Nvidia pitches 1GW compute network at Brightspeed sites

By Grace Kim

Nvidia estimates that Brightspeed’s roughly 130 telecom buildings across the United States could deliver about 1 gigawatt of compute capacity. The estimate could accommodate more than 2,200 Supermicro Vera Rubin NVL72 racks. The proposal remains at the talks stage, with Nvidia seeking more site and operating data.

If the company can reuse the buildings’ existing power infrastructure, the distributed model could compress deployment schedules compared with constructing a new data-center campus.

What would Nvidia and Brightspeed build?

According to The Information, Nvidia pitched Apollo executives on converting several Brightspeed buildings into small data centers. Apollo, Brightspeed’s parent, asked Nvidia to submit written plans to the internet and telecommunications provider. Nvidia complied and now seeks additional information, data and site visits to some of Brightspeed’s existing infrastructure.

Brightspeed operates roughly 130 buildings around the country, according to the report. Many already handle data and voice traffic. They also have the power infrastructure needed to run Nvidia servers.

The proposal would create a distributed set of compact sites rather than one large campus. Nvidia’s roughly 1GW estimate applies to the collective potential of the buildings, not confirmed operating capacity at any single site.

Each Supermicro NVL72 rack contains 72 Rubin GPUs and 36 Vera CPUs. More than 2,200 such racks could fit within the estimate, although the figure does not establish when any site would enter service.

What would still need to happen?

The existing buildings can power the proposed GPUs, but Nvidia cannot simply install servers and switch them on. The company still must add cooling systems and other supporting architecture. Each conversion would also require permits and local government approval.

According to the report, converted sites would likely face less red tape and opposition than a data center built from the ground up. That expectation is not guaranteed. Existing communities could still object, and construction of the missing infrastructure could delay deployment.

In practice, the proposal removes the need to establish entirely new sites, but it does not eliminate construction, permitting or local review at each converted building.

Why would Apollo use the sites?

Apollo has made substantial AI hardware commitments. The asset manager is involved in a $40 billion deal with SpaceX that would give it around 360,000 Vera Rubin GPUs across 5,000 NVL72 racks, The Information reported.

Ownership of the chips installed at Brightspeed remains unclear. The proposal could give Apollo a place to activate Nvidia chips it has ordered without waiting years for a new data center to receive approval. Community resistance can complicate that schedule even after officials approve a project.

The timing advantage would depend on completing cooling and support installations while avoiding the prolonged opposition that has delayed some otherwise approved data-center projects. Ownership and operating arrangements therefore matter as much as the initial capacity estimate.

Why an inference network?

OpenAI and Anthropic are also considering multiple small data centers as large facilities encounter delays and public backlash. A single, homogeneous cluster remains preferable for training, according to the report. Nvidia representatives described the cooperating sites as an “inference network” for running existing AI models.

The distinction targets continuously serving deployed models rather than training them from scratch. Distributed sites can place inference capacity in locations where power and telecom connections already exist, while the smaller sites work together as one network for model serving.

AI agents could accelerate demand for that capacity. Agents use five to 10 times more tokens than humans, the report said, and their wider use could increase sustained inference consumption across the network.

The Brightspeed plan remains a proposal. The report did not provide a deployment timetable, commercial terms or a confirmed division of ownership between Nvidia, Brightspeed and Apollo.

What could local politics change?

Renewed interest in reuse followed Americans’ overwhelming rejection of data centers during the midterm elections. U.S. jurisdictions have enacted more than 300 data-center bans. So far, those measures have delayed only three existing projects.

The immediate disruption has been small, but bans and moratoriums could make new development harder. The concern is whether AI companies can secure enough compute capacity to power future models as approval and community resistance constrain new sites.

If approvals and cooling installations proceed, the distributed sites could add inference capacity sooner than new campuses, although Nvidia’s 1GW figure remains an estimate rather than a deployed total.