SoftBank bets on sovereign AI infrastructure as Japan pushes back against hyperscaler dependence

SoftBank has set an October 2026 commercial launch date for its AI Data Center GPU Cloud, a service built around NVIDIA’s GB200 NVL72 hardware and a proprietary software layer called Infrinia AI Cloud OS. As part of the rollout, a beta version went live alongside the announcement, with SoftBank running it internally across its group companies first.

The practical pitch is straightforward. Japanese enterprises that want to train models or run inference workloads without routing data through US-based hyperscaler regions now have a domestic option. In recent years, regulatory pressure on cross-border data flows has tightened across Asia-Pacific, and as a result, demand for compute that stays within a defined jurisdiction has grown considerably. SoftBank is not subtly positioning itself here. Rather, sovereignty is the main argument.

Infrinia AI Cloud OS handles the orchestration layer, covering Kubernetes-based cluster management for multi-tenant environments. Additionally, it offers an inference API service that lets users deploy a model simply by selecting it, without touching the underlying infrastructure. Meanwhile, the GB200 NVL72 racks provide the compute, with Blackwell GPUs connected via NVLink, which matters for memory-intensive LLM workloads that strain conventional network setups between nodes.

Junichi Miyakawa, President and CEO of SoftBank Corp, framed the launch in terms of where AI competition is heading. “The source of competitiveness is expanding beyond AI itself to include the computing power and operational software that support it,” he said.

What makes SoftBank’s position unusual is how many layers of that stack it is trying to control at once. Earlier this year, the company contributed $30 billion to OpenAI’s $110 billion funding round and holds one of the largest stakes in the company. Beyond that, it is spinning out a new entity called Roze, targeting a $100 billion US IPO focused on AI infrastructure construction. On the energy side, SoftBank also launched a battery business in Japan to support the power demands that large-scale GPU deployment creates.

Furthermore, the telco angle adds another layer. Because SoftBank is connecting this GPU cloud to its AI-RAN infrastructure, it can optimize workloads closer to the network edge, something hyperscalers without telecom assets cannot easily replicate. Whether Japanese enterprises ultimately choose this over established alternatives remains open, but SoftBank has clearly decided the bet is worth making.