NVIDIA has proposed converting existing facilities owned by Brightspeed, a telecom company backed by U.S. investment firm Apollo, into small AI data centers, The Information reported on November 5, 2026. The plan would put roughly 130 Brightspeed telecom facilities, and the roughly 1 GW of power said to be available there, to work as AI infrastructure. However, the figure of about 130 buildings does not mean that many facilities will actually be retrofitted, and the estimate of about 1 GW of power cannot be converted directly into AI computing capacity. Repurposing existing telecom facilities as AI data centers requires looking beyond the power terms to each site's power distribution and cooling capacity, as well as the type of AI workload to be run there.
NVIDIA's Telecom Facility Conversion Idea Is Still at the Proposal Stage
According to The Information's report, NVIDIA approached Apollo about building a network of small AI data centers and submitted a written proposal to use Brightspeed's facilities. This is a report based on information from sources about an undisclosed plan, however. NVIDIA, Apollo and Brightspeed have not officially announced a signed agreement or the start of any facility retrofits. It is also not yet clear whether the plan will become an actual business.
The figure of about 130 buildings describes the scale of Brightspeed's telecom facilities, and it has not been decided that all of them will be converted into AI data centers. There is no concrete information on which buildings would be chosen or how many would be retrofitted. The roughly 1 GW of power is likewise an estimate that NVIDIA is said to have judged to be available, not a number confirmed through facility audits or power supply contracts. The number of GPUs that could be deployed, or the AI processing capacity, therefore cannot be estimated from it. The first step is to confirm the power and equipment conditions actually available at each site.
Reusing Brightspeed's Regional Telecom Infrastructure for AI
To understand the plan, it matters what kind of telecom infrastructure Brightspeed owns. In its announcement of the start of operations on October 3, 2022, the company said that the acquisition of Lumen's regional telecom assets and related businesses by funds managed by Apollo had closed and that it had begun operating. The acquired business spanned 20 U.S. states and mainly offered fiber and copper-based communications services in rural and suburban areas. Brightspeed took over these assets and set out a policy of upgrading regional networks while expanding its fiber network.
The figure of 20 states refers to the scope of assets disclosed when operations began in 2022, and it does not mean the AI data center plan covers all 20 states. Still, the fact that Brightspeed holds telecom infrastructure spread across a wide area is important background for this plan. If existing buildings and telecom networks can be used, it may be possible to repurpose equipment already placed in communities, rather than building new sites for AI processing.
That said, the existence of a telecom facility does not mean the building or its power has room to spare. The report and public materials do not show the operating status of each candidate building, nor do they confirm any history of surplus power resulting from the scaling back or shutdown of copper services. A facility that still supports local residents' communications would need a design that lets AI equipment run alongside existing services. Because this is a plan to reuse telecom infrastructure, power and space would have to be allocated on the premise of coexisting with current telecom equipment, not just adding AI equipment.
Three Conditions for Turning About 1 GW of Power Into Computing Capacity
To judge how much AI processing capacity a power estimate of about 1 GW could translate into, it helps to consider three separate elements: the power allocation, each facility's power distribution and cooling equipment, and the assignment of AI workloads. A useful reference is NVIDIA's technical documentation on DSX Flex, which describes power-use constraints, facility-side equipment, and compute power management and workload assignment as distinct roles. Applying these to the Brightspeed conversion idea, the conditions that need to be checked can be organized as follows.
| Item to check | Role described in NVIDIA's documentation | Information needed to evaluate Brightspeed's plan |
|---|---|---|
| Power allocation | What Is NVIDIA DSX Flex? explains that power-use programs may set usage limits for certain periods | Power actually available at each site, available time windows, contract terms |
| Power distribution and cooling | Facility Systems: the facility side handles power distribution to equipment, cooling and power metering | Each site's power distribution capacity, cooling capacity, and the operating status of existing equipment |
| AI workload assignment | Workloads and Compute Power Management covers assigning where workloads run and managing power at the GPU and rack level | Types of AI workloads to run at each site, required compute resources, equipment operating conditions |
This table organizes items for evaluating Brightspeed's plan based on the capabilities described in NVIDIA's technical documentation. It is not the result of an equipment survey of Brightspeed's facilities, and it does not indicate that the company will adopt DSX Flex. The conditions listed under each item are information to be confirmed about actual facilities in the future, not equipment specifications that have already been published.
First, even if a power allocation exists, not all of it can necessarily be used at all times. If a contract limits the available time windows or maximum power draw, the equipment operation plan has to be built around those limits. Also, being able to deliver power to a building is a different matter from being able to distribute that power stably to equipment such as GPU servers. If a facility's internal power distribution and cooling equipment are insufficient, the power said to be available cannot be fully used for AI processing. The aggregate figure of about 1 GW alone does not tell us how large a computing setup each site can host.
After the equipment is in place, there is also the question of which AI workloads to assign to each site. Managing the power consumption of GPUs and racks is a different role from routing the workloads sent by users to the appropriate sites. Decisions about which workloads to run where must take into account each facility's computing capacity and its power and cooling capability. Even if GPUs are placed at multiple facilities, that alone does not mean the whole can be operated like a single giant computing facility.
These three categories are also not a formula for converting power into a GPU count, nor are they sufficient conditions guaranteeing that a data center can be opened once all are met. To judge whether this plan will lead to actual AI services, it is necessary to confirm concretely not only the scale of available power but also each facility's equipment conditions and the content of the workloads to be run there.
In Distributed AI Data Centers, Matching Sites to Workloads Is Key
One possible use for small AI sites spread across regions is inference, in which a trained AI model processes inputs. For distributed inference infrastructure, Akamai set out a concrete vision in its announcement of October 28, 2025. The company said it would run general inference workloads at the network edge, close to users, and handle more complex inference at centralized sites. Rather than concentrating all AI processing in one place or distributing all of it, the idea is to choose where to run it based on the characteristics of the workload.
Akamai CEO Tom Leighton, drawing on the history of internet infrastructure expanding closer to users, has expressed the view that proximity to users will matter in the next wave of AI as well. This is an argument that geographic distance to users should be considered when deciding where to place AI computing infrastructure. However, placing computing equipment close to users does not by itself guarantee shorter response times or better performance. Actual results depend on conditions such as network paths, computing resources and the AI models being processed.
In that announcement, Akamai set an initial goal of rolling out to 20 sites. That is a plan as of October 2025, however, and not a figure showing that 20 sites are operating now. It is also necessary to distinguish between the scale of a company's existing network and the number of sites that can actually deliver GPU-based AI services. Owning telecom infrastructure spread across regions does not mean all of it can be used as AI computing sites.
Akamai's vision shows that another company has already put forward a concrete version of the idea of assigning work to regionally distributed computing facilities according to the type of processing. But NVIDIA and Brightspeed will not necessarily adopt the same approach, and the reported plan does not make clear what AI workloads would run at each facility. The purpose of reusing existing telecom facilities could be not only to replace large-scale AI training clusters with small sites but also to distribute processing such as inference across regions. What configuration will actually be adopted will have to wait for future announcements.
NVIDIA's idea of converting Brightspeed's facilities explores the possibility of using existing telecom infrastructure as a computing base for AI. But the reported 130 or so facilities and roughly 1 GW of power do not by themselves mean usable AI infrastructure. If specific facilities for conversion become known, it will be important to check each site's power supply conditions and its power distribution and cooling capacity, as well as how the plan coexists with existing telecom services and what AI workloads would run there. If those conditions are met and workloads suited to each region can be assigned, the existing facilities that have supported telecom networks could become new nodes of distributed AI infrastructure.