A 2.45GW data center. That’s the power of two nuclear reactors, or roughly the peak draw of 3.5 million homes. But the electricity source is not nuclear—it’s natural gas fuel cells. And the project is stuck in regulatory quicksand. Ledgers don’t lie, but energy permits do.
The project is Oracle’s “Project Jupiter,” a custom-built AI supercomputer for OpenAI, originally planned to run on gas turbines. In April, they pivoted to Bloom Energy’s solid-oxide fuel cells, pushing the microgrid capacity to 2.45GW and adding billions in costs. Analysts estimate the power solution alone could hit $8 billion—more than double the typical spend for a conventional gas plant at that scale. The twist? New Mexico’s state attorney general is investigating a petition that allegedly used residents’ names without consent to support the air permit. And a key fuel pipeline route was rejected, threatening fuel supply stability.
As an on-chain data analyst who cut his teeth auditing EOS ICO contracts in 2017, I’ve learned to follow the flow of funds and energy. This data center is no different—it’s a massive pool of computational power, and its energy input is the single most important variable for its output reliability. I’ve seen this pattern before in crypto mining farms: a low-hype but high-capital project hits a regulatory wall, and the costs cascade down to the end user. Here, the end user is OpenAI, and the cost overrun is billions—equivalent to a 30-50% increase in the total project CapEx. For context, a 100MW Bitcoin mining farm typically costs $100-150M in power infrastructure. Oracle’s 2.45GW is 24x that, but the cost per MW is inflating faster than any mining rig depreciation schedule.
The core insight is that the energy sector is now the bottleneck for AI scaling, just as GPU shortages were in 2022. On-chain data from the project’s financing—though not public—can be inferred from Bloom Energy’s stock movements and Oracle’s earnings calls. Since the pivot announcement, Bloom’s market cap has risen 60% (now ~$10B), pricing in an order of this magnitude. But that order is contingent on permits. The microgrid architecture also introduces a single point of failure: fuel delivery. Without a dedicated pipeline, trucks or rail would be needed to supply an estimated 100,000 MMBtu of natural gas per day—roughly the daily consumption of a small city. History repeats, if you read the chain. The 2021 Texas grid collapse taught us that energy infrastructure fragility can cascade; this data center is essentially its own mini-grid.
Contrarian angle: The narrative says fuel cells are cleaner and more efficient (60% vs. 40% for gas turbines). But efficiency gains are eaten by cost overruns. The real blind spot is that this project represents a financial engineering problem, not a technical one. Oracle is effectively building a captive power plant for a single customer—OpenAI. That’s like a Layer2 chain building its own sequencer for one app. It defeats the purpose of shared infrastructure. Moreover, the correlation between energy cost and AI adoption is not causation. OpenAI can pay the premium, but it will price out smaller players. This mirrors the fragmented liquidity in Layer2s—dozens of chains, same small user base. Here, dozens of data center proposals, same handful of hyperscalers.
Takeaway: Over the next quarter, watch two signals. First, Oracle’s Q3 earnings—if CapEx guidance jumps more than 15%, the cost overrun is real. Second, Bloom Energy’s manufacturing capacity updates—if they can’t deliver 5,000+ fuel cell modules on schedule, the project timeline slips. The energy chain determines the computation chain. In crypto, we say “follow the gas, not the hype.” Here, literally follow the gas molecules. If they can’t flow, the AI model training stalls. That’s the new scaling law: not compute, but current.