Nvidia Says AI Data-Center Power Can Take 2–3 Years; Bloom Claims a 55-Day Route

Nvidia CEO Jensen Huang has put a hard timeline on a growing constraint for AI infrastructure: securing the land, power and building shell for a data center can take two to three years. That delay creates an opening for power suppliers that can bring generation to a site more quickly, but it does not eliminate the permitting, water and legal hurdles surrounding major AI campuses.

Huang’s definition of an AI factory extends beyond servers and accelerators: operators first need land, electricity and a building. Nvidia is addressing that constraint by providing credit support for the initial 4.25 gigawatts of land, power and shell capacity at SB Energy’s Ohio data-center campus, alongside a $1.5 billion investment in the developer.

secure the land, power, and shell

Jensen Huang, Nvidia CEO

Onsite generation becomes a timing proposition

Evercore analyst Nicholas Amicucci identified Bloom Energy, Constellation Energy, Talen Energy, NRG Energy and American Electric Power as potential beneficiaries of AI-driven power demand. The common thesis is that the physical buildout required for AI compute is becoming commercially important alongside the chips themselves.

Bloom’s strategy is to use onsite fuel cells so data centers can operate without waiting for new grid connections. The company said the first system under its broader Oracle partnership became fully operational in 55 days, illustrating why deployable power equipment matters even though it does not match every stage of Huang’s two-to-three-year site-development timeline.

Oracle supplies the scale test

The Oracle relationship gives Bloom a substantial commercial test of that proposition. Oracle has contracted an initial 1.2 gigawatts of Bloom capacity, under an agreement that permits it to procure as much as 2.8 gigawatts. At Project Jupiter in New Mexico, Oracle and BorderPlex Digital Assets said a campus intended to serve OpenAI could use up to 2.45 gigawatts of Bloom fuel cells.

The buildout has two distinct constraints

• Capacity: Oracle’s initial Bloom contract covers 1.2 gigawatts, with potential procurement up to 2.8 gigawatts.
• Project approvals: Project Jupiter’s air-quality permit proceeding and an emergency authorization to use water from a new construction well were temporarily frozen by the New Mexico Supreme Court.

Faster equipment does not settle project approval

The Jupiter proceedings show the boundary of the onsite-power argument. The New Mexico Supreme Court temporarily froze both the air-quality permit process for the project’s microgrid and the emergency water authorization. A quicker generation system can change the electricity-delivery schedule; it cannot, by itself, resolve the approvals needed to build and operate a large campus.

Bloom also faces a proposed securities class action alleging that it misled investors about reliance on Chinese-sourced scandium. Bloom has rejected that allegation and said its scandium supply is not dependent on China. Separately, the company reported second-quarter revenue of $1.07 billion, up 165.5% year over year, and raised its 2026 revenue outlook to $3.9 billion to $4.2 billion. Those company results indicate growing activity, but the decisive question for the AI-infrastructure thesis is whether large, permitted projects can convert contracted power capacity into operating data centers on the required schedule.