensure that hashrate flows immediately to any coin that dares to break out and beats it back down.
Many miners seem to be pinning their hopes on the next bull market in crypto to restore profitability. Putting aside that a future bull market in crypto is not guaranteed, on the surface that is certainly a logical belief – but idled hardware throws a wrench into the mix. Pre-merge, when returns were solid and virtually all hardware that could mine was mining, a 50% increase in the price of crypto translated directly into a 50% increase in revenue in the short term. Of course, these higher returns would prompt miners to deploy more hashrate, but in the past, this process would take a considerable amount of time. But nowadays, what would have taken months will now take just minutes to just power idle rigs back on.
With the number of idled miners likely being several times as large as active miners, a 5X increase in the price of crypto will not translate into a 5X increase in mining revenue per GPU. It will increase the overall revenue flowing to miners, but the idled GPUs will pile back on to soak it up to suppress revenue per GPU. How can we be so sure there are huge amounts of idle rigs? Because all GPUs used for mining were not created alike. The mining-specific GPUs such as NVidia P10X, CMP, HX, and many AMD Radeon models completely lack display ports or have minimal VRAM and bandwidth constraints, which makes them impossible to repurpose for gamers or AI. There are simply no willing buyers for these GPUs at the industrial scale where they were deployed. The only choices these miners have are to scrap them for a few dollars or sit tight and wait for an opportunity to mine with them. We personally know of many large miners with tens to hundreds of thousands of units that are sitting tight, and zero of them have taken that final step to recycle them thus far to our knowledge. Those GPUs are out there waiting for mining returns to recover, and they’re going to stand in the way of anyone making a decent profit per GPU for many years.
Without an accurate way to measure idle GPUs, it’s impossible to be more specific than “enormous.” There may be some uncertainty about exactly how much idle hashrate is out there, but there shouldn’t be any doubt that it exists and is huge. Just because hard numbers can’t be placed on a variable doesn’t mean you can entirely ignore it. We can still know for sure that revenue per card in the next bull market will be highly suppressed by idled miners, even if we can’t be certain by exactly how much.
Beyond that, there isn’t much else to say about the state of GPU mining. If we’re being honest, there aren’t any GPU mineable coins that appear poised to break into the mainstream to the extent required to bring the glory days back. Post-merge, ETC has lost value relative to ETH and is flooded by ASICs. ETHW was a total flop. Everything else you can mine is a minor project with minimal revenue. And so on and so on. We wish we could be more hopeful than that, but between the lack of a significant project with a $50B+ market cap like ETH to drive revenue up and all the idle hashpower ready to drive revenue down, the dark ages are here to stay for a while.
To put it bluntly – if you still have GPUs and aren’t ready to take them to the local e-waste recycler, you realistically have two options moving forward – repurpose them for something other than mining such as AI or sell them.
Repurposing GPU mining rigs
One of the key advantages of GPUs is their flexibility, and there are technically many other alternatives to mining with them – Distributed Compute, Rendering, AI/ML, and who knows what else people will come up with? Let’s look at them one by one, and see where they can fit in.
The most obvious use case is to rent out your CPU and GPU power to others to do some useful calculations instead of mindlessly hashing away. The most popular project to try to bring this to miners is the Golem Network. And it was an utter failure. Based on data that I could scrape from sites tracking actual usage of the network, it’s currently outputting a whopping $20 revenue per day for the entire network. It’s a neat idea and the supply side is there, but the demand side isn’t. There are a few newer, more centralized options popping up, but they suffer from the same demand-side scaling problem.
The next best hope is distributed 3D rendering, with the Render Network and the RNDR token leading the charge. This project at least has some clout behind it, but does it show miners the money? Hard to say. It’s next to impossible to find public data about how much usage it has and income it produces. You can’t just sync up to the network to find out. There’s a waitlist to join, and many people have been waiting for over a year to get in. Common sense dictates that the demand isn’t there for this use case yet either. Either way, it doesn’t matter – you’re not invited to the party anyway.
With those false starts out of the way, let’s move on to the elephant in the room. We’re huge proponents of AI, and we have total conviction that generative AI will continue to deliver on its promise. There are two places where mining GPUs could potentially fit in – training (creating the AI models) and inference (running the AI models.)
