Science & Technology9 Nov 2026 · about 6 min

Nvidia hits back at US data center gloom with plans to repurpose old telephone and telecom buildings owned by Brightspeed into small AI data centers

The brief

Repurposing an old telecom building means adapting an existing property for AI computing. Nvidia would install modern server systems and supporting equipment inside buildings originally used for telephone or broadband operations. The goal is to create smaller regional facilities rather than enormous training campuses. The proposed sites would form an inference network. They would run AI models after those models have completed their initial training. Existing buildings could provide power connections, physical space, and access to established networks, while upgrades would prepare them for newer systems built around Nvidia’s Vera Rubin platform. This approach could reduce the need to start every project on an empty plot. It would not eliminate construction or approvals, however. Local permission would still be needed for cooling and other upgrades. Older equipment could also require more maintenance and raise operating costs.

What does it mean to repurpose an old telephone or telecom building as a small AI data center?

The proposed sites would form an inference network. They would run AI models after those models have completed their initial training. Existing buildings could provide power connections, physical space, and access to established networks, while upgrades would prepare them for newer systems built around Nvidia’s Vera Rubin platform.

This approach could reduce the need to start every project on an empty plot. It would not eliminate construction or approvals, however. Local permission would still be needed for cooling and other upgrades. Older equipment could also require more maintenance and raise operating costs.

How large is Nvidia's proposed network, in terms of properties, electrical capacity, and states covered?

Nvidia has reportedly contacted Apollo Global Management about converting more than 130 Brightspeed properties into smaller AI facilities. The properties are spread across as many as 20 American states. Together, they reportedly offer approximately 1 gigawatt of electrical capacity.

That capacity could support substantial investment in Nvidia’s next-generation server systems. The sites would not necessarily all become data centers. The article says it remains uncertain how many properties would ultimately be converted, which customers would lease capacity, and who would own the equipment.

Nvidia has reportedly sent Brightspeed a written proposal and requested more information. It also sought opportunities to inspect possible locations. Renovations with suppliers could begin in the first quarter of 2027, using systems built around Nvidia’s upcoming Vera Rubin platform, if the plan proceeds.

Which companies are involved, and what roles could Nvidia, Brightspeed, and Apollo Global Management play?

Nvidia is the proposed technology and expansion lead. It reportedly wants to inspect Brightspeed properties and convert them into smaller facilities for AI applications. Nvidia would likely provide or arrange the server systems and oversee the infrastructure strategy, but the article does not confirm final responsibilities.

Brightspeed is a broadband operator backed by Apollo Global Management. Its property portfolio contains the older telephone and telecom buildings being considered. Apollo’s role could involve supporting Brightspeed and participating in discussions, although the article does not establish a final deal or commercial terms.

The arrangement remains at the proposal stage. It is unclear how many sites would be converted, who would own the equipment, and which customers would lease computing capacity. Possible users mentioned include Axiom, Paradox, and SpaceX, but no discussions with them are confirmed.

Why might using existing telecom properties be faster or easier than building entirely new data centers?

Using existing telecom properties could give Nvidia a physical starting point. The buildings are already present, and their locations may connect to established power and communications infrastructure. That could reduce the time and effort needed to assemble every part of a new site from scratch.

The article says retrofitting may face fewer planning obstacles than entirely new developments. Nvidia could inspect Brightspeed locations, install newer server systems, and upgrade cooling and related equipment. The proposed sites would support regional inference, placing computing closer to customers and existing networks.

This is not an automatic shortcut. Local approvals would still be needed for cooling and other upgrades. The buildings’ age could also create maintenance problems. The proposal remains unconfirmed, and Nvidia has not established how many properties would ultimately be converted.

What could happen to local communities and operating costs when older telecom buildings are converted for AI computing?

Converting older telecom buildings could reuse properties that might otherwise remain underutilized. It could bring AI computing facilities into communities across as many as 20 states and support regional services closer to customers. The article does not establish the number of jobs, construction effects, or other local economic outcomes.

Communities would still need to consider changes required for cooling and other upgrades. Those changes could require local approvals. The sites might also add demand for infrastructure, although the article gives no specific local impacts beyond the need for approvals and upgrades.

Operating costs are a key concern. Zac Smith of Datum said older telecom buildings can unlock underutilized power, but warned that aging equipment may require extra maintenance and replacement. That could make older facilities more expensive to run than newer purpose-built sites.

What is an AI inference network, and how is it different from the enormous data centers used to train AI models?

An AI inference network is a group of computing sites that runs an already-trained AI model when users or applications request results. Nvidia reportedly describes the proposed Brightspeed network this way. Its purpose would be to operate models after their initial training is complete.

That differs from enormous data centers built for training. Training requires the model to learn from large amounts of data, while inference uses the finished model to answer requests. The article presents Nvidia’s proposed facilities as smaller regional sites, not giant training campuses. Their locations could place computing nearer to customers and existing networks.

The distinction may make older telecom properties useful to cloud operators and AI developers needing nearby capacity. Possible users include Axiom, Paradox, and SpaceX, but the article does not confirm discussions for this network. The number of sites remains uncertain.

Why do AI data centers require so much electricity, cooling, and network connectivity?

AI data centers require electricity to run servers, storage, networking equipment, and facility systems. AI workloads can use especially powerful server systems, so a site’s available electrical capacity helps determine how much computing it can support. The article reports that the Brightspeed properties collectively offer approximately 1 gigawatt.

Cooling is needed because operating servers produce heat. A facility must remove that heat to keep equipment working reliably. Network connectivity is also essential because AI systems need to receive requests, access data, and return results. The article highlights the value of placing inference capacity near customers and existing networks.

These requirements explain why Nvidia is examining telecom properties rather than only empty land. Existing sites may offer underutilized power and network access, but older equipment can need additional maintenance and replacement. Local approvals would still be required for cooling and other upgrades.

This brief was written by AI from the original reporting and checked by other models. Names, figures and quotes come from the source; read it for full context.