Bitcoin Miners Urged to Focus on Infrastructure Rather Than GPUs

During the Energy Investors Forum in Dallas, Mike Alfred, founder of Alpine Fox, expressed his belief that the current surge in artificial intelligence (AI) infrastructure should not be seen as a bubble, according to TheEnergyMag.

However, he cautioned bitcoin miners that diversifying into AI hinges on the choice of operational model—whether to sell energy and facilities or to risk investing in GPUs.

Alfred is confident that the demand for AI infrastructure will remain strong for the next 20 to 30 years, even amid potential market corrections and economic downturns.

He outlined a critical decision for miners regarding their operational model. The first option involves owning GPUs at their facilities, which can lead to higher revenues and greater control but significantly increases the stakes. This approach requires financing for equipment purchases, managing the risk of rapid chip obsolescence, and ensuring consistent power usage.

“If you decide to own them, you must have a compelling reason,” Alfred emphasized.

The second option is colocation, where clients place their servers and chips in a facility that provides space, electricity, water, and cooling. Alfred described this model as more conservative, likening it to real estate, which makes it easier to attract financing.

Alfred serves as a non-executive director on the board of IREN, noting that approximately 80% of Alpine Fox's investments are concentrated in IREN and Cipher Digital.

Both companies initially focused on bitcoin mining before transitioning to the AI sector. Recent agreements include a five-year partnership between IREN and Microsoft valued at around $9.7 billion for cloud infrastructure using Nvidia GB300 chips, and a 10-year contract for Cipher with Fluidstack for a 168 MW facility at Barber Lake in Texas.

Many Bitcoin Mining Facilities Are Ill-Equipped for AI

Panelists at the Energy Investors Forum concluded that many existing bitcoin mining operations are difficult and costly to convert into AI data centers. They discussed a model where mining acts as a flexible load, temporarily monetizing unused electricity during the construction of AI campuses or absorbing excess generation, and shutting down when power is required for more expensive computations.

Jay Zapata, founder and CEO of SATOKIE, noted that in the competition for suitable substations, cryptocurrency mining will always fall short of AI in terms of business attractiveness.

According to observations from TheEnergyMag, investors are increasingly assessing miners based on their available power and project portfolio rather than just bitcoin production. For instance, a July agreement between TeraWulf and Anthropic for a 20-year campus in Houseville, Kentucky, includes about 401 MW of critical IT power and an expected revenue of around $19 billion. Similarly, Galaxy Digital's 15-year deal with CoreWeave for the Helios facility in West Texas involves 133 MW and approximately $4.5 billion in revenue. Both examples indicate a shift toward AI rather than a sustainable coexistence of the two models.

Zapata estimated the cost of building a mining facility at $200,000 to $300,000 per MW, compared to over $10 million per MW for AI infrastructure, excluding chip costs. He stated that an intermediate model is only viable with cheap electricity; at $0.06 to $0.08 per kWh, mining economics become challenging, prompting developers to either transition to AI or simply retain access to power without mining revenue.

Panelists highlighted that electricity for AI projects is just the beginning. Typical mining sites are often located in remote areas, built to simplified standards, and have limited reserve capacity. AI tenants require redundant fiber-optic lines, backup generation, sophisticated cooling, water, large plots for construction, and service availability of around 99.999%.

Shanon Squires, director of mining at Compass Mining, stated that after evaluating numerous new Tier 3 data center sites, it became clear that the vast majority are unrealistic for this purpose. He noted that the intersection of the two models primarily concerns land, substations, and step-down transformers.

Speakers identified sites with their own generation capabilities as a more sustainable hybrid model. Steve Barbour, head of Upstream Data, pointed out that gas power plants are often built with excess capacity that mining can utilize while AI facilities maintain guaranteed round-the-clock loads. In this scenario, cryptocurrency mining remains a secondary consumer and can quickly shut down without violating service obligations.

However, panelists agreed that expensive GPUs are less suited for such flexibility, and long-term AI contracts carry the risk of concentration. All four speakers estimated that by the next bitcoin halving in 2028, at least one major tenant, creditor, or landlord in the AI segment may face default or contract termination.

It is worth noting that publicly traded bitcoin miners are increasingly converting their energy capacities and data centers into infrastructure for artificial intelligence, according to data from Stanford University.