Core Scientific’s 75% return on AI deal isn’t the template for bitcoin miners, Bernstein says

The Mirage of Easy AI Pivots

In the wake of the last halving, every bitcoin mining executive suddenly started sounding like an AI data center developer. The narrative shifted almost overnight. We were told that the years spent securing cheap power and building out electrical substations were the perfect foundation for the generative AI boom. But if you look closer at the math, most of these teams are selling a dream that doesn't hold up to a basic balance sheet audit.

A recent report from Bernstein has put some cold water on the fervor surrounding Core Scientific and its massive deal with CoreWeave. While the market treated this 75% return on assets as a new benchmark for the sector, the reality is that Core Scientific is an outlier, not a template. For the average miner trying to transition into high-performance computing (HPC), the margins aren't nearly as pretty.

Why the 75% Return is a Unicorn

To understand why Core Scientific’s numbers look so good, you have to look at who is paying for the equipment. In their specific deal, CoreWeave is footing the bill for a significant portion of the capital expenditure. This changes the math entirely. If a partner is buying the GPUs and the specialized cooling systems, your return on the assets you actually own sky-rockets. It is a landlord’s dream, but it’s not a repeatable business model for everyone else.

When Bernstein looked at other major players like TeraWulf and Cipher Mining, the numbers dropped off a cliff. We are talking about stabilized returns on assets in the 4% to 5% range. That is a massive delta. It’s the difference between a high-growth tech play and a slow-moving utility company. For builders in this space, this is a wake-up call. You cannot simply swap out an Antminer for an H100 and expect the same capital efficiency.

The Hidden Cost of HPC Infrastructure

Building a bitcoin mine is relatively simple. You need power, some fans, and a shell of a building. It is a gritty, industrial process. Building a tier-three data center for AI is a completely different beast. You need redundancy on top of redundancy. You need liquid cooling, ultra-low latency fiber, and a level of physical security that most mining sites simply weren't designed for.

Most miners are sitting on "stranded" power. This was their edge in crypto. But AI companies don't want to be in the middle of nowhere if it means it takes them six months to get a technician on-site to fix a rack. The capital required to retrofit these sites is often higher than the cost of just starting from scratch. When you factor in those costs, that 5% return Bernstein projects starts looking very realistic, and arguably, a bit risky given the volatility of the AI market.

The Strategy Gap

What we are seeing now is a divide in the mining sector. On one side, you have the pure-play miners who are doubling down on efficiency and trying to survive on thin margins by scaling their hash rate. On the other, you have the "pivoters" who are trying to rebrand as AI infrastructure companies to save their stock price from the post-halving slump.

The problem is that the market is starting to sniff out the difference between a company with a real AI strategy and a company that is just buying time. TeraWulf has been praised for its low-cost power, but even they are facing a steep climb to prove that their infrastructure can compete with the likes of Equinix or Digital Realty. If the return on capital is only 5%, investors might decide they are better off holding the underlying assets rather than the infrastructure providers.

What This Means for Builders

If you are building in the crypto-infrastructure space, the lesson here is about specialization. The "jack of all trades" model where a single facility hosts bitcoin miners, rendering farms, and LLM training is likely a path to mediocrity. The operational requirements for these tasks are diametrically opposed. Bitcoin thrives on interruptible, cheap power. AI requires 99.999% uptime and high-cost reliability.

Founders need to be honest about their site’s capabilities. Is your facility actually capable of hosting HPC, or are you just courting venture capital with a buzzword-heavy slide deck? The 75% margin dream is likely dead for anyone who didn't get in early with a partner willing to subsidize their build-out.

The Takeaway

Core Scientific’s success is a result of timing and a very specific partnership structure. It is not a signal that the entire bitcoin mining industry is about to become an AI gold mine. For most operators, the shift to AI will be a capital-intensive grind with single-digit returns. The miners who survive the next two years will be the ones who focus on what they are actually good at—managing power and securing networks—rather than chasing a pivot they aren't equipped to handle.

The market is finally realizing that having a power plug doesn't make you an AI powerhouse. The infrastructure gap is wider than the hype suggests.

We need to stop looking at these two industries as interchangeable. One is a high-risk, high-reward digital commodity game. The other is a low-margin, high-reliability real estate play. Trying to bridge that gap without a massive partner to subsidize the risk is a dangerous bet for any founder or investor.

Read the original at The Block →