Overview
Bloom Energy is a solid-oxide fuel-cell manufacturer providing on-site power generation. Until 2025–2026, Bloom’s installed base was concentrated in mid-scale commercial and industrial deployments. The April 2026 Oracle Project Jupiter announcement (up to 2.8 GW agreement, 1.2 GW already contracted) moves Bloom into the hyperscaler-primary-power category for the first time at GW scale.
Key facts
- Fuel-cell technology operates without combustion; minimal water use Oracle Project Jupiter Bloom Fuel Cells — DCK - 2026-04-29
- NOx emissions reduced ~92% versus conventional gas turbine generation Oracle Project Jupiter Bloom Fuel Cells — DCK - 2026-04-29
- Oracle Project Jupiter agreement: up to 2.8 GW total, 1.2 GW already contracted Oracle Project Jupiter Bloom Fuel Cells — DCK - 2026-04-29
- Project Jupiter campus: up to 2.45 GW on-site microgrid capacity (single integrated microgrid) Oracle Project Jupiter Bloom Fuel Cells — DCK - 2026-04-29
- Operating constraint: fuel cells “have very limited capacity to handle overloading” and operate optimally at steady power consumption (Prithpal Khajuria, Intel) Oracle Project Jupiter Bloom Fuel Cells — DCK - 2026-04-29
Newsletter relevance
Bloom is the operating partner for the most explicit hyperscaler-exits-the-grid move documented so far. The Oracle agreement makes fuel-cell microgrids the primary power architecture rather than the backup. This is the technology behind the Speed to Power story when the speed is being prioritized over grid resilience and overload protection.
Connections
- Oracle — Project Jupiter anchor customer
- BorderPlex Digital Assets — Project Jupiter co-partner
- Speed to Power — concept the deployment exemplifies
- Behind the Meter Generation — architectural category
- AI Buildout Grid Constraint
Source appearances
Open questions
- What’s Bloom’s full 2026 backlog beyond the Oracle 2.8 GW commitment? Are other hyperscalers in similar-stage negotiations?
- How does Bloom’s fuel-cell economics compare to behind-the-meter natural gas at GW scale? (The Enki AI piece’s $1.6B for ~1 GW of Williams gas suggests fuel cells must be cost-competitive even before counting emissions reductions.)
- The Khajuria comment on overload limits is significant — has Oracle’s design accounted for AI-workload spikes (training-run peaks, inference burst)? Or is the bet that AI workloads are steady enough that overload protection is no longer the binding constraint?