The U.S. government has formally partnered with SoftBank on what could become the largest AI infrastructure project in American history. The Department of Energy announced on March 20 that SB Energy, a SoftBank subsidiary, will develop a co-located AI data center and power generation complex at the former Portsmouth Gaseous Diffusion Plant uranium enrichment site in Piketon, Ohio. The project is part of the “Stargate” AI infrastructure initiative, which also involves OpenAI and Oracle.
What’s confirmed matters here, because the confirmed numbers are already extraordinary.
Reuters confirmed that the project is supported by $33.3 billion in Japanese funding tied to the U.S.-Japan Strategic Trade and Investment Agreement. SB Energy plans to develop up to 10 gigawatts of new power generation capacity, 9.2 gigawatts of which will come from natural gas, to support what multiple outlets have described as a 10-gigawatt-scale data center facility.
The scale needs context. A typical hyperscale data center consumes somewhere between 100 and 500 megawatts. One gigawatt powers roughly 750,000 average American homes. A 10-gigawatt data center would be in a category that does not yet exist. Whether the project reaches that scale, and on what timeline, remains to be seen.
Bloomberg confirmed that SoftBank CEO Masayoshi Son publicly stated the project could channel $500 billion into a single campus and called it potentially the “largest construction project in the country.” That framing belongs to Son. The committed figure, the one backed by signed agreements and Japanese government involvement, is $33.3 billion.
Construction is expected to begin in 2026, according to reports citing SoftBank. The early 2028 completion target and a reported $30 to $40 billion first-phase cost could not be independently verified against primary sources and are not confirmed facts.
Why it matters for AI infrastructure markets. This is not a speculative announcement. The $33.3 billion funding commitment, the DOE partnership, and the Stargate affiliation with OpenAI and Oracle represent a meaningful federal signal: the U.S. government is treating AI compute infrastructure as a national priority asset. For infrastructure investors, data center operators, and energy sector analysts, the Ohio announcement changes the competitive baseline. Federally backed projects at this scale attract different capital, different regulatory treatment, and different timelines than private-sector-only builds.
The natural gas dependency is worth watching. Nine-point-two gigawatts of gas generation powering AI infrastructure will draw scrutiny from environmental regulators and ESG-focused investors. Utility Dive’s coverage of the DOE energy component frames this as a deliberate policy choice, prioritizing speed and reliability of power delivery over energy mix considerations.
What to watch. The gap between Son’s $500 billion vision and the $33.3 billion commitment is where the real story lives. Watch for: formal construction contracts and procurement announcements in 2026; regulatory and environmental permitting for 9.2 GW of new natural gas capacity in Ohio; Stargate partner commitments (OpenAI, Oracle) that would lock in the larger investment vision; and competing state-level bids for the next federally backed AI infrastructure project.
The Ohio announcement follows reporting that Jeff Bezos is reportedly exploring a $100 billion fund aimed at AI-driven manufacturing transformation, a separate event covering distinct capital flows, but part of the same pattern of very large AI-aligned commitments concentrating in a short window.
TJS synthesis: The confirmed $33.3 billion is real. The 10-gigawatt ambition may also be real. But the $500 billion headline is a CEO’s articulated vision, not a signed check. The more important signal is the DOE’s willingness to anchor a project of this scale with federal partnership, that changes how other AI infrastructure projects will seek funding, site selection, and political cover. The Ohio campus is as much a policy statement as a construction project.