This is a use case that should be taken seriously – but it comes with a few very big caveats. Unlike the previous two examples, there are viable markets for renting out your rigs for AI such as Vast.ai, among other startups. There are also very realistic prospects for growth here. But you can’t just bring your typical GPU rig to the party. If you have powerful Nvidia GPUs with 12GB+ VRAM, that’s a good start. But AI training requires a lot more than raw GPU power. You need a fast CPU with lots of cores, tons of ECC memory, high bandwidth data links, lots of fast SSD storage, and a fiber optic internet connection doesn’t hurt. In other words – you need a real server to host your GPUs, not a mining enclosure, and certainly not a rickety open-air rig. And not a cheap server either if you hope to compete with the guys who have been at this for a long time. That’s a steep hill to climb…but it’s not an impossible hill to climb.
And we know many have climbed it – but we are seeing signs of the potential tapping out. As of this writing, there are now waitlists on AI rig rental sites, and even once your rig is approved it’s still nowhere near the simplicity of mining. The demand side of the market is increasingly demanding the most powerful GPUs available, with waning interest in using lower-end cards with less VRAM for training. While there’s always the possibility this changes with advances in training and distribution of the workload, it appears to increasingly be a use case that you build specifically for, rather than repurpose existing hardware. If you have a farm full of 24GB RTX 3090s or A5000s you have a much more viable starting point because VRAM is king for training, but this market is simply not accessible for most miners.
Compared to training, inference has significantly lower requirements, making it potentially much more suitable for repurposing mining rigs. It was never going to be possible to train something like ChatGPT with GTX 1070s – the truly groundbreaking advancements really do require tens of thousands of 80GB H100s. But once trained, running a model is not necessarily off the table. Progress in alternative models that work on lower-end hardware is progressing rapidly. Inference generally doesn’t require nearly as much horsepower in the rig – but VRAM is still a significant limitation. At this stage, LLMs that can run on typical mining GPUs are more experimental than useful. Beyond LLMs, generative Image AI such as stable diffusion are producing viable images for real-world content on mining-class GPUs. As with LLMs it’s still not as high quality as bigger models such as MidJourney, but it’s getting closer every day.
While inference seems like it should be promising, what we have yet to see is a viable market form around distributed inference. Unlike mining or AI training, which utilize a GPU 100% 24/7/365, users of generative AI are instead making requests that require a short burst of activity while the content is generated. Due to this a small cluster of GPUs can serve generative AI for a very large quantity of users. And with the output currently being limited to simple images and text, this use case can be efficiently served by data centers so inexpensively that it is given away for free. If a user can generate a high-quality essay or image for free on ChatGPT or Bing, they’re simply not going to be interested in paying even a penny for a lower-quality essay. They might be interested in paying a small amount for even higher quality output as with MidJourney and ChatGPT Plus, but that’s not something that mining-class GPUs can generate. And to whatever extent running models locally becomes more viable, it just makes it easier for users to run them on their own computer, so it’s not clear why they would need to use a distributed network anyway.
As it stands right now, while there are a few use cases for which typical mining GPUs are technically viable, the demand side is minimal to non-existent. Doubling down and upgrading your farm to support AI training appears to be the only viable path forward to repurpose existing mining farms and it’s only truly suitable for farms built exclusively from 16GB+ GPUs. It also requires significantly more expertise and capital and is far from a sure thing that there will be a positive ROI on your investment. It’s ultimately still a tiny niche and far less accessible compared to the magnificence of the ETH network at its prime, but the barrier to entry can work in favor of those who take the chance on it by keeping the growth rate and competition in check.
Beyond that, we do not believe that there is a plausible way forward for most mining GPUs. Alternatives have thus far not come to fruition, and it looks highly unlikely that the next bull run in crypto (if it even happens) will be able to drive up per-GPU mining returns. The only realistic choice that remains for anyone concerned about recovering value going forward is when and how to sell your GPUs. It might not be a terrible idea to hold on to the rigs without the GPUs, as those parts have been so thoroughly devalued already that there isn’t much room left for them to fall, and in the off chance that mining comes back in a few years, it’s quite possible you can just plug in some RTX 6080s and fire it back up. The market for current GPUs on the other hand continues to be quite dynamic and while they’ve lost tremendous amounts of value compared to the peak, there’s still a very healthy market for them.
