Core Scientific is expanding its AI deal with CoreWeave.
Russell Cann, Core Scientific’s chief development officer, talked to CoinDesk about the particulars of the bitcoin miner’s multi-billion dollar infrastructure agreement with CoreWeave.
AI fleets and bitcoin mining operations have very different requirements, Cann said, with AI clusters being much more capital intensive.
Bitcoin miner Core Scientific (CORZ) revealed a deal on Tuesday to help cloud-computing firm CoreWeave expand its artificial intelligence capabilities.
It’s easy to think that Core Scientific is simply jumping on the AI bandwagon with this agreement (which it hopes will generate $8.6 billion in revenue over 12 years) to build 500 megawatts of infrastructure. Mining has gotten brutal following the fourth Bitcoin halving in April (which hammered profits), and other firms – such as Hut 8 and HIVE – are now dedicating significant resources to AI computation instead of single-mindedly extracting bitcoin (BTC).
The truth, however, is that due to its unique business model – a heavy focus on building application-specific infrastructure for data centers – Core Scientific was almost always guaranteed to have a leg up on other bitcoin miners with AI.
Not only that, but Core Scientific had a head start. The company entered the AI game in 2019, far ahead of the rest of the industry, and while it briefly entered Chapter 11 bankruptcy protection following the crypto market’s collapse in 2022 before re-emerging at the beginning of 2024, it still has technical know-how.
“Core Scientific started out as a data center operator primarily providing colocation services rather than for proprietary mining. So it may have a relatively longer history than other public peers in terms of constructing and operating proprietary infrastructure,” Wolfie Zhao, head of research at mining advising firm BlocksBridge Consulting, told CoinDesk.
“The investment requirements for purpose built infrastructure and GPUs for AI workloads are so high that capital allocators cannot afford to take operational risks by partnering with inexperienced operators,” CJ Burnett, chief revenue officer at bitcoin mining marketplace Compass Mining, told CoinDesk. “A track record of performance and deep industry relationships can create a defensible moat for the first movers in the space.”
Core Scientific’s second ace in the hole? This multi-billion-dollar deal only marks another step in its fruitful relationship with CoreWeave.
“It wasn’t something that just came out of nowhere,” Russell Cann, the miner’s chief development officer, said in an interview with CoinDesk.
“We were their largest GPU hosting provider,” he said, referring to graphics processing units, the high-performing computers that form the basis of AI clusters. “We’ve had a long-standing relationship with CoreWeave. There’s been a lot of trust there. We worked on a lot of engineering pieces together as they developed their GPU cloud business.”
So it isn’t so much that Core Scientific is diversifying into AI, but rather that the seven-year-old company – one of the biggest mining operations in the world in terms of hashrate, per TheMinerMag data – is growing in a way that plays to its core strengths.
The agreement separates responsibilities between Core Scientific and CoreWeave, with the former being in charge of building a data center suited for AI clusters, while the latter – which ultimately owns the machines, and the software stack – provides high computational services to its clients. Any modification to Core Scientific’s existing infrastructure is funded by CoreWeave.
“We build our infrastructure, in most cases, from the substation all the way down to the chip itself that mines bitcoin,” Cann said, and the same level of detail applies to building the machines dedicated to AI.
There’s a lot to take into consideration. Data centers for AI fleets and bitcoin mining may look the same from the outside, but once in the building, everything changes, said Cann. That’s because Core Scientific doesn’t believe in one-size-fits-all – the firm’s competitive advantage stems from its ability to customize every bit of infrastructure to make it application-specific.
For example, the power infrastructure is completely different for mining and AI applications. If the power goes down in a bitcoin mine, it’s not the end of the world, because the ability to switch your machines on and off depending on the price of electricity is an essential facet of bitcoin mining, to the point that many firms use that function in a strategic manner.
But AI fleets need constant uptime, and that means implementing expensive power redundancy measures, including batteries, backup generators and uninterruptible power supply, or UPS, systems.
Another example: Whether you’re mining bitcoin or running an AI cluster, machines must be cooled to avoid overheating – but the optimal cooling method will depend on the application, Cann said. Most bitcoin rigs are cooled by fans or by being submerged in a pool of dielectric fluid, which comes at little cost. The GPUs used for AI, on the other hand, require some form of air conditioning, or to pass fluid over the machine’s silicon chip – two methods that consume large amounts of energy. Other differences emerge when you look at the kind of hard drives used for AI fleets compared to bitcoin mining operations, or the amount of fiber and connectivity that is needed for each site.
Big picture: AI sites tend to be much more expensive and less versatile than bitcoin mining operations, which can be plopped in all kinds of remote locations all around the world. That’s why Core Scientific, as a rule of thumb when examining a new site, will determine whether the economics of it makes sense for AI purposes first, and, if not, look at it from a bitcoin mining perspective.
“If I built a world-class air-cooled bitcoin mining site, I’m going to spend probably between $500,000 to $750,000 per megawatt,” Cann said. But that cost climbs up to $10 million to $12 million by megawatt for AI-related GPUs, he said.
That certainly puts the size of the 500-megawatt deal with CoreWeave into perspective. By that measure, it will cost roughly $5 billion to build the infrastructure CoreWeave needs. The project will be big enough that it could power roughly 100,000 homes, according to data from the Electric Reliability Council of Texas – the Lone Star State’s electrical grid operator. Most of that power will be coming online in 2025 and 2026.
It’s still just a fraction of Core Scientific’s total footprint. The firm currently owns 1,600 megawatts worth of heavy electrical infrastructure and has 1,200 megawatts of power purchase agreements.
According to Cann, the reason Core Scientific is so good at building application-specific infrastructure is because the firm has a wealth of experience using all kinds of high compute machines.
For example, the company used GPUs to mine ether (ETH) back before Ethereum switched its consensus mechanism from proof-of-work to proof-of-stake in 2022. Using similar GPUs, Core Scientific started hosting AI clusters in early 2019, Cann said, way before ChatGPT and chipmaker Nvidia shook the world.
“A lot of people talk about us pivoting into AI, but we’ve been doing AI for a long time,” Cann said. “We’re taking advantage of market conditions and expanding our AI piece back again.”
And the fast-paced world of bitcoin mining has shaped the team – which counts people with decades of experience building traditional data centers – in a way that gives it an edge over their competitors who’ve never been involved with bitcoin, Cann said.
“Bitcoin is a 24/7, 365 days a year, global market. It’s always moving,” Cann said.
“Some of the traditional guys that don’t have experience mining bitcoin, they’re just going to be a little bit behind, because they have to catch up on that fast iteration,” he added. “AI is the only thing I’ve ever seen that iterates as fast as bitcoin mining.”