Oracle Jupiter Swaps Gas Turbines for Costlier Bloom Fuel Cells

Oracle's most ambitious AI data center site keeps changing shape mid-build, and the latest change is expensive. The Information reported that Oracle is facing multibillion-dollar cost surprises at its AI megacampuses, with Project Jupiter in Doña Ana County, New Mexico pivoting away from natural gas turbines toward a fuel-cell-based microgrid supplied by Bloom Energy, a redesign publicly confirmed in Oracle's own April announcement that pegs the campus at up to 2.45 GW of on-site capacity.

The context matters. Jupiter is part of the Stargate super-campus push and has been described as a long-term investment that could reach up to $165 billion, and Oracle has indicated it expects to spend about $50 billion on AI infrastructure this fiscal year. Swapping combustion turbines for fuel cells does buy Oracle a cleaner story, with the company saying the new design will reduce emissions by approximately 92% compared to the previously planned gas turbines and use a negligible amount of water, but fuel cells at that scale are not a like-for-like substitute on cost or schedule.

They also do not, on their own, solve the underlying problem. The site still needs methane to feed the Bloom units, and Oracle recently told regulators that "without a reliable source of natural gas, the power plant that will generate electricity for the data center campus cannot function." Per reporting summarized by Hunterbrook, the 17.77-mile gas pipeline required to feed the campus is stuck, there is no approved air permit for the fuel cells, and SemiAnalysis has pushed its base case for first power at Jupiter from 2027 to 2029, with interviews suggesting further slippage into the 2030s is possible.

The honest caveat is that The Information's exact dollar figure sits behind a paywall, and a chunk of the corroborating detail comes from a short-oriented investigation of Bloom that flags a wide gap between the roughly $20 billion backlog Bloom markets and the $492.6 million in binding contractual obligations it discloses to the SEC. Take the specific cost delta as reported, not settled. What is clear is direction: permitting friction and community pushback are forcing an on-site power redesign, and the new stack is slower and more capital-intensive than the plan Oracle underwrote.

For anyone with revenue booked against Jupiter's original 2027 timeline, starting with OpenAI's Stargate capacity commitments, the useful move now is to treat the fuel-cell path as the base case rather than the optimistic scenario, and price counterparty risk accordingly. Bloom gets a marquee anchor deployment, Oracle gets a cleaner permitting story, and the schedule, along with the bill, is where the surprise lands.