SoftBank Launches SB Neo as CoreWeave Slides 48%: Energy Is the New Moat

On July 2, 2026 — one day after a Bloomberg report sent CoreWeave down nearly 14% and Nebius Group down 17% in a single session — SoftBank Corp. and SoftBank Group Corp. announced SB Neo, Inc., a new U.S. neocloud subsidiary that will begin renting GPU compute to American enterprises and hyperscalers. The timing was deliberate. By stepping into the AI cloud race at the precise moment that market confidence in the pure-play GPU rental model is wobbling, SoftBank is signaling that it sees a structural opportunity — not in renting NVIDIA racks like everyone else, but in owning the energy supply, the software stack, and a strategic stake in the dominant AI lab.

The decision to enter a market that McKinsey described as having “fragile” economics is not naive. It is calculated — and the calculation is this: the commodity in neocloud is the GPU, but the moat is the gigawatt.

SoftBank’s New U.S. Entity

SB Neo will be incorporated in Delaware in July 2026, owned 51% by SoftBank Corp. and 49% by SoftBank Group Corp., and consolidated onto SoftBank Corp.’s balance sheet. Commercial neocloud services are scheduled to launch in fiscal year 2027, the period ending March 31, 2028. The target: 10 gigawatts of deployed AI data center capacity by around 2030 — a figure that would place SB Neo among the largest cloud infrastructure operators on the planet.

For context, one gigawatt of continuous data center power can support the equivalent load of roughly 750,000 homes simultaneously. Ten gigawatts is a commitment of a different order: roughly the entire power capacity of a mid-size state’s grid on a peak summer day.

SoftBank Corp. CEO Junichi Miyakawa framed the timing around power visibility, not GPU pricing. The launch decision followed what he described as “steady progress toward securing 10 gigawatts of power” in the United States — a statement that signals SoftBank’s edge is upstream of the GPU rack, in the infrastructure layer that makes the GPU rack possible at all.

How Infrinia AI Cloud OS Works

The technical foundation SB Neo will bring to the U.S. market is Infrinia AI Cloud OS, a proprietary software stack that SoftBank’s Infrinia team began deploying in Japan in beta form in May 2026 on NVIDIA GB200 NVL72 hardware.

Infrinia operates across the full GPU infrastructure stack — from BIOS and RAID configuration through operating system, GPU drivers, networking, Kubernetes controllers, and storage — automating tasks that require specialized engineering teams to perform manually in competing deployments. The result is two services delivered on top of the automated hardware layer: Kubernetes-as-a-Service (KaaS) for multi-tenant GPU environments, and Inference-as-a-Service (Inf-aaS) that exposes large language model inference through OpenAI-compatible APIs, allowing existing AI applications to connect without modification.

The hardware Infrinia runs on is engineered for exactly this purpose. The NVIDIA GB200 NVL72 rack integrates 72 Blackwell GPUs across 18 hosts, connected by NVLink — a high-speed interconnect that allows all-to-all GPU communication at 1.8 terabytes per second within the rack. That architecture effectively turns each rack into a single addressable memory pool rather than 72 independent GPU nodes, which matters for the largest LLM training and inference jobs, where inter-GPU communication latency is often the performance bottleneck.

The engineering tradeoff Infrinia makes is deliberate: by automating everything from BIOS upward and standardizing on NVIDIA GB200 NVL72, SoftBank reduces total cost of ownership compared with bespoke builds, but binds SB Neo’s infrastructure to NVIDIA’s hardware roadmap. For customers, the benefit is an OpenAI-compatible inference API that works on day one, without Kubernetes expertise; the constraint is that the software stack is not designed to be hardware-agnostic.

The Physical Infrastructure: Ohio and Texas

SB Neo’s compute will draw on two major infrastructure projects already under development by SoftBank Group.

The anchor site is the Portsmouth AI Technology Campus in Pike County, Ohio, where SB Energy — SoftBank’s energy subsidiary — is building a data center campus on leased U.S. Department of Energy land at the former Portsmouth Gaseous Diffusion Plant, a Cold War-era uranium enrichment facility. The project, announced March 20, 2026, alongside U.S. Energy Secretary Chris Wright and Commerce Secretary Howard Lutnick, envisions as much as $500 billion in total investment across data center halls, AI semiconductors, and associated power infrastructure at full 10GW build-out. The first 800-megawatt phase will be backed by $10 billion from SB Energy and is expected to come online around 2028. SB Energy is simultaneously building $4.2 billion in new electrical transmission infrastructure in southern Ohio with American Electric Power to ensure the project pays for its own grid impact rather than passing costs to residential ratepayers.

