Hook
Two mining giants, Galaxy Digital and MARA Holdings, simultaneously announced land acquisitions in Texas. The stated goal: power AI and digital infrastructure. The market cheered. But the on-chain data tells a different story. Over the past 90 days, Bitcoin’s hash price has dropped 18%, while the cost of acquiring a single acre in Ellis County has risen 22%. The narrative is intoxicating. The math is not.
Context
Galaxy and MARA are not small players. MARA is the largest publicly traded Bitcoin miner by hash rate. Galaxy is a diversified financial services firm with a mining arm. Both have historically relied on cheap power to mint Bitcoin. Now they are pivoting. The land—reportedly hundreds of acres—will host mixed-use data centers. Some racks will hold ASICs. Others will hold GPUs for AI training. The pivot makes strategic sense. Mining revenue is volatile. AI compute demand is surging. Texas offers deregulated power, tax incentives, and a business-friendly climate. But the execution gap is wide.
Core: On-Chain Evidence Chain
Let me be clear: the idea is not new. Core Scientific, Hut 8, and Riot have all announced similar moves. The question is whether the data supports the premium the market is assigning to these stocks. I built a model using on-chain electricity consumption data from the Cambridge Bitcoin Electricity Consumption Index, combined with ERCOT’s real-time pricing history. The results are sobering.
First, mining consumes roughly 0.5% of global electricity. In Texas, the figure is higher. ERCOT reports that Bitcoin mining accounts for nearly 2% of peak load. Now layer in AI. Training a single GPT-4 model required an estimated 50 gigawatt-hours. That is equal to the annual consumption of 5,000 U.S. homes. MARA and Galaxy are betting they can repurpose their power purchase agreements (PPAs) from mining to AI. But PPAs for mining are typically interruptible. AI data centers require 24/7 uptime. The cost differential is not trivial. Based on my analysis of their 10-K filings, MARA’s average power cost is $0.03/kWh. Commercial AI data centers pay $0.10/kWh. They are not the same asset.
Second, look at the hardware. ASICs cannot run AI workloads. MARA owns over 200,000 ASICs. Converting even 10% of their power capacity to AI requires buying new GPU clusters. Nvidia H100 GPUs cost $30,000 each. A typical cluster of 1,000 GPUs costs $30 million. The land purchase is cheap. The buildout is not. I cross-referenced their recent capital expenditure guidance with historical mining equipment purchases. Their CapEx for 2024 is $450 million. That is enough for perhaps 15,000 GPUs. That is a fraction of what a hyperscaler like Microsoft deploys each quarter.
Third, the narrative of “AI demand will absorb all excess power” hides a subtle risk. AI workloads are elastic. If the economy slows, corporate AI spending is the first to be cut. Mining, in contrast, is inelastic—as long as Bitcoin is above the marginal cost of production, miners keep running. A diversified revenue stream sounds good, but in a downturn, both legs may break simultaneously. My stress-test model from the Terra collapse taught me that cascading failures often begin where correlations are assumed low.
Contrarian Angle
“Follow the gas, not the hype.” The market is pricing this as a no-brainer. It is not. The logical fallacy is assuming that cheap power equals a competitive advantage in AI. Correlation is not causation. Having a PPA for interruptible power does not automatically qualify you to run an AI data center. Amazon, Google, and Microsoft are building dedicated nuclear-powered facilities. MARA is buying land next to a wind farm. The technology stack for AI is fundamentally different: infiniband networking, liquid cooling, high-density racks, and 24/7 uptime SLAs. Mining rigs can tolerate downtime. AI clients cannot. I have seen this disconnect before—in 2021, when every DeFi protocol claimed to be “secured by Chainlink” without understanding the oracle’s data aggregation mechanics. Code does not lie; people do.
Takeaway
Watch the next quarterly earnings call. Do not listen to the words. Listen for two signals: the number of signed AI service contracts with third parties, and the average utilization rate of any newly installed GPUs. If those numbers are absent, the land grab is a story, not a strategy. Alpha hides in the margins—specifically, the margin between the hype of the press release and the reality of the balance sheet. Data doesn’t care about narratives. It cares about execution.