The GPU Market Post Merge
Even though mining rigs are more than just GPUs, we’re only going to focus on GPUs. This is probably going to hurt to hear, but we need to say it – as a complete unit, your mining rig is now worth less than the sum of its parts. It doesn’t matter how good it is at mining. No one wants to buy a machine that burns money, and that’s all a mining rig can do right now. The individual parts will have value to gamers, but those parts being bundled together in a mining rig is a problem that needs to be solved, not a value-add that commands a premium price. Unfortunately, most of the other parts typically used in mining rigs aside from the GPUs have been devalued so much that they’re not even worth discussing.
Before we dive into the price charts, there are a few key differences between the pre-merge and the post-merge market for GPUs that we should highlight to contextualize the data.
Crypto prices are now completely irrelevant to the GPU market. Many miners have told us that they’re holding on to their hardware because they don’t want to “sell the bottom.” The problem is that they’re bringing a crypto market mentality to a market with completely different dynamics. For this to be “the bottom” of the GPU market, there must be a reason to believe that supply won’t be able to keep up with demand. From a mining perspective, we’ve already gone over why there is still a glut of supply unlike any previous cycle, so it is unlikely that crypto prices will be able to drive up GPU prices the way it has in the past. GPUs are not collector’s items that age like fine wine. They just bleed value as they increasingly become obsolete over time as newer models are released. The real “bottom” for most GPUs is about $3, which is roughly the value of raw materials for scrap, and they all inch closer to that bottom every day.
Gamers only want exactly one GPU. Even a small mining rig is six times more GPU than any gamer wants. A gamer will not buy your entire rig or farm. If you choose to sell, realistically you have two choices right now – break it all down yourself and sell the GPUs individually to many different gamers or liquidate it in bulk to a company that is going to do that for you. If you try and sell the rig or farm whole, especially for a premium price, chances are you will be waiting a very long time before anyone makes a serious offer for it.
The market for used GPUs is very limited. The most popular cards like the RTX 3060 sell 100-200 units per week in the entirety of US eBay, which is the largest online market for used GPUs by far. While eBay isn’t the only place to sell, not even a thousand eBays can absorb the millions of gaming GPUs that were used for mining without prices crashing. In 2021, anyone could easily find a miner willing to buy 1000 RTX 3060s at once at market price. In 2023, it would take months to move that many units one by one, while the market price continually decreases over that time. Just because you saw a 3060 sell for $250 yesterday doesn’t mean you can sell 1000 for $250K today, especially when the other guy is trying to sell his 1000 3060s at the same time. Supply still dramatically outstrips demand.
No one cares how much money you lost. Especially gamers. Miners ruined their hobby, and gamers still hold a grudge. Gamers think miners are greedy and evil and run their hardware into the ground while killing the planet. If they catch even a whiff of the fact that your GPU was previously used for mining, most gamers won’t take a second look at it. And while you may have spent $2000 for that RTX 3080 at the peak of the market, no one is going to pay you more for your GPU out of pity for your losses. Would you pay someone more for their BTC just because they bought at the peak? Of course not. It’s a tough pill to swallow but markets just aren’t concerned with anyone’s emotional baggage. Refusing to sell your GPUs because you don’t want to lock in your losses is applying the sunk cost fallacy to the highest degree. The loss in value has already happened, whether you are ready to accept it or not. The longer it takes to move past that mental barrier, the more value you will lose moving forward.
The current state of the GPU market
With that in mind, let’s take a look at some data:
Nvidia RTX 3070 – sales for the last 24 months
This chart shows the average selling price and quantity sold of used RTX 3070s on eBay over the past 24 months. Sales volume was low pre-merge at first because there simply weren’t many people selling used 3070s yet. Miners that had them were holding on, and the price for them was sky-high at $1200/ea. As the bear market destroyed returns, the average selling price rapidly decreased to roughly $400 by merge day as miners stopped buying and started selling. Immediately post-merge, cards began to flood the market.
Nvidia RTX 3070 – sales for the last 12 months
Zooming in to the 1yr chart (which begins shortly after the merge), despite the significantly increased supply, prices didn’t suddenly crash. Why not? Two reasons: Pent-up demand from gamers, and the holidays. Similar to how the pent-up supply of idle hashrate will suppress mining returns, there were years of pent-up demand from gamers who were effectively locked out of the market for affordable GPUs by miners willing to pay obscene amounts for them. On top of that, there is always a bump in demand starting in October as the biggest games of the year release and Santa starts wrapping gifts. These two factors temporarily kept the price afloat, but once the new year arrived, the market fell apart again, with the average price dropping from $400 in January to about $270 by July. This represents a roughly 80% drop from the peak in 2021. From there, prices appear to have leveled off for the time being.
