In the 2020s, artificial intelligence (AI) has been the only buzzword to make more noise in investor circles than blockchain. Unsurprisingly, publicly traded bitcoin miners are pivoting toward the sector. But as I’ve covered before, the pivot is not easy, and there are multiple routes a bitcoin miner can take to stand up a high-performance compute (HPC) AI operation.
Core Scientific and Bit Digital have jumped to the front of the pack in the bitcoin miner AI race. Core Scientific is leading because of its 500 MW, multi-billion dollar deal flow with CoreWeave. On the other hand, Bit Digital has delivered the biggest margins of any bitcoin miner dabbling in AI.
They may have early leads, but the next two years will indicate whether they can remain ahead of the pack with their differing strategies.
KEY BACKGROUND
For most bitcoin miners entering the AI market, leasing an AI-grade data center is the easiest route at the outset. Most bitcoin miners have bootstrapped their AI operations this way, either leasing out the GPUs in these centers, hosting a third party’s computers or hosting their equipment.
So far, the scale of these rented operations is small, and they are not generating much revenue for these miners yet. According to SEC filings, the five public bitcoin miners with active HPC business lines earned $40.3 million in revenue in the first half of 2024. Additionally, given the relatively high costs of leasing equipment, only Bit Digital has been able to produce a comfortable margin. Bit Digital, Core Scientific, Hut 8, and Hive’s profit margins for the first nine months of 2024 were respectively: 60%, 13%, 24%, 14%.
H.C. Wainwright Managing Director of Equity Research Kevin Dede says that Bit Digital’s margins are much larger because it already owns roughly half of the GPUs in its leased datacenter, so it only has to pay power and management costs. Bit Digital did not report how much it spent specifically on its Nvidia H100s in 2023, but it spent $81.5 million on property and equipment in 2023. Additionally, Dede said that Bit Digital has favorable power costs at this datacenter, which is in Iceland and uses geothermal energy. Clear Street Managing Director Brain Dobson, who covers Bit Digital among other bitcoin miners, assented to this view.
In addition to its successful leasing model, Bit Digital also recently acquired an HPC business, Enovum, for $46 million. Enovum currently operates 4 MW of HPC capacity according to a Bit Digital investor presentation, and according to a press release, it has 288 MW in the pipeline for expansion over an unspecified timeframe, of which Bit Digital “tentatively [plans] to bring an additional 20MW online by year-end 2025.”
Core Scientific plans to take its wholly-owned operations to the next level by retrofitting its existing and in-development bitcoin mines for AI operations through 2026, following in the footsteps of Iris Energy (IREN).
Each company will chart its own course, and each management team has a different vision for which strategy will position it for success in the fast-paced, burgeoning industry of AI computing services. So far, none of these approaches is a clear winner, but it is clear that the leasing and acquisition models are the safest bet, while the retrofitting model carries the most risk – but also the most upside. For the former model, Bit Digital is emerging as an early winner, while Core Scientific has positioned itself as a potentially lucrative play with the latter model.
Bitcoin Miners Have Four Options When Building AI Businesses
Bitcoin miners have a few entry points into AI: leasing a data center, acquiring one, building one from scratch, or retrofitting their existing bitcoin mines.
As mentioned above, leasing is straightforward: the company rents a data center that is rated for HPC and AI services, and they either outfit it with their computers, a third party’s computers, or rent out the computing power that’s already in place. This strategy is the datacenter equivalent of someone leasing an apartment and renting it out on Airbnb – for better or worse.
The margins for this model are thin, as the price to lease a top-tier data center is high and rising. According to a real estate investment firm CBRE Group report, the average cost per megawatt to rent a data center in the US rose 18.6% in 2023 to $163 per MW per month. For comparison, Core Scientific’s 16 MW HPC pilot in Austin costs roughly $102 per MW per month to operate, according to the company’s Q2 results.
The acquisition route is easy to parse as well. Hut 8, for example, purchased five datacenters in Canada from Terrago in 2022 for $23.7 million. This is a small deal in terms of data center acquisitions, as some of the largest deals often range in the tens of billions of dollars.
The details get trickier when miners consider building a data center from scratch or retrofitting an existing one, the most expensive option to deploy AI at scale. Consider a joint venture between private equity giant Blackstone and Digital Reality to understand the cost. The venture plans to develop 500 megawatts (MW) across four data centers for an estimated cost of $7 billion (or $14 million per MW).
The data centers that Blackstone and Digital Reality are building are called hyperscalers – the business’s largest, most complex centers. These hyperscalers require expensive, sophisticated infrastructure to ensure they are fail-proof and have near or uninterrupted uptime. A data center’s uptime capabilities designate its tier. Tier one data centers have no more than 28.8 hours of downtime per year, tier two no more than 22 hours, tier three no more than 1.6 hours, tier four no more than 26 minutes, and tier five can have no downtime.
The higher up the tier list a data center, the more networking and electrical redundancy the center must-have, which equates to a greater construction and operating cost. Miners like Core Scientific and IREN hope to reduce these costs by retrofitting existing mines, but whether or not this is cost-effective at scale – or even possible – is to be determined.
Retrofitting a Bitcoin Mine for AI is High Risk, High Reward
IREN has already retrofitted some of its Prince George, British Columbia site for AI computing. This pilot, which runs 816 H100 GPUs from Nvidia, has generated modest revenues for the miner through an agreement with Poolside, and it plans to expand its AI operations by retrofitting its Childress, Texas, site by the end of the year. IREN’s agreement with Poolside lapsed at the end of August, and the company has said that its AI capacity is being “redeployed with other customers.” IREN did not specify any revenue for Q2 in its SEC filings.