The second project is a 1.2GW data center in Milam County, Texas, built in partnership with OpenAI as part of the Stargate joint venture.

The power strategy for both sites relies primarily on gas-fired generation. SB Energy has committed to 9.2GW of new natural gas generation capacity as part of the Ohio project. Miyakawa made the energy sourcing explicit: SoftBank’s competitive edge comes from “the ability to secure sources of power, mainly from gas-fired plants” — a statement that identifies gigawatt-scale power control as the differentiator, not GPU procurement alone.

Why the Neocloud Model Is Under Pressure

The sector SB Neo is entering generated over $25 billion in revenue in 2025, with revenues growing more than 200% year-over-year in the fourth quarter. But the business model underpinning that growth is structurally precarious.

McKinsey’s November 2025 analysis found that bare-metal-as-a-service (BMaaS) — the model most neoclouds use — carries gross margins of roughly 55–65% before depreciation, but can fall to as low as 14–16% after labor, power, and hardware depreciation. The critical threshold is utilization: if it slips below 80%, returns flatline. With GPU rental prices for H100-class hardware falling roughly 50% from their 2023 peak, and the market expecting further erosion as supply normalizes, the pure procurement-and-resale model becomes progressively harder to defend.

The existential risk McKinsey identified is the Cloud 1.0 precedent: in the early 2000s, specialized compute startups filled gaps that Amazon, Microsoft, and Google couldn’t address quickly enough — and nearly all were eventually acquired, sidelined, or forced into niche roles as the hyperscalers caught up.

The week of June 30, 2026 compressed that warning into three days. On July 1, Bloomberg reported that Meta Platforms was developing “Meta Compute,” an initiative to commercialize its surplus AI infrastructure — offering hosted model access and raw GPU capacity to third-party customers. That single report sent CoreWeave (NASDAQ: CRWV) down 13.92% to close at $85.69, and Nebius Group (NASDAQ: NBIS) down 17.01% to $229.18. The selloff added to losses that had already taken CoreWeave roughly 48% below its 52-week high of $166.22.

CoreWeave’s Q1 2026 results illustrate the structural tension: revenue more than doubled year-over-year to $2.08 billion, and the revenue backlog reached $99.4 billion. But net losses widened to $740 million, and total debt reached approximately $24.9 billion. Interest expense for the quarter alone — $536 million — consumed roughly 46% of adjusted EBITDA. The company is generating cash but spending capital to stay ahead of its own growth at a rate that leaves nearly no margin for competitive disruption.

Nebius Group, which grew revenue 684% year-over-year to $399 million in Q1 2026, faces the same structural question. Its Meta agreement is worth up to $27 billion over five years — but that is the same Meta now reportedly developing infrastructure to compete with it.

SoftBank’s Vertical Integration Bet

The standard neocloud play is to procure NVIDIA GPUs, sign take-or-pay contracts with AI labs, borrow against those contracts, and race hardware depreciation to the bank. SoftBank is doing something architecturally different.

Holger Mueller, VP and Principal Analyst at Constellation Research, noted after the SB Neo announcement that the market is shifting from AI training toward sustained production inference — and that SoftBank has entered at exactly this inflection. “There’s an opportunity for SB Neo to redefine what an inference-focused neocloud for enterprises actually looks like,” Mueller said. “The key will be having access to enterprise data in as close to real-time as possible, or else having extremely fast data replication access to relevant enterprise data.”

SoftBank’s structural answer to the commoditization trap has four components that no pure-play GPU landlord can easily replicate.

Energy control. SB Energy owns the power generation — gas-fired, secured at scale — rather than purchasing power from utilities at market rates. In a market where data center operators are scrambling for contracts, that upstream control is a genuine cost advantage that compounds at 10GW scale.

Federal land and a DOE partnership. The Portsmouth Site partnership grants SB Energy access to infrastructure — grid connections, remediated federal land — that a company like CoreWeave or Nebius could not replicate through a private transaction. The DOE relationship also functions as a form of political insurance for a project that spans multiple presidential administrations.

Proprietary software. Infrinia AI Cloud OS is SoftBank’s attempt to build a software moat into what is otherwise a hardware commodity business. By automating the full stack and by designing Inf-aaS to accept OpenAI-compatible API calls, SoftBank is positioning Infrinia not just as internal infrastructure but as a platform that could eventually extend to other data center operators.

