SoftBank to launch GPU cloud powered by Infrinia AI Cloud OS

SoftBank has announced the October 2026 commercial launch of its “AI Data Center GPU Cloud”, a cloud service sitting at the intersection of high-density GPU compute and a purpose-built software layer called Infrinia AI Cloud OS.

Infrinia AI Cloud OS, developed by SoftBank’s own Infrinia team and first unveiled in January 2026, handles Kubernetes as a Service (KaaS) for multi-tenant environments and Inference as a Service (Inf-aaS) for large language model inferencing via API. The claim is that it lowers the total cost of ownership compared with custom-built or in-house stacks, while allowing GPU resources to be managed centrally and automatically. A beta version launched the same day as the announcement, with SoftBank beginning to run the service internally across its group companies.

The underlying hardware (NVIDIA GB200 NVL72 systems deployed inside SoftBank’s Japan-based data centres) provides the compute muscle. The GB200 NVL72 racks Blackwell GPUs connected via NVLink, which makes them well-suited to memory-intensive LLM training and complex inference tasks that would otherwise require significant network bandwidth between nodes.

Junichi Miyakawa, President & CEO of SoftBank Corp, commented, “As AI becomes more deeply integrated into society, the source of competitiveness is expanding beyond AI itself to include the computing power and operational software that support it,” he said.

“SoftBank will provide integrated computing infrastructure and software that can be securely used within Japan as a neocloud provider. ‘Infrinia AI Cloud OS’ and ‘AI Data Center GPU Cloud’ will serve as core services in this initiative, strongly supporting customers’ AI development and real-world deployment.”

Charlie Boyle, vice president of DGX systems at NVIDIA, was equally direct about what the partnership is designed to solve. “The transformation of telecommunications into an AI-native architecture requires a new foundation of AI infrastructure capable of handling the most complex sovereign AI workloads,” he said.

“SoftBank’s deployment of the NVIDIA GB200 NVL72 and ‘Infrinia AI Cloud OS’ gives Japanese enterprises a high-performance, secure, and scalable platform to accelerate their industries.”

Sovereignty, latency and the telco advantage

The sovereignty angle is central to SoftBank’s pitch, and it is one that resonates beyond Japan. Regulatory pressure on cross-border data flows is intensifying across the Asia-Pacific region and in Europe, and the demand for AI compute that stays within a defined jurisdiction is rising accordingly. SoftBank is positioning itself as the answer for Japanese enterprises that cannot, or do not wish to, send training data or inference workloads to US-based hyperscaler regions.

Longer term, the company is linking the AI Data Centre GPU Cloud to its “Telco AI Cloud” vision, announced in March 2026, which aims to integrate AI data centre capacity with its AI-RAN (radio access network) infrastructure. The intent is to optimise AI processing across training and inference workloads by exploiting proximity to the network edge – a capability that hyperscalers, without telco assets, cannot easily replicate. The roadmap calls for a sovereign, distributed AI infrastructure delivering low latency and high reliability, with the telco backbone acting as a differentiator.

The three headline capabilities of the new service are straightforward. The platform supports training and inference on a single GPU pool, handling everything from compute-intensive LLM development to latency-sensitive inference without customers needing to manage separate infrastructure for each workload type. Kubernetes-based orchestration automates scaling, load balancing and failure recovery. And the Inf-aaS layer lets users stand up inference APIs simply by selecting a model, without managing the underlying deployment machinery.

A company moving on multiple fronts simultaneously

The GPU cloud launch does not exist in isolation. SoftBank has been one of the most active investors and builders in the AI infrastructure space over the past 12 months, and the pace has been striking.

In May 2026, SoftBank reported that its quarterly net profit had more than tripled to $11.6 billion, driven in large part by a $25 billion rise in the value of its OpenAI stake. The company is simultaneously spinning out a new AI and robotics entity called Roze, targeting a $100 billion US IPO – a vehicle designed to own the construction layer of the AI infrastructure buildout for the long term.

In February 2026, SoftBank contributed $30 billion to OpenAI’s $110 billion funding round alongside Amazon and Nvidia, cementing its position as one of the defining backers of the world’s most prominent AI company. Reports from January 2026 suggested SoftBank was in discussions to commit a further $30 billion to OpenAI as part of a round that could value the company at approximately $830 billion, underlining just how deeply Son has aligned SoftBank’s future with the AI economy.

On the energy side, SoftBank has launched a battery business in Japan to build next-generation power infrastructure to support electricity demand driven by the AI boom, with the company targeting annual domestic battery revenue of over 100 billion JPY by FY2030. The batteries are designed without lithium or cobalt, a deliberate move to sidestep geopolitical supply chain risk, something that will be of interest to any operator planning infrastructure at scale.

SoftBank subsidiary SB Energy has also partnered with OpenAI under the Stargate initiative, with OpenAI selecting SB Energy to build and operate a 1.2-gigawatt data centre in Milam County, Texas.

Taken together, the picture is of a company that is not simply investing in AI – it is attempting to own the full stack from energy and physical infrastructure through to GPU compute, orchestration software and, through its OpenAI relationship, the models themselves. The Infrinia AI Cloud OS launch is the software piece that ties the compute layer together. Whether SoftBank can execute across that many fronts simultaneously (and persuade Japanese enterprises to commit to a neocloud provider when hyperscaler alternatives remain available) is the question the market will now put to the test.

Sam Altman takes the stand as Musk’s OpenAI lawsuit reaches its pivotal week

SoftBank, OpenAI invest $1bn in SB Energy to expand data centres