Let’s start with a few basic facts. Fact one: Despite a recent rally, the price of bitcoin is still far below the bullish heyday of 2021, when many bitcoin miners bought their hardware.
Fact two: Much of the hype (and capital) in the startup space has shifted from Web3 to AI.
Fact three: Both cryptocurrency mining and AI development — such as the training of large language models — require high-end computer chips.
This story is part of CoinDesk’s 2023 Mining Week, sponsored by Foundry.
Fact four: There’s a global shortage of these chips.
All of this poses a natural question. Are crypto miners pivoting to AI?
The short answer is that while most bitcoin-only miners are not, larger shops are exploring their options. “I haven’t thought about that or really heard of anyone doing this,” says an influential home bitcoin miner who goes by the alias Econoalchemist.
The reason is simple. Most bitcoin miners use specialized ASICs chips, such as the Antminer S19 Pro, that are designed for SHA-256 hashing (Secure Hash Algorithm). They are excellent at mining bitcoin but lousy for doing anything else. They can’t be repurposed. (Every bitcoin miner I reached out to said the same thing.)
The math is different for larger operations. While the actual ASICs can’t be converted from bitcoin to AI, the infrastructure that the companies have already built — cooling systems, security, access to cheap energy — can be used for an expansion into AI.
So they’re starting to expand. Applied Digital, a Texas-based crypto miner, recently announced a $460 million deal to host AI cloud computing in its data center. (Wall Street approved; shares immediately jumped 17%.) Iris Energy, another Texas-based mining company, announced an expansion and revitalization of its high-performance computing (HPC) data center strategy, which was generally viewed to mean a push towards AI. (Again Wall Street approved; shares surged by 21%.)
Skeptics might see this as a ploy to capitalize on a fashionable trend – like how in 2017, Long Island Iced Tea made the natural pivot to “Long Blockchain Corp.” But the companies view it as a way to reduce systemic risk. Mining profits correlate with bitcoin prices. So adding other services — like hosting AI computing — lessens the reliance on bitcoin.
“The business is looking for a diversified revenue stream. If bitcoin goes to $10k or $20k, we’re in a less stressed position, and can still act strategically compared to our peers,” says Josh Rayner, VP of High Performance Computing at Hut 8, which until recently had been exclusively a cryptocurrency mining company.
Hut 8 began its pivot early. In January of 2022, well before the hype of ChatGPT, Hut8 invested in five data centers and two cloud regions that could be devoted to HPC. Unlike the laser focus of bitcoin-mining ASICs, these data centers are packed with Nvidia GPUs that could perform a wider range of workloads — gaming, virtual reality, AI, machine learning. And they still continue to operate bitcoin rigs. “Our core thesis is that we picture a world where mining and data centers and HPC workloads [used for AI] come together,” says Rayner, “and we’re starting to see that more and more.”
Back in early 2022, it occurred to Hut 8 that they’ve already done some of the heavy-lifting to service AI customers. “We have the staff. We have the compliance. We have the expertise to operate traditional data centers,” says Rayner. “Mining really goes hand in hand with that. You have a lot of the same synergies.”
Then there’s the pivot from Ethereum mining to AI.
When Ethereum switched from Proof of Work (which requires mining) to Proof of Stake (which does not), suddenly the ETH-mining equipment had nothing to do. And while the bitcoin mining ASICs can do nothing but mine bitcoin, the chips they used to mine Ethereum — Nvidia A40s — are more versatile. “They’re very capable of VFX rendering, of gaming, and they can do AI/ML [machine learning] workloads,” says Rayner. So the chips that once mined Ethereum have now been redeployed.
(Rayner clarifies that the old ETH chips are “limited in scope in what they can do,” and that they “come with low memory and low storage capacity.” But better to put them to use than to let them sit idle.)
So how exactly do the chips of Hut 8 power the development of AI? Here’s a real-world example. XYZ AI, a startup, wants to let people convert plain text to 3D imagery. They need to train their models using a massive data set, and that requires chips for extensive processing. If you’re a startup like XYZ, you can either purchase the chips yourself — and chips are hard to come by — or outsource the computing to a cloud provider like Hut 8. (It’s a similar model to parking data on the cloud with Amazon Web Services.) So XYZ is essentially renting out computing from Hut8.
In the future, this will allow XYZ’s users to type in something like “show me a sword that is dripping with melted strawberry ice cream,” and then magically see the image. Generative AI is just one of the AI categories that require processing horsepower. “There are a zillion of these specific types of applications that these chips are being used for,” says Rayner. “That’s why demand is absolutely exploding.” He cites medical technology, gaming, biology, and CAD drawing as other potential customers for Hut 8. Each of them “start with a big-data set, and that’s where model training comes into play.”
Mintgreen, a Canadian company that recovers heat from bitcoin mining and then uses it to harness power, is in the early “conceptual” stages of exploring an expansion into AI. “We’ve had questions from our investors about the possibility of a pivot,” says Colin Sullivan, Mintgreen’s CEO. Sullivan clarifies that this is still just theoretical and says “I need to investigate the economics deeper,” but he acknowledges that “down the road, it’d be wise to diversify into other computer-intensive electronics.”
That’s the key word here: diversify. For mining-focused companies like Hut 8, Applied Digital, Iris Energy, and potentially Mintgreen, the idea is not a complete pivot from bitcoin to AI, but rather creating a broader portfolio that’s more durable in a bear market. “We need to have multiple revenue streams,” says Rayner, “so that we’re the most efficient miner and can stay nimble and strategic if margins compress.” And “margins compress,” of course, is a euphemism for another crash in bitcoin prices. That could wipe out some bitcoin-only miners; for others, AI could mean survival.
