Roughly the output of ten large nuclear reactors, this capacity targets the cloud GPU rental market, where SoftBank will join CoreWeave, Lambda, and the GPU equipped arms of AWS, Microsoft Azure, and Google Cloud.
SoftBank is putting roughly 10 gigawatts of US data center capacity behind a new cloud GPU rental subsidiary, and the next real news is which paying customers actually show up.
SB Neo, the US neocloud entity SoftBank announced on July 2, is the opening move. SoftBank Corp owns 51% of SB Neo and SoftBank Group Corp owns 49%, with SB Neo reporting as a consolidated subsidiary of SoftBank Corp itself. US operations are scheduled for fiscal 2027, which begins April 1, 2027 and ends March 31, 2028, a timeline that puts any paying customer at least nine months out from spinning up US capacity through this entity, even on the optimistic procurement timeline (The Register).
Cloud GPU rental is the business of renting AI training chips, the GPUs that train large language models, by the hour to AI labs and enterprises that do not want to build their own data centers. The incumbents SoftBank is joining are CoreWeave, Lambda, and the GPU-equipped arms of AWS, Microsoft Azure, and Google Cloud. SoftBank's product hook is Infrinia AI Cloud OS, a SoftBank Corp platform that has been running in beta in Japan since May and that packages Kubernetes-as-a-Service and Inference-as-a-Service behind APIs (Yahoo Finance, Mobile World Live).
SoftBank Group founder Masayoshi Son set the tone with the line that "the SoftBank Group will work together to deploy world-class AI infrastructure and drive the AI revolution" (The Register). It is the only direct SoftBank voice in the reference set, and the framing is portfolio-shaped rather than tenant-shaped. That distinction matters in a market where press-release customers tend to outrun signed contracts.
The competitive warning already on file comes via The Register's reading of McKinsey, which described the neocloud business model as "fragile and commoditized." McKinsey is one consultancy. The structural concern, however, is concrete: GPU rental margins are thin, spot pricing tracks hyperscaler capacity decisions, and customer commitments often run project-by-project rather than as multi-year contracts. Ten gigawatts of US capacity is a meaningful pile to land in that market.
A US build at this scale also brings a power story with it. The Department of Energy has flagged AI-driven data center demand as one of the defining load-growth drivers for the US grid, and an agency fact sheet on Ohio frames the trade-off between data center build-out and affordable retail electricity (DOE Ohio fact sheet). SoftBank has not disclosed where the US capacity will sit. Landed power cost will decide whether SB Neo can undercut CoreWeave or end up matching hyperscaler pricing.
The next checkpoint for readers is concrete. Whether SoftBank discloses a named anchor tenant, a hardware-of-record supplier, or a US site with a power agreement attached will tell the market whether SB Neo is a strategic hedge on contracted AI training demand or a re-run of the Vision Fund pattern of building capacity ahead of customers. The build is the easy part. The signed orders are not.