An anchor tenant relationship with OpenAI. SoftBank Group’s cumulative investment in OpenAI has reached approximately $65 billion. That stake gives SB Neo a plausible first major customer once commercial services launch — OpenAI has separately been reported in discussions about leasing capacity at the Ohio site. The anchor-tenant relationship is the mechanism that makes the multi-year take-or-pay debt structure work; without it, a new entrant borrowing at scale against an unproven customer base faces the same lender skepticism that pushed SoftBank’s OpenAI margin loan from $10 billion down to $6 billion earlier this year, and back up to $10 billion — with new corporate-guarantee terms — as of July 1.

According to people familiar with SoftBank’s plans who spoke to the Japan Times, a successful U.S. neocloud operation could triple or quadruple SoftBank Corp.’s annual operating income, raising it to somewhere between $18.5 billion and $25 billion — what Miyakawa described as profit “on a different order of magnitude” from SoftBank’s current telecom business.

What This Means for the Sector — and for Enterprise Buyers

For investors, SB Neo’s arrival simultaneously validates and complicates the neocloud investment thesis. It validates the thesis by confirming that a well-capitalized conglomerate with access to federal land and gas-fired power believes 10GW of AI cloud demand is real and durable. It complicates the thesis by adding a well-capitalized new competitor at a moment when existing players are already watching their biggest customers consider becoming their rivals.

For enterprise buyers and AI developers evaluating compute vendors, the SB Neo announcement is premature: commercial services do not begin until fiscal year 2027, meaning the earliest real service availability is April 2027, with the Portsmouth data center not expected online until around 2028. Today, SB Neo is a strategic signal, not a procurement option.

The underlying question the sector is working through — whether vertical integration at the power and infrastructure layer creates a defensible position, or whether hyperscalers will simply build their own everything eventually — will not be answered before SB Neo’s first customer goes live. What the July 2 announcement established is that SoftBank believes the answer is yes, and is committing the capital, the federal relationships, and the software development resources to prove it.

Whether that conviction is better founded than McKinsey’s skepticism is what the next two years will determine.

Frequently Asked Questions

What is SB Neo and when does it launch?

SB Neo, Inc. is a U.S. neocloud subsidiary established jointly by SoftBank Corp. (51%) and SoftBank Group Corp. (49%), incorporated in Delaware in July 2026. It plans to begin commercial GPU cloud services for American enterprises and hyperscalers in fiscal year 2027, the period ending March 31, 2028. Its long-term target is 10 gigawatts of deployed AI data center capacity by around 2030.

What is a neocloud and how does it differ from hyperscalers like Amazon Web Services or Google Cloud?

A neocloud is a specialized cloud provider built exclusively around GPU compute for AI training and inference workloads, rather than offering the broad software ecosystems and managed services that hyperscalers provide. Neoclouds typically offer bare-metal GPU access at pricing that has historically run 50–85% below hyperscaler rates for equivalent silicon. The tradeoff is fewer integrated services, fewer global regions, and greater vulnerability to price erosion as the GPU rental market matures.

Why did CoreWeave stock fall nearly 48% from its 52-week high?

CoreWeave’s stock decline reflects both market-wide pressure on neocloud stocks and a specific selloff triggered by Bloomberg’s July 1, 2026 report that Meta Platforms was developing infrastructure to sell its own AI computing capacity. The concern is structural: Meta is simultaneously one of CoreWeave’s largest customers — with a $21 billion committed contract through 2032 — and potentially a future direct competitor in the same GPU rental market. CoreWeave also carries approximately $24.9 billion in total debt and reported a $740 million net loss in Q1 2026, leaving limited financial cushion for competitive disruption.

What makes Infrinia AI Cloud OS different from standard cloud infrastructure software?

Infrinia AI Cloud OS automates the full GPU infrastructure stack — from hardware-level BIOS and RAID configuration through the operating system, GPU drivers, networking, Kubernetes controllers, and storage — removing the need for specialized engineering teams to configure each layer manually. It delivers Kubernetes-as-a-Service for multi-tenant GPU environments and Inference-as-a-Service via OpenAI-compatible APIs, so existing AI applications can connect without code changes. SoftBank designed the stack specifically for NVIDIA GB200 NVL72 hardware, which interconnects 72 Blackwell GPUs within a single rack at 1.8 terabytes per second via NVLink.