One other thing to point out – there were many large drops and spikes in the price of crypto throughout this period. If the price of crypto still had any influence on the price of GPUs, we would see it reflected in these charts. In fact, the price of ETH nearly doubled between Jan and May – all the while prices continued to crash. RVN had a few good spikes too, and you’d never be able to tell. GPU prices just don’t care about crypto anymore.
Other cards had it much worse. Take the 5700 XT for example – a card highly favored by miners, but less desirable to gamers:
AMD RX 5700 XT – 24 mos
While following the same general pattern as the RTX 3070, peak to trough it went from $1000 down to $120 – nearly a 90% decline in price.
Some cards faired a little better. Although the RTX 3090 had as rough of a fall pre-merge as any other card, post-merge the decline has been much less dramatic:
Nvidia RTX 3090 – 12 mos
Unlike the 5700 XT, the RTX 3090 has some utility to the AI market with its 24GB VRAM, and although still declining, the price hasn’t moved much in the last few months.
And to put an even finer point on the influence of AI in this market, the only cards that have increased in price over the last year are professional cards such as the A4000, which are directly targeted at AI:
Nvidia A4000 – 12 mos
Overall, holding on to GPUs and waiting for prices to increase has been a losing strategy. While most of the damage has already been done, the only GPUs showing any sign of increasing in price are the professional cards. We’re on the cusp of the holidays potentially keeping prices stable until January, but there’s every reason to believe that prices resume dropping in the new year.
While we wish we had more good news to share, but there doesn’t appear to be anything on the horizon to put most GPU miners back to work doing something else, and there is a horde of idled GPUs out there ready to soak up any mining revenue that returns. But that being said – the post-merge situation ultimately turned out to be far less catastrophic for the price of GPUs than we expected. We expected an unprecedented flood of GPUs to drag prices down to unthinkable lows, and that didn’t happen. An 80-90% drop in value from the peak is substantial, but the fact remains that there are still many millions of gaming GPUs sitting around doing nothing, and the supply/demand dynamics could still push prices down much further. But many miners either aren’t giving up hope that mining revenue will recover, or they’ve simply not bothered to sell them, and that supply isn’t hitting the market at the rate we thought it would. It’s not good news, but it could have been worse, and this story is far from over.
If you’re a small miner with a rig or two and you don’t have much financially at stake, you can continue to mine as a hobby or tinker with AI, and there isn’t much to lose that hasn’t already been lost. If you decide to sell, you have plenty of options. DIY is viable at this scale on eBay or Facebook Marketplace, but if you don’t want to deal with that, there are companies that will purchase even small quantities of hardware and make the process super easy if you’re willing to take enough of a haircut to leave them some profit.
For the medium-sized miners, with a basement or small warehouse, there is likely to be considerably more at stake. Especially for those who must maintain a separate facility, not only is the hardware bleeding value, but the fixed overhead costs add an extra burden on top. At this scale, DIY can still be viable, but it tends to be a much more arduous process than most miners expect going into it. Parting out the farm entirely will mean hundreds of individual transactions – you will come away with more in the end, but you’ll have to take on an extra part-time job for a few weeks and there will be significant expenses incurred that add up, and take a much bigger chunk out of the payout than people expect. For those without the time or inclination to deal with any of that, we can make that problem go away.
For the truly large miners, with hundreds to thousands of GPUs, DIY is not a particularly realistic option. Unfortunately for those who built farms entirely with mining-only GPUs like P102s that can’t be re-used by gamers, there really isn’t anyone in the world interested in buying them, and we’re no exception. For miners with large quantities of gaming GPUs, there is still a healthy market to sell them at scale over time, but it’s still a very different market than 2021 where vast quantities of GPUs could be sold with ease on the open market. They’re worth far less than they were just a year ago, but they’ll also be worth even less as time goes on. Given the time and expense required to sell large quantities of GPUs one by one, the haircut required by liquidators will be even steeper, but the quick capital can be readily re-deployed to utilize existing facilities with a new project such as AI or ASIC mining, which may ultimately have a positive ROI.
If you have more than a handful of GPUs and/or you’re just looking for a quick, safe, and easy way to turn hardware into cash, give us a shout.
If you want to be notified when we write about the GPU market, please feel free to subscribe to our blog below. We don’t post frequently, but when we do, we try to make it count.