The new Telco AI Cloud architecture integrates large-scale GPU data centers with edge AI-RAN
In sum – what we know:
- Technical architecture – Combines large-scale GPU data centers for training with edge AI-RAN for real-time inference, managed by Infrinia AI Cloud OS.
- AITRAS platform – Orchestrator acts as a “central nervous system,” dynamically allocating resources between telecom and AI workloads based on real-time demand.
- Strategic goal – Positions network infrastructure as a competitive asset for robotics and data sovereignty, challenging centralized hyperscalers.
SoftBank wants to play more of a role in telco-related AI, and at MWC 2026, it revealed what it calls the Telco AI Cloud — framed as “next-generation social infrastructure,” but really a play to redefine the company from a traditional telecom operator into something closer to an AI infrastructure provider. The initiative rests on three tightly integrated pillars that span the full AI pipeline — large-scale GPU data centers built for large-scale model training, an AI-RAN-powered Multi-access Edge Computing (MEC) platform designed to push inference right to the network edge for low-latency decision-making, and Infrinia AI Cloud OS, a unified software layer that ties cloud and edge management together under one roof.
The big idea SoftBank is betting on here is distribution. Hyperscalers like AWS, Azure, and Google Cloud run their operations out of centralized data center regions. Telco AI Cloud takes different approach — embedding AI infrastructure directly inside the telecom network itself. On paper, this gives SoftBank a structural edge in latency, reliability, and data sovereignty, all of which matter enormously when it comes to real-time applications like industrial automation. Whether that structural edge turns into a genuine competitive advantage remains to be seen.
Of course, there’s a wide gap between unveiling a vision and shipping it at scale. AI-RAN as a category is still early, with real technical obstacles still in the way, and SoftBank is essentially wagering that its existing network footprint can be transformed into something it was never originally designed to be.
The role of AITRAS orchestration
Sitting at the core of this architecture is AITRAS, which is SoftBank’s proprietary AI-RAN product, paired with what the company calls the AITRAS Orchestrator. The orchestrator’s job is to monitor compute demand in real time across two domains that have historically lived in completely separate worlds — AI processing workloads and Radio Access Network control. It looks at resource availability, application requirements, and projected power consumption, then dynamically shifts compute to wherever it’s most needed.
The interesting part is that AITRAS doesn’t treat the RAN as some separate, siloed telecom function — it treats it as another AI application. Instead of maintaining rigid boundaries between network control and inference tasks, the orchestrator manages everything from a single resource pool. SoftBank’s framing is that this cross-domain control turns the network into a “central nervous system” for computation — one that can fluidly reallocate capacity between things like wireless signal processing during rush-hour traffic and robotics inference models when demand drops off.
None of this is trivial to engineer. Dawid Mielnik, General Manager of Telco at Software Mind, makes an important distinction about where AI-RAN actually stands today, noting that “the problem is the industry is using one label for two completely different things, and nobody’s being honest about which one they mean. AI-assisted RAN — ML models doing energy optimization, traffic steering, beam management inside existing infrastructure — that’s real and commercial. Operators are getting 15–30% energy savings through intelligent sleep modes. It’s in production. It works.”
The more ambitious flavor, or AI-native RAN, is a little different — it involves traditional signal processing gets replaced wholesale by AI models. As Mielnik puts it, “the NVIDIA-SoftBank program is serious, I’m not dismissing it. But it’s one operator, it needs GPU clusters with power and cooling requirements that frankly don’t exist in most base station environments right now.” The dynamic orchestration SoftBank is pitching is a real and worthwhile goal, but the physical infrastructure needed to support it across an entire fleet of base stations hasn’t caught up to the vision yet.
Use cases
The headline use case SoftBank is pushing for Telco AI Cloud is what it calls “Physical AI” — which is essentially the intersection of artificial intelligence and robotics. The company has teamed up with Yaskawa Electric Corporation to deploy robots in real-world settings, and it ran a proof-of-concept with Ericsson showing how robots with limited onboard GPU power can offload heavier AI model processing to mobile edge GPUs over the network.
On the strategic side, SoftBank is leaning into data sovereignty as a differentiator. By keeping AI processing within domestic network infrastructure instead of routing it through foreign-owned hyperscaler clouds, the company positions itself squarely for security-conscious enterprises and government customers. The distributed architecture also tackles scalability from a completely different angle than what hyperscalers offer. Rather than funneling all inference through a handful of massive centralized facilities, SoftBank can spread workloads across edge locations already woven into its existing network. That doesn’t remove the need for centralized compute — those gigawatt-scale GPU clouds exist for a reason — but it creates a complementary layer that hyperscalers genuinely can’t replicate without carrier partnerships.
