Nothing stops this trAIn.
When Core Scientific announced its landmark AI/HPC deal with CoreWeave in June last year, it joined Hut 8 and Hive as one of the few public bitcoin miners with an incipient AI business line.
Fast forward a year, and now AI/HPC pivots are the rule – not the exception – for public miners. Out of all the leading public miners, only Cleanspark, MARA, BitFuFu, and Cango haven’t made the leap into AI.
For those flocking to AI, it’s easy to see why. These companies could earn anywhere from $1-4 million/MW/year hosting infrastructure for hyperscalers, or if they were to run the equipment themselves, they could bag more than $10 million/MW/year.
Core Scientific’s $10 billion deal with CoreWeave, for example, is projected to net the company $1.4 million/MW/year according to the contract’s 12-year term for 590 MW of critical IT load. Compare that to the $1.1 million/MW/year that an S21 can earn at current hashprice levels.
“There’s no question that the CoreWeave–Core Scientific deal last year was a shot across the bow to a lot of people — to think about, wow, there’s probably something more to investigate here,” Cipher Mining CEO, Tyler Page, said on a recent Mining Pod.
And that’s not just because the money is good (and that, unlike the volatile bitcoin mining industry, the income is much more stable). It’s also because there’s a bottleneck for readily available power in the United States. Bitcoin miners have been stacking power purchasing agreements and power allotments faster than most, so they are uniquely positioned to cash in on the AI wave.
“We have seen power as the bottleneck…So we’ve made a big bet on the desire for these large interconnects in potentially non-traditional areas, and that’s really where the industry is right now. It’s a combo platter of figuring out ways to get in front of the people you’re feeding,” Page continued.
A testament to the AI opportunities available for miners, Cipher came out of the woodwork with its AI aspirations when it announced that Japanese tech giant SoftBank made a $50 million investment in the company. Now, Cipher is evaluating Phase 2 of its Black Pearl site in Texas for up to 150 MW of AI/HPC load, as well as its forthcoming 300 MW Barber Lake and 100 MW Stingray sites.
Cipher joins Terawulf as one of the bitcoin mining companies that has come out swinging in 2025 with its AI/HPC strategy. Both companies have been working quietly in the background, but now they are becoming more vocal about their AI/HPC data center projects.
Let’s dive into some takeaways from our recent Mining Pod conversions with either company.
The powershell approach
Bootstrapping an AI data center business is easier said than done. First, there’s the obvious cost burden to construct one of these data centers, which are several magnitudes more expensive than bitcoin mining data centers.
“When we’re building Bitcoin mining facilities, we spend about $500,000 per megawatt to build the Bitcoin mining data center. We’re spending $7 to $8 million per megawatt to build the data center for AI. So they’re fundamentally different buildings … and our customer partner is spending $30 to $40 million per megawatt [on GPUs and infrastructure],” Terawulf COO and Co-Founder Nazar Khan said on the Mining Pod, adding that the all-in cost for a 100 MW AI/HPC project could eclipse $4-5 billion.
And then there’s the less obvious – but still challenging – question of the business model. A bitcoin miner could decide to simply host the equipment, an approach many of them are taking, including TeraWulf, Cipher, and Core Scientific. Or they could purchase the equipment and run it themselves as we’ve seen with IREN and Hive.
But that’s only one part of the equation. They also have to decide which clients they want to serve and which types of computing services they want to offer. Remember, “AI” and “HPC” are blanket terms; whereas bitcoin mining only means one thing, AI/HPC could entail inference services, LLM training, general cloud computing, and a litany of other tasks.
As such, bitcoin miners have to strike a balance between what is lucrative and what is feasible.
“And a question that we were getting kind of early in 2024 was, “Well, what’s your model?” Right? Some of our peers buy GPUs, some of our peers do other things. And so we had this kind of recurring question of, “Hey, how does this all work?” And we thought that there was a lack of understanding on the various business models,” Khan said.
Ultimately, Terawulf has settled on a hosting model similar to Core Scientific (if you build it, they will come) in its HPC deal with Core42. This “powered shell” model, as Khan called it, has lower revenue upside, but it’s low risk as the tenant shoulders most of the capital expenditure requirements.
“The least capital-intensive is just effectively delivering a powered shell to a hyperscaler. And so, you know, it’s a couple million bucks per megawatt. Power’s there, there’s a shell [for the GPUs]. And then that hyperscaler comes in and outfits the internals — the cooling, the chilling mechanisms, whatever they want. They bring in their own equipment, their hardware, to use it,” he said.
Designing AI data centers with flexibility in mind
Cipher appears to be taking a similar approach with its Black Pearl Phase 2 and Barber Lake sites. It hasn’t publicized an anchor tenant yet, but the company’s May 2025 presentation said that “multiple tenants [are] under NDA.”
In our interview, Page stressed that, as Cipher designs these sites, modularity is front of mind. Ideally, these sites can be flexible to deliver both inference and training load, depending on the tenant’s needs.
“You know, we’re setting up our sites such that, at least from our perspective, we think are going to be suitable for both applications — training and/or inference. You know, what we’re seeing on the technical requirement side is not necessarily two very distinct brackets of technical requirements,” he said, adding that clients may also demand different levels of fiber connectivity or ask for specific cooling solutions.
In order to keep themselves open for whatever comes, Page said that “it makes more sense not to build something on spec.”
“Because you can get radically different requirements. Someone might — for example, we have had one potential tenant say they want to do full evaporative cooling, and one say literally the opposite…So that’s something that obviously has tons of downstream knock-on effects — and not just the physical and design infrastructure, but like the lease and pricing and things like that,” he continued.
Bottlenecks and opportunities
Securing the contracts and devising a business model for AI are one thing, but actually breaking ground on the data center is a challenge in its own right.
And if access to power is one bottleneck, access to construction labor is another. Cipher Mining Chief Construction Officer, Chris Totin, said that data center businesses are projected to grow 12-30% year-over-year for the foreseeable future. Meanwhile, the construction labor force is growing 1% year-over-year – and that’s all construction, not just teams with acumen for data centers.
“On the load side, you’ve got 80 gigawatts roughly in queue. We’ve heard rumors that it’s about 50 gigs of data centers in queue,” Totin said. “We’re seeing this. We’re seeing the numbers start to creep. But we don’t have the labor force. So we have to go and find opportunities outside the state in high-performing, highly efficient groups of folks that are capable of turning out high capacity in a short amount of time.”
The miners that can lick both the energy and labor bottlenecks should have their fair share of opportunities – if they can execute.
Page espoused that these opportunities include previously untapped synergies between bitcoin mining and AI load. The thesis is simple: you outfit a data center with AI and HPC capabilities and bitcoin mining.
That way, if the GPUs on site are taking a breather or drawing less power than is optimal to monetize a site, you can flip on bitcoin miners to fill in the gap. It’s a perfect foil to how bitcoin miners already partake in demand response; instead of turning off for demand response, in this scenario, the miners are turning on in response to slackened demand for GPUs.
“And so, all the sort of tricks that you learn to curtail your data center in response to price signals in Texas — you can imagine the same sort of thing, looking at the variability of the power draw and what is excess to be soaked up. So as Bitcoin and Bitcoin mining gain broader acceptance and broader understanding, it’s a completely natural fit,” Page said.
“We have thought a lot about this. Black Pearl is a site that we’re currently marketing — the second half of the site — to HPC tenants. So that will be, I hope, if we get the tenant to the finish line, a live-action demonstration of this,” he concluded.
