The ‘Mullet’ Mining Pitch Has a Catch: Most Bitcoin Mines Aren’t AI-Ready

The bitcoin mining industry has found a colorful label for its attempt to participate in the AI infrastructure boom without abandoning its original business: the mullet.

AI is in the front, backed by long-term contracts and expensive infrastructure. Bitcoin stays in the back as a flexible load that monetizes unused electricity and shuts down when the higher-value computers need the power.

At the inaugural Energy Investors Forum in Dallas, a panel of mining and energy executives said bitcoin can monetize power while an AI campus is developed or absorb surplus generation, but the two businesses share much less infrastructure than investors often assume.

Their discussion also exposed a more consequential point: most existing bitcoin mines are not realistic candidates for conversion into AI data centers, and when both businesses want the same suitable substation, AI will take priority.

“Bitcoin will never win that fight,” Jay Zapata, founder and chief executive officer of energy and computing company SATOKIE, said during the session.

The limitation matters as investors increasingly value mining companies according to megawatts and development pipelines rather than their ability to produce bitcoin. AI leases can support longer-duration, dollar-denominated revenue, while mining income changes with bitcoin’s price, network competition and electricity costs.

TeraWulf (NASDAQ: WULF) Inc.’s July agreement with Anthropic offered the most dramatic recent example. The 20-year lease covers about 401 megawatts of critical IT capacity at a planned campus in Hawesville, Kentucky, and is expected to generate roughly $19 billion of contracted revenue, according to the company.

Galaxy Digital Holdings Ltd. has made a similar transition at its Helios campus in West Texas. Its initial 15-year lease with CoreWeave (NASDAQ: CRWV) covers 133 megawatts and is expected to generate about $4.5 billion of revenue. Galaxy has been winding down bitcoin mining at the site as it develops the campus for AI and high-performance computing.

Those projects validate the value of power-secured land. They do not necessarily validate the hybrid model, because each is closer to a full conversion than a permanent balance between AI and mining.

The EIF panel was moderated by Backbone Digital CEO Dave Perrill and included Zapata, Compass Mining Chief Mining Officer Shanon Squires, Upstream Data CEO Steve Barbour and Satoshi Energy’s Conor Beatson.

Zapata said bitcoin’s clearest role is as a bridge during the long development period for an AI facility. A miner can install comparatively simple equipment and generate revenue while the owner secures fiber, completes engineering work and finds a tenant.

He estimated that building a bitcoin mine can cost about $200,000 to $300,000 per megawatt, compared with more than $10 million per megawatt for AI infrastructure before the chips are included. Those figures were presented as industry estimates and will vary substantially by location and facility design.

“Bitcoin allows you to at least monetize that power and either run a break-even operation or run slightly profitable,” Zapata said. “You can use it as a bridge to get your AI site up and running.”

That bridge only works where electricity is cheap enough to mine profitably. Zapata said power costing six to eight cents per kilowatt-hour would make the economics difficult. A developer at such a site would need to move directly into AI or carry the power position without mining revenue.

The larger obstacle is that access to electricity is only the beginning of an AI project. Bitcoin mines typically use inexpensive structures with limited redundancy. They can operate in remote locations because the machines require little data connectivity and can tolerate frequent interruptions.

AI tenants may require redundant fiber, backup generation, sophisticated cooling, water, large parcels of developable land and contractual uptime approaching 99.999%. Legal details such as mineral rights, easements or proximity to existing infrastructure can stop a project even after months of negotiation.

Beatson said his team had worked on a site initially expected to become a bitcoin mine before a hyperscaler emerged as the customer. The diligence process was “drastically different,” he said, with small site details capable of derailing the transaction.

Squires said Compass had reviewed numerous greenfield opportunities and then assessed which could support a Tier 3 data center.

“The vast majority of them, it’s not realistic,” he said.

The overlap largely ends after the land, substation and step-down transformers, Squires said. Cooling, networking, buildings, backup systems and operational requirements diverge from there. That means a mining company may control a valuable power position without possessing an AI-ready facility.

The more durable version of the mullet may occur at sites with their own generation. Barbour said natural-gas plants are typically built with excess generating capacity. A flexible mining load can consume that surplus while an AI facility takes the firm, around-the-clock output it needs. Mining can then shut down when the generators are required for backup power.

“Any time you have power generation, there’s going to be a fit for bitcoin compute,” Barbour said, adding that the synergy depends on generation being located at the site.

That configuration gives each workload a distinct role. AI is the priority customer. Mining improves utilization of the energy asset and accepts interruption without violating a service agreement or affecting a specific end user.

The same flexibility is harder to replicate with GPUs. AI developers and technology companies have begun experimenting with demand response, moving workloads among locations or temporarily reducing consumption. Panelists were skeptical that those systems would match bitcoin mining’s ability to shut down quickly.

The reason is economic as much as technical. Costly GPUs generate no return while idle, and AI customers pay for availability. Bitcoin miners sell computation to an open network rather than a tenant operating under a service-level agreement. Turning off a group of mining machines marginally slows the network; turning off production AI servers can interrupt a customer’s business.

Zapata and Beatson said the collateral and interconnection requirements also make large grid-connected projects more difficult for miners to finance. AI projects can rely on leases and credit support from hyperscalers or other well-capitalized counterparties, while mining revenue remains exposed to bitcoin and hashprice.

But long AI contracts introduce a different concentration risk. All four panelists said they expected at least one significant AI tenant, lender or landlord to default or break a lease before Bitcoin’s next scheduled halving in 2028.

A relatively small group of hyperscalers, neocloud companies and financial backers supports a growing volume of announced infrastructure. Beatson said many gigawatt-scale campuses were overpromising relative to supply-chain and financing realities.

“There’s going to be a lot of big gigawatt-scale sites that fail in a spectacular fashion,” he said.