SoftBank Launches SB Neo for a 10GW US Neocloud Push

The list of companies trying to rent out AI compute keeps getting longer — and richer. On July 2, 2026, SoftBank Corp. and SoftBank Group Corp. announced they are establishing SB Neo, a new company to operate a “neocloud” business in the United States. The venture will be 51% owned by SoftBank Corp. and 49% by SoftBank Group Corp., making it a consolidated subsidiary of the former. Its ambition is stated plainly: to build toward 10 gigawatts of AI infrastructure capacity and sell it to America’s largest enterprises and hyperscalers.

What a “neocloud” actually is

The term “neocloud” describes a new class of provider that does one thing: rent out GPU compute for AI training and inference at scale. Unlike the general-purpose hyperscalers — with their sprawling menus of databases, storage tiers, and managed services — a neocloud is purpose-built around clusters of accelerators, high-speed networking, and the power and cooling to run them hot.

SB Neo fits that mold. According to SoftBank, the company will “provide major U.S. enterprises, including hyperscalers, with the computing resources needed for large-scale AI model training and inference.” In other words, its customers may include the same cloud giants that ordinarily are the compute supply — a sign of just how far demand has outrun what any single provider can build alone.

The plan and the timeline

SB Neo will lean on the SoftBank Group’s broader 10-gigawatt-scale energy and AI infrastructure, which the group says is currently under development. Services are slated to launch in the fiscal year ending March 31, 2028 (FY2027), and capacity will be added in phases, expanding toward that 10-gigawatt target over time.

The technology foundation comes from work already underway in Japan. Since May 2026, SoftBank Corp. has offered a beta version of a GPU cloud service powered by “Infrinia AI Cloud OS,” its software stack for AI data centers. SB Neo is designed to carry the expertise from that Japanese beta into the far larger U.S. market — a pattern of proving the operating model at home before exporting it.

10 gigawatts, in context

Ten gigawatts has quietly become the unit of ambition in this era. It’s the same figure at the center of the OpenAI–NVIDIA compute partnership, and it recurs because it marks the point where AI infrastructure stops being an IT problem and becomes an energy problem. Ten gigawatts is the output of roughly ten large power plants. Committing to that scale means securing land, grid interconnects, and years of power contracts long before the first customer signs.

That’s why the economics are so demanding. As we’ve written in our look at AI data-center economics, the binding constraints on this buildout are increasingly power, siting, and financing rather than the GPUs themselves. SoftBank’s advantage here is structural: through its group companies and long-running investments, it can marshal the capital and the energy commitments that a pure-play startup simply cannot.

A crowded — and well-funded — field

SB Neo does not arrive to an empty market. Industry trackers noted it was the second major neocloud entrant announced within a week, underscoring how quickly capital is pouring into the compute-rental layer of the AI stack. Established GPU-cloud specialists, hyperscalers expanding their own fleets, and sovereign-backed projects are all chasing the same demand.

The competitive logic is straightforward. Frontier AI labs need vastly more compute than they can build fast enough, and they’d rather rent flexible capacity than tie up their own balance sheets in concrete and transformers. That has created a wholesale market for AI compute — and everyone with access to power, land, and financing wants a piece of it. SoftBank’s move mirrors the wave of national-scale commitments we’ve tracked, from China’s $295 billion AI infrastructure plan to the hyperscalers’ record capital budgets.

What it means

SB Neo is a bet that the shortage of AI compute is durable enough to justify building a business around renting it out for years. It’s a reasonable bet, but not a riskless one.

The bull case is that demand for training and inference keeps climbing, that hyperscalers remain net buyers of external capacity, and that SoftBank’s access to energy and capital lets it undercut smaller rivals on the metric that matters most: cost per token served. If the SoftBank Group’s 10-gigawatt pipeline materializes on schedule, SB Neo could become a meaningful wholesale supplier to the very companies that define the market.

The bear case is timing and commoditization. Neocloud capacity is fungible — one operator’s GPUs look much like another’s — so the business competes largely on price and utilization. With multiple heavyweights entering within days of each other, there’s a real risk of overbuild: a wave of capacity landing around FY2027 into a demand curve that may or may not have kept pace. The FY2027 launch window is far enough out that a lot can change, in either direction.

What to watch. First, the anchor customers — a neocloud lives or dies on securing large, long-term contracts before it pours concrete, so watch for named commitments. Second, the power deals, which are the real gating factor at 10-gigawatt scale. Third, whether the crop of neoclouds announced this summer starts to consolidate or compete each other’s margins away. SoftBank has the balance sheet to play a long game. The open question is whether the market it’s building for needs quite this many players.