|
The Gap Nobody Budgeted For
The AI capital expenditure cycle is measured in hundreds of billions. Microsoft, Google, Amazon, and Meta have collectively committed over $700 billion to data-center buildouts through 2030. Each new facility requires 50 to 200 megawatts of firm, around-the-clock power. The grid was not designed to deliver that on the timeline the hyperscalers need it. Natural gas plants take four to six years to permit and build. Nuclear takes a decade. Solar and wind are intermittent. The firms building the AI infrastructure cannot wait for the grid to catch up.
The Shale Playbook, Applied Underground
In this tape, the capital rotation is not theoretical. Enhanced geothermal systems pump cold water into hot rock miles below the surface, heat it, and return it as steam to drive turbines. The technology uses horizontal drilling, directional steering, and hydraulic stimulation — the exact techniques that built the shale revolution. The DOE’s FORGE project in Utah proved the concept. Drilling rates improved more than 500 percent between 2017 and 2024. Fervo Energy commercialized it, achieving 30 meters per hour, nearly four times what FORGE demonstrated in its early wells.
Fervo’s May 14 IPO on Nasdaq raised $2.2 billion after upsizing the offering three times — the largest primary energy and power IPO in recent memory. The company has 658 megawatts in binding power purchase agreements worth $7.2 billion in contracted revenue. Its Cape Station in Utah, the world’s largest enhanced geothermal project, is commissioning its first 100-megawatt phase this quarter with 4.3 gigawatts of total resource potential at the site.
Where the Hyperscaler Money Is Going
Our view: the market is pricing AI as a semiconductor trade. The hyperscalers are pricing it as an energy trade. Google signed a 3-gigawatt geothermal framework agreement with Fervo, enough to power roughly 2.4 million homes. Ormat Technologies and NV Energy signed a 150-megawatt PPA in February specifically for Google data centers in Nevada. Meta contracted 150 megawatts of waterless geothermal with XGS Energy in New Mexico. These are not pilot programs. These are long-duration, binding infrastructure commitments from the same firms spending hundreds of billions on GPU clusters.
The cost curve supports the thesis. Wood Mackenzie estimated Fervo’s Cape Station can produce electricity at $79 per megawatt-hour without subsidies. New nuclear runs $142 to $222 per MWh. The U.S. Geological Survey estimates 150-plus gigawatts of EGS resource potential nationally. Current installed geothermal capacity is 3.97 gigawatts. The ratio between what exists and what is possible is 38 to 1.
What to Watch This Week
Worth watching: JOLTS data lands tomorrow. Jobs Friday closes the week. Both readings feed the macro environment that either supports or compresses the AI capex cycle. If hiring decelerates in construction and manufacturing, the buildout timeline stretches and the power gap widens. If it holds, the infrastructure trade accelerates. Either way, the capital allocation decision has already been made at the hyperscaler level. The question for the portfolio is whether the market prices the power side at the same conviction it priced the compute side — or whether the gap stays open long enough to become the trade.
|