Telcos are not adding a GPU product to the catalogue. They are reorganising their capital budgets around GPUs.
In February 2026, SoftBank extended the orchestrator inside AITRAS, its AI-RAN product, so that GPU servers processing its own radio traffic could also run external AI workloads. The integration, built with Nokia Bell Labs, was the first live implementation of a use case the AI-RAN Alliance had only described on paper. A month later at MWC Barcelona, SoftBank described its network as a platform that brokers AI workloads across its distributed edge.
The same month, China Mobile told investors it will cut total 2026 capital expenditure by 9.5 percent, to RMB 136.6 billion, while raising investment in computing power by 62.4 percent. Its computing power and intelligent services businesses already make up 20.2 percent of revenue.
Read those two facts together and the pattern is hard to miss. Telcos are not adding a GPU product to the catalogue. They are reorganising their capital budgets around GPUs. For infrastructure buyers, that means a new class of GPU supplier is arriving: one that owns its hardware, controls the network your data travels over, and answers to national regulators. Whether that is good news depends on what you are buying it for.
Coverage of telco AI tends to blend two very different businesses. Separating them is the first step to evaluating either.
This is a conventional GPU data center that happens to be owned by a carrier. Deutsche Telekom's Industrial AI Cloud in Munich, opened in February 2026 with NVIDIA and partner Polarise, holds up to 10,000 Blackwell GPUs and represents roughly one billion euros of investment. SoftBank's own AI computing platform passed 10,000 GPUs and 13.7 exaflops in July 2025, and the company says it will sell that capacity as a service to Japanese companies and research institutions. SK Telecom launched a sovereign GPU-as-a-service (GPUaaS) platform on more than 1,000 NVIDIA B200s in August 2025.
Commercially these compete with neoclouds like CoreWeave, Nebius and Lambda Labs. Their pitch is sovereignty: data and models stay in-country, under a domestic operator the government already trusts.
AI-RAN puts accelerated compute inside the radio access network (RAN) itself, the base station and baseband layer that connects phones to the network. The same GPU server runs radio signal processing and AI inference. SoftBank's AITRAS runs both on the NVIDIA GH200 Grace Hopper platform. NVIDIA's ARC-Pro reference design, launched alongside a $1 billion investment in Nokia in October 2025, is meant to make the same architecture available to every operator, with T-Mobile US running field trials from 2026.
Here the GPUs are small, numerous and distributed: in cell sites and regional switching centres, not a single hall. The compute is there first to run the network. Selling the spare cycles is the business case.
The list of operators is longer than most buyers realise. ABI Research counted roughly 17 telcos building NVIDIA-based AI factories and projects more than $21 billion in telco GPUaaS revenue by 2030. The table below covers the five with the most concrete public numbers.
| Operator | Platform | Scale | Buyer angle |
|---|---|---|---|
| SoftBank Japan |
AI computing platform + AITRAS AI-RAN | 10,000+ GPUs, 13.7 EFLOPS (2025) | Sovereign Japanese capacity; AI-RAN running external workloads since Feb 2026 |
| China Mobile China |
National AI computing network | 107+ EFLOPS, 1,500 edge nodes (2026) | Domestic chips by 2028; not an option for buyers outside China |
| Deutsche Telekom Germany |
Industrial AI Cloud, Munich | Up to 10,000 Blackwell GPUs (2026) | EU sovereignty; industrial and public sector workloads |
| SK Telecom South Korea |
Sovereign GPUaaS, Gasan AI data center | 1,000+ B200 GPUs (2025) | National AI model programme capacity |
| Singtel Singapore |
RE:AI sovereign AI factory | Not disclosed | Multi-country Southeast Asia coverage |
China Mobile deserves particular attention because of its scale and its constraints. At the 2026 World Artificial Intelligence Conference in July, the company reported more than 107 exaflops of AI compute across multiple 10,000-GPU clusters, 12 regional AI computing centres and 1,500 edge computing nodes. In October 2025 its chairman committed to running its national AI computing network entirely on domestic chips by 2028. For buyers outside China, China Mobile is not a supplier option. It is a signal: when the largest operator in the world shifts its investment mix toward compute, the rest of the industry follows the template.
Almost every AI-RAN press release quotes the same number. Omdia projects the AI-RAN market to exceed $200 billion in cumulative value by 2030, and NVIDIA and Nokia cited it when announcing their partnership.
