Bloom Energy told investors on its second-quarter earnings call that it now has visibility on 25 gigawatts (GW) of deployments — 25 GW of onsite fuel cell power systems that the company sees a credible path to delivering to customers, primarily AI data centers, over the coming years. The company shared this figure with investors on its second-quarter earnings call, noting that it’s not just interpreted as a growth story, but as a strong data point about the reliability of Bloom’s core technology.
A Pipeline, Not a Pilot
Twenty-five gigawatts is not a pilot program. It is roughly the scale of dozens of large data center campuses’ worth of committed and pipeline power. These systems are built on solid oxide fuel cells that convert natural gas or hydrogen to electricity onsite, without the multi-year interconnection queues that have made grid power a bottleneck for AI buildouts.
This figure sits alongside a set of results that are hard to wave away: record second-quarter revenue of $1.065 billion, product revenue up 215% year over year, and full-year guidance raised to $3.9–$4.2 billion. Bloom’s arrangement with Brookfield for onsite data center power was expanded from $5 billion to $25 billion, a 2.8 GW master agreement with Oracle is on the books, and Intel has turned to Bloom’s fuel cells to help power the expansion of its own HPC data centers.
Scale of Adoption as a Reliability Signal
The reason this matters for reliability is straightforward: Hyperscalers do not repeatedly commit at this scale to unreliable hardware. Oracle, Brookfield, and the other operators behind that 25 GW pipeline are sophisticated buyers with deep engineering benches and enormous capital at risk if their power infrastructure falters. A single campus going offline can cost these companies and the AI workloads they support tens of millions of dollars a day. That is not an environment where procurement teams sign expansion after expansion on faith. Sustained adoption across multiple quarters and operators functions as a market-validated proxy for field performance in a way that a vendor’s own claims cannot.
A Sector Under Scrutiny
That distinction matters more now because the broader off-grid power industry is under real scrutiny. Reports of cracked turbines and broken engine cranks at gas-fired sites and a campus forced onto diesel backup after a critical power failure have raised questions about whether some onsite power systems are ready for the spiky load profiles AI training and inference create. Industry experts have noted that nobody has enough operating history yet to say definitively how this generation of equipment holds up under sustained AI workloads.
Bloom’s fuel cells sit in a different category from the combustion turbines and reciprocating engines at the center of those reports. Solid oxide fuel cells generate power electrochemically rather than mechanically, without the rotating parts most exposed to the load swings AI data centers produce.
That doesn’t make Bloom’s systems immune to stress, but it means the specific failure modes reported elsewhere — cracked shafts and turbine housings from oscillating demand — aren’t a direct match for how a fuel cell stack degrades. Bloom’s systems are also built from independent modules rather than one large generating unit, so a single module’s failure doesn’t take a whole site offline. Bloom credits this design with enabling availability in the 99.9%–99.999% range. Unlike the 25 GW pipeline, this isn’t purely forward-looking: Bloom already has roughly 1.5 GW running at more than 1,200 sites worldwide — a working track record the pipeline builds on, rather than a promise standing alone.
The Supply Chain Behind the Number
But one question hanging over any technology asked to scale rapidly is whether the supply chain can support it. Bloom reports that its diversified global sourcing of scandium oxide, a dopant used in trace amounts in its fuel cell electrolyte, can support up to 25 GW of annual production capacity. Moreover, the pipeline is not dependent on any single country, including China. Bloom recovers scandium as a byproduct of titanium, nickel, cobalt, and uranium processing already happening at industrial scale worldwide rather than relying on dedicated mining. This procurement model avoids the single-point-of-failure risk that has made other critical mineral supply chains geopolitically fragile.
The Takeaway
None of this means onsite power is now a solved problem. The industry is still young enough that long-run performance data, spanning multiple years and multiple generations of equipment under real AI load conditions, simply doesn’t exist yet for any company in the space.
However, visibility into 25 gigawatts of deployment — backed by expanding commitments from sophisticated operators, a working 1.5 GW installed base, and a diversified supply chain built to match that scale — is itself a meaningful signal. In a sector where the real test is whether the lights stay on when a training run spikes demand by hundreds of megawatts in seconds, a sustained, repeat commitment from the buyers with the most to lose is one of the more credible reliability indicators available.