IREN’s operations are small, but its rival, Core Scientific, has grand aspirations. It has inked three deals with CoreWeave this year to deliver the company 500 MW of AI capacity by 2026 (for reference, Core Scientific’s current total MW capacity is 830 MW). CoreWeave will foot the bill for the expansion as part of the deal. Core Scientific has not disclosed which sites it plans to retrofit, although its under-construction 100 MW site in Muskogee, OK, seems like a safe bet. Core Scientific projects that its 12-year contract with Core Weave could generate $8.7 billion in revenue.
The question is, can Core Scientific pull off the retrofit? Core Scientific’s c-suite is stacked with data center professionals with pedigrees from Equinix and Hewlett Packard, so the team has experience with hyperscale data centers. However, the infrastructure needed to ensure these data centers are state-of-the-art is several magnitudes greater than what’s needed for a bitcoin mine, which requires no redundancy.
J.P. Morgan analyst Reggie Smith, who has published multiple research reports with partner analyst Charles Pearce on the intersection of bitcoin mining and AI, believes it’s still too early to say definitively whether or not retrofitting is a fool’s errand. “We’re still learning how realistic it is. There have been some smaller bitcoin mining operators that have announced much smaller deals with companies that are not necessarily household names in the AI space, and so I think we’re figuring out how realistic it is.”
Assuming it is realistic, the cost savings are attractive compared to building an AI data center from scratch (what folks in the industry call “greenfielding”) and could offer companies a quicker return on investment. Smith and Pearce’s most recent report estimates that retrofitting a bitcoin mining facility for AI could cost $5-7 million per megawatt, compared to $10-20 million for a greenfield data center. A retrofit saves infrastructure costs and cuts down on lead time for grid interconnections and power purchasing agreements, which are already in place for a functioning bitcoin mine. The capital expense to furnish an AI data center with GPUs is much higher than a bitcoin mine. But, the computers can last much longer under optimal conditions, anywhere from 5-10 years according to some estimates.
Per a prior J.P. Morgan report that Smith and Pearce published in June, the majority of the savings from retrofitting comes from the power infrastructure and construction that are already in place. The bulk of remaining costs fall on HVAC and IT infrastructure, which are much more costly for HPC data centers than for bitcoin mining sites.
Too Early to Call a Winner in the AI Race, But Early Movers Have the Best Shot
Bit Digital CEO Sam Tabar believes that part of Bit Digital’s recipe for success includes leasing. “The opportunity is now,” he emphasized in an interview. With this model, the company projects it will pull in $700 million over five years from a contract it signed with gaming company Boosteroid.
The jury is still out on whether or not this strategy is the best long-term. In Core Scientific’s case, it’s hard to imagine that Core Weave would commit to its contract with Core Scientific if retrofitting an existing mine is impossible at worth and infeasible at best. Still, retrofitting will come with substantial challenges: it will be expensive, and the infrastructure will entail a level of complexity that is greater than a bitcoin mine. If Core Scientific can pull it off, though, they will produce larger profit margins than if they were leasing and they will have great control over their operations.
As Dobson put it, the “answer really isn’t so simple” regarding which strategy will win out in the long run. “It really matters what you’re starting with … and what datacenter you want to build,” he continued.
“What we’ve learned from bitcoin mining is that it’s probably better to own the data center, because then you can do whatever you want with it and you can try to drive efficiencies. Also, since you are entering a business segment that you haven’t been active in before, you might want to buy something that’s working and that’s run by people who know how to do it so it will make it that much easier for you to scale, rather than starting at zero. But that said, it really depends on where you’re starting and your expertise on the management team,” dobson concluded.
Both Core Scientific and Bit Digital exist on opposite ends of the spectrum Dobson described, and they are the current leaders of the miner AI race, given the revenue they have currently booked and the deal flow they have in their pipelines – now, they just have to execute. Investors looking for a higher risk, higher reward play may want to allocate to Core Scientific, while those who are looking for safer exposure to AI through bitcoin mining equities might favor Bit Digital. Both of these miners, however, are probably the best situated out of the bunch, given the deals they have already inked and the fact that they are already generating decent revenue from these business lines. Hut 8 is also in a prime position given it already owns five data centers and has expanded its HPC business line with its newly announced HPC-as-a-service (although there is no clear, disclosed client pipeline yet).
In our conversation, Smith told me that, for underexplored opportunities, Cipher is “probably the largest and most immediate and visible opportunity.” He also mentioned that he thinks IREN is a good opportunity, but he cautioned that its AI business is “much harder to scale.”
The juiciest acquisition or development opportunity for a hyperscaler in Smith’s mind, however, comes from a miner who hasn’t made any effort to break into the AI/HPC market: Riot Platforms.
“We think [that Riot] has the best assets. Now, they’ve not been super bullish on the HPC opportunity, and they’re very particular about who they would do a deal with…But if you look at their locations and just the sheer size of their sites, I think they probably have the best assets,” he said, explaining that they could be a ripe acquisition target or partner for a hyperscaler or other HPC company. Smith pointed out that legacy hyperscaler and HPC firms have not yet brokered deals with bitcoin miners to develop existing assets into HPC infrastructure, and that massive miners like Riot are hitherto underexplored opportunities that could be attractive to these mainstream tech companies.
So Riot might be a dark horse candidate should a deal with a hyperscaler crop up that they can’t refuse. Core Scientific’s deals with CoreWeave currently provide the highest potential upside but also the most unknowns (and thus, the most risk). Conversely, Bit Digital’s recent acquisition of Enovum makes it a more conservative, safer bet, since the operation is already up and running, but one which has much less upside.
