The Texas Power Grab: When Mining Hype Meets AI Reality

The Texas Power Grab: When Mining Hype Meets AI Reality Hook Two weeks ago, Mara Holdings and Galaxy Digital quietly filed 8-K forms disclosing land acquisitions in Texas. Combined square footage? Roughly 1,200 acres near the Permian Basin. The press releases spun it as a strategic pivot: “high-density computing infrastructure for AI and digital assets.” But read the fine print. Neither company has a single binding AI compute contract. What they have is a power purchase agreement (PPA) with ERCOT, a state grid that has already suffered three winter blackouts in five years. The ledger remembers what the marketing forgets: land without a kilowatt is just dirt. I’ve seen this before—in 2021, a similarly hyped “AI mining” project in Norway collapsed when curtailment clauses kicked in. The math doesn’t care about narratives. Context MARA Holdings (formerly Marathon Digital) operates the largest fleet of S19 XP ASICs in North America. Galaxy Digital is a diversified crypto financial services firm with a mining division. Both are publicly traded, reporting to the SEC. Over the past eighteen months, every major miner—Core Scientific, Riot Platforms, Hut 8—has announced plans to pivot toward AI hosting. The reasoning is straightforward: Bitcoin halving reduces block rewards by 50% every four years, while AI compute demand grows at 60% CAGR. By converting baseload power capacity into GPU racks, miners can stabilize revenue streams. Texas is the epicenter: cheap wind and solar, deregulated market, and a state government that courts capital. But there’s a catch. The pivot isn’t a software upgrade. It’s a capital-intensive rebuild. A single Nvidia H100 cluster costs $250,000 per rack. A 100-megawatt facility requires $400 million in upfront hardware. And the power needed for AI cooling is 2.5 times higher per square foot than mining. The marketing says “seamless transition.” The balance sheet says “massive dilution.” I’ve audited four such transitions as a risk consultant. Three missed their go-live dates by over six months. Trace every byte back to the genesis block—the code doesn’t lie, but developers do. Core Let’s stress-test the numbers. MARA currently operates 23.3 EH/s, consuming roughly 350 megawatts. To convert even 30% of that capacity to AI compute, they need to replace ASICs with GPUs and install liquid cooling infrastructure. Estimated CapEx: $600 million. Their last 10-Q shows $475 million in cash and equivalents. That means either debt or equity issuance. Assume 20 million new shares at $15 per share (current price: $17). Dilution: 35%. For Galaxy, the situation is similar but worse: their mining arm is a subsidiary, and the parent’s balance sheet is tied to volatile crypto holdings. The risk is a number until it becomes a breach. Let’s examine the AI narrative from a purely mathematical angle. The market expects these facilities to earn $15–$20 per GPU hour. But the actual spot price on platforms like Vast.ai for H100s is $3–$4 per hour. The delta is covered by long-term contracts with hyperscalers (Microsoft, Google). Yet as of today, neither MARA nor Galaxy has disclosed a single signed contract with a Fortune 500 cloud customer. The term sheets I’ve seen from Core Scientific required 3-year commitments and guaranteed power costs. No guarantee, no contract. So what are they buying? Options on a dream. The metadata is not ownership; it is merely a pointer. The real asset is not the land—it’s the ERCOT interconnection agreement. And that agreement can be revoked or repriced. In 2022, a similar deal in West Texas collapsed when the local co-op raised transmission fees 400% after infrastructure repairs. The project never launched. Greed optimizes for yield, not for survival. Contrarian Now, let me play the other side. The bulls argue that miners have unique advantages: existing substations, licensed electrical infrastructure, and access to interruptible power rates. They point to Core Scientific’s 200MW deal with a “leading AI hyperscaler” as proof of concept. And they are not entirely wrong. In a world where new AI data centers face 4-year lead times for grid connections, miners can flip a switch in 12 months. The storage-first philosophy applies here: if you already have the physical conduit, you own the bottleneck. I saw this firsthand during my 2020 audit of a DeFi protocol that mispriced oracle latency—the value wasn’t in the smart contract; it was in the node access. Similarly, the value here isn’t in the AI compute itself, but in the real estate and power leases. If MARA and Galaxy can secure long-term PPAs at $0.02/kWh and rent out GPU hours at $5/GPU hour, the math works even without the crypto cycle. The counter-argument is not that the strategy is flawed, but that the execution is priced as if it’s already achieved. The market is assigning a premium to these companies based on optionality, not realized revenue. That’s fine, until the option expires worthless. A mirror reflects the face, not the value. Takeaway The next 12 months will separate the engineers from the marketers. Watch two things: (1) binding AI compute contracts with named customers, and (2) the CapEx-to-revenue conversion ratio in quarterly filings. If MARA’s Q3 2025 report shows AI revenue below $10 million with $400 million in AI-related CapEx, the story is toast. If Galaxy signs a multi-year deal before year-end, the land grab was smart. But don’t mistake a shovel for a treasure chest. The ledger remembers what the marketing forgets—and in Texas, the wind doesn’t always blow. I’ll be tracing the bytes from these PPAs back to the ERCOT settlement statements. Code doesn’t lie, but consultants do. I’d rather trust the meter than the press release.