Why AI Data Centers Are Starting to Build Their Own Power Plants

If the grid can't deliver power fast enough, can AI data centers bring their own? Some US developers are ordering turbines, contracting for fuel cells, and planning campuses around dedicated generation. But getting enough megawatts is only the beginning. Reliability, equipment availability, fuel supply, permitting, and economics all shape whether onsite power can actually get a campus online faster. In this episode of UNDERLYING, we examine Crusoe's GE Vernova turbine orders, Microsoft's planned Pecos campus, Oracle and Bloom Energy's fuel-cell agreement, and the IEA's analysis of onsite generation and reliability. This is episode 2 of the Power Arc, following “The AI Race Just Hit the Power Grid.” CHAPTERS 00:00 Can AI bring its own power? 00:36 Crusoe's turbine orders 01:12 Microsoft's Pecos plans 01:33 Who builds and operates the plant? 02:09 Turbines and fuel cells 02:19 Oracle and Bloom Energy 02:55 Two routes to power 03:27 The value of opening earlier 04:00 Capacity versus reliability 05:30 The shared fuel dependency 05:58 The wider infrastructure system 06:46 Why the grid still matters 07:37 Can onsite power be faster? SOURCES & FURTHER READING GE Vernova — Crusoe's 29-unit turbine agreement (July 22, 2025): https://www.gevernova.com/news/press-relea... Microsoft — Planned Pecos campus and dedicated energy supply (June 22, 2026): https://blogs.microsoft.com/blog/2026/06/2... Bloom Energy — Oracle partnership expansion (April 13, 2026): https://www.bloomenergy.com/news/bloom-ene... IEA — Key Questions on Energy and AI (April 16, 2026), especially Chapter 6: https://www.iea.org/reports/key-questions-... EVIDENCE & VISUAL NOTES Orders, contracts, and announced plans are distinguished from operating facilities throughout. Company announcements establish what companies reported, not independent verification of every deployment milestone. The examples discussed are primarily in the United States. The IEA capacity graphic is an adapted explanatory illustration based on Key Questions on Energy and AI (2026), pp. 52–54 and Figure 6.2, licensed CC BY 4.0. The modeled range depends on assumptions and is not a universal sizing rule. https://creativecommons.org/licenses/by/4.0/ The thumbnail is an AI-generated illustrative composite, not a photograph of a named facility. System diagrams are conceptual, not electrical design drawings. They do not represent exact site layouts or dispatch behavior. Research cutoff: October 18, 2026. Real photographs and stock footage are labeled; they do not depict Pecos or the announced Crusoe/Oracle deployments. UNDERLYING explores the physical systems behind technological change. #AI #DataCenters #Energy VISUAL CREDITS Google campus, Council Bluffs (2017): Chad Davis, CC BY 2.0. Cropped/animated. https://commons.wikimedia.org/wiki/File:Go... Bloom installation at eBay (2010): Bloom Energy; photo Jakub Mosur, CC BY 2.0. Cropped/animated. https://commons.wikimedia.org/wiki/File:Bl... CC BY 2.0: https://creativecommons.org/licenses/by/2.0/ Simmering gas turbine (Austria, 2010): Wdwd, CC BY-SA 3.0. Cropped/animated; adapted image portions under the same license. https://commons.wikimedia.org/wiki/File:Kr... https://creativecommons.org/licenses/by-sa... Data center drone footage: Jakub Zerdzicki / Pexels, clip 38081600. Network cables: Dima Krivoy / Pexels, clip 1085656. https://www.pexels.com/license/ Gas power plant aerial: Akarawut Lohacharoenvanich / Vecteezy, asset 7995834. https://www.vecteezy.com/video/7995834-aer...