SoftBank Corp. Pursues Telco AI Cloud to Become AI Infrastructure Provider

SoftBank has launched a program called Telco AI Cloud that combines a centralized GPU cloud for large‑scale model training with an AI‑RAN driven MEC layer for low‑latency inference, fronted by a software stack named Infrinia AI Cloud OS. The company described orchestration via an AITRAS Orchestrator and a Dynamic Scoring Framework that schedules workloads across distributed sites based on latency targets, resource availability and power constraints. SoftBank is positioning the stack to support operator‑hosted edge inference, industrial on‑prem deployments and an open AI‑RAN ecosystem by opening key orchestrator components and prioritizing O‑RAN interworking. Demonstrations and partner agreements revealed in late February and early March 2026 indicate the program is targeting near‑term pilots rather than purely long‑term research.

Concurrently, a separate industry track has surfaced: the GSMA’s Open Telco AI initiative emphasizes a model‑and‑dataset‑centric approach, publishing validated telecom model libraries, fine‑tuning datasets, and a public leaderboard of seven telecom‑specific benchmarks to make model evaluation reproducible across operators. Founding contributors named at launch include AT&T and AMD, and compute routing support was noted for partner TensorWave; GSMA’s portal is designed to seed reproducible baselines and community activity such as competitions and synthetic data pipelines. That effort is explicitly complementary to, but different from, an NVIDIA‑anchored consortium that publicly focuses on embedding accelerators, telemetry pipelines and low‑latency inference primitives into radio and edge stacks.

The coexistence of these approaches creates a practical tension: benchmark‑and‑model centricity (GSMA) aims to standardize objective metrics and datasets, while the NVIDIA consortium pushes reference implementations that tightly integrate NPUs/DPUs and telemetry into operator stacks. SoftBank’s Telco AI Cloud sits between these poles — it signals openness at the software and interface level while simultaneously depending on deterministic hardware behavior and telecom‑grade accelerators to meet latency and safety commitments. This means real‑world operator pilots will likely act as demand signals that steer semiconductor and orchestration roadmaps toward the accelerators and form factors that satisfy both reproducible model performance and hard latency/safety requirements.

Operationally, SoftBank’s architecture separates centralized training pools from edge inference and introduces a scoring/orchestration layer designed to factor energy and site heterogeneity into placement decisions; but heterogeneous hardware, site power limits and GPU scheduling remain top practical risks. From a regulatory and safety perspective, embedding learned models into RAN control loops raises questions about spectrum safety, deterministic latency guarantees and model auditability — constraints that will push participants toward staged field trials, curated datasets and joint lab validations before broad commercial rollout. If operators can coordinate standards, benchmarks and hardware references, value may shift to carriers offering managed, low‑latency inference; if not, the market risks bifurcating between hyperscaler managed edge services and narrowly focused industrial on‑prem deployments.