Dell'Oro Group, the firm that tracks RAN vendor revenue quarter by quarter, published a very different view in June 2026. It projects $35 billion in cumulative AI-RAN revenue from 2026 to 2030, and its base case assumes AI-RAN will not expand the overall RAN market at all. In other words, it is mostly existing RAN spend relabelled. Dell'Oro expects the near-term market to centre on AI-for-RAN, single-purpose deployments and non-GPU architectures. GPU-based RAN specifically is forecast to pass $1 billion only by the end of the period.
The economics behind the bullish case also come from interested parties. When AITRAS launched in November 2024, NVIDIA and SoftBank suggested operators could earn about $5 in AI inference revenue for every $1 of AI-RAN capital expenditure, and SoftBank estimated a return of up to 219 percent per AI-RAN server. Both figures come from the companies selling the platform. They may prove right, but they have not been independently validated, and Dell'Oro's numbers suggest the market is not yet pricing them in.
The practical reading: telco AI factories are real capacity you can buy today. GPU-based AI-RAN capacity sold to third parties is early, small and mostly in trials. Plan around the first, watch the second.
If a carrier offers you edge inference on AI-RAN infrastructure, understand the hardware's first job. The radio network has priority. SoftBank built the AITRAS orchestrator precisely to watch demand for both RAN processing and AI processing in real time and reallocate resources between them. When mobile traffic peaks, the network takes the cycles back.
That makes AI-RAN capacity behave more like spot instances than reserved capacity: attractive when idle, preemptible when the primary tenant needs it. It suits latency-sensitive inference that can tolerate graceful degradation, such as robotics control, video analytics and on-site industrial AI. It does not suit model training, long batch jobs, or anything with a hard availability SLA unless the contract guarantees dedicated slices.
Utilisation is the other open question. One analyst assessment published by RCR Wireless in October 2025 put realistic GPU utilisation at 30 to 50 percent for regional telco edge clusters, against 60 to 70 percent for global hyperscalers. Low utilisation is a problem for the telco but can be an opportunity for the buyer, because idle capacity is negotiable capacity. It fits a broader shift we have been tracking: the GPU shortage era is giving way to a market where buyers with flexible workloads hold more leverage.
Telco GPU announcements are long on GPU counts and short on storage. That is the gap buyers should probe first.
A sovereign AI factory is, by design, regional. If your training data and model artefacts live in a hyperscaler region in another country, every training run on the telco's GPUs starts with a data transfer, and egress fees on the source side can erase the price advantage of the cheaper GPU hour. Data gravity applies to telcos the same way it applies to neoclouds: the compute you can actually use is the compute next to your data.
Edge inference changes the shape of the problem rather than removing it. Models must be distributed to many sites, updated without breaking running services, and fed with local data that may itself be subject to residency rules. Ask where model weights are stored, how they are replicated to edge nodes, and what it costs to move results back to your core environment.
This is where zero-egress object storage earns its place in the architecture. Keeping the canonical copy of training data and checkpoints in an S3-compatible store with no egress fees, such as Backblaze B2, Wasabi or Cloudflare R2, lets you send data to whichever GPU supplier, telco or neocloud, is cheapest or most available that quarter without paying a toll on every move. Sovereignty requirements may dictate that the store sits in-country, which narrows the field, but the principle holds: own the data layer so you can rent the compute layer.
Telcos pass the first vetting question we apply to every neocloud. They own their GPUs, and their support chain does not run through a reseller. Beyond that, the questions are different from a standard neocloud evaluation.
Confirm whether you are buying dedicated AI factory capacity or shared AI-RAN capacity. The pricing model, availability guarantees and preemption terms differ completely.
For any capacity that shares hardware with RAN workloads, get the reclaim policy in writing: notice period, frequency during peak hours, and whether you are billed for interrupted jobs.
Ask for storage pricing, ingress and egress fees, and whether the operator offers private interconnects to the hyperscaler region holding your data. Model the full transfer cost for one training run before comparing GPU hour prices.
In-country hosting is not the same as in-country operations, support and key management. If sovereignty is the reason you are buying, verify each layer.
Benchmark the quoted GPU hour against current market rates for the same GPU. Telcos are new to price competition with neoclouds, and early contracts are often negotiable, especially for committed volume on clusters running below target utilisation.
The carriers spent two decades being called dumb pipes. Now they own the GPUs. Whether they become real clouds depends on whether they solve the data problem, not the compute one.
Compare AWS, Google Cloud, Azure, and alternatives like Backblaze B2 Discover how much you could save in seconds
