SoftBank launches SB Neo to rent AI compute in the US at 10GW scale

Strip away the branding and SB Neo is a wager on a single scarce input: electricity. SoftBank Corp. President Junichi Miyakawa tied the US launch directly to the group making steady progress toward securing 10 gigawatts of power ^[1], and SB Neo's stated goal is to grow US data center supply to 10 gigawatts by around 2030 ^[1]. Bloomberg framed the launch squarely around that 10-gigawatt capacity ^[3]. The logic, echoed across the market chatter, is that GPUs can be bought and money can be raised, but power - the permits, the grid interconnects, the multi-year energy allocations - is what actually gates how much compute you can stand up. SoftBank's group-level energy buildout, already underway to feed its Stargate commitments, hands SB Neo a supply advantage that rivals scrambling for electricity cannot easily replicate. In this reading, SB Neo is not a new capital outlay so much as a way to monetize infrastructure the group was already committed to building.

The easy caricature of a neocloud is a warehouse of GPUs billed by the hour, and McKinsey (cited by The Register) warned exactly that these models are 'fragile' because they rest on commodity hardware ^[2]. SoftBank's counter is that SB Neo is not shipping bare metal but a full operating layer: its proprietary Infrinia AI Cloud OS, which supports Kubernetes-as-a-Service and LLM inference APIs ^[1][2]. Crucially, this is not vaporware. SoftBank has run a beta GPU cloud in Japan on the same stack since May 2026 ^[1], giving it operating know-how before it touches a US customer. That is the intended moat against the commodity-hardware critique - the argument that a managed platform with real inference and orchestration tooling, hardened in production, is stickier than a raw GPU rental.

The market read the entry as a threat first. Following the SoftBank and Meta AI-cloud news, CoreWeave (CRWV) fell about 13.9% and Nebius (NBIS) dropped roughly 17% intraday ^[5]. Bloomberg noted SoftBank will compete directly with both CoreWeave and Nebius ^[3]. Yet the more interesting fight is over whether the market is saturating or still starving. Aggregated Wall Street analysts argued the SoftBank and Meta moves are not a bet on computing overcapacity but rather a strategic wager on the long-term expansion of AI training and inference demand ^[4], pointing to a neocloud market that topped $25B in revenue in 2025, grew more than 200% year over year in late 2025, and is projected to approach $400B by 2031 ^[4]. Investor sentiment on social channels split along the same line - skeptics warned of a crowded field, while bulls read a deep-pocketed new entrant as validation that demand is real, with the recurring argument that permits, zoning, and equipment supply, not capital, are the true constraints and there is enough pie for everyone.

For all the ambition, SB Neo is a late mover with a levered balance sheet behind it. The service will not go live until FY2027 (the year ending March 31, 2028) ^[1], while CoreWeave, Nebius, and the hyperscalers are already selling capacity and signing enterprise clients today ^[4]. The funding picture adds risk: SoftBank secured a $10B loan backed by its OpenAI stake, reportedly with personal guarantees, tying infrastructure financing to OpenAI's valuation ^[2]. That intertwines two bets - the neocloud business and OpenAI itself - into one exposure. If AI demand compounds as SoftBank projects, the power moat and the head start on a $500B, roughly 10GW Ohio-focused buildout whose first 800MW phase targets 2028 operations ^[2]could pay off. If it stalls, a late launch on leveraged, OpenAI-linked capital is precisely the fragile setup skeptics have been circling.