Nvidia weighs converting Brightspeed telecom sites for AI inference
Nvidia has approached Apollo Global Management about turning aging Brightspeed telecom buildings into small AI data centers, according to a report. The proposed network could have roughly 1 gigawatt of electrical capacity across as many as 20 U.S. states.

Key takeaways · 4
- 01
Treat the project as a proposal under consideration: Nvidia has asked Brightspeed for information and site visits, but the report does not say a conversion deal is settled.
- 02
Assess the opportunity as inference capacity, not training infrastructure; the facilities are meant to run already-trained models.
- 03
Use the reported 1 GW figure as a potential portfolio capacity, not a confirmed deployment: the number of buildings to convert remains unclear.
- 04
Factor in retrofit constraints: permits are still needed for upgrades such as specialized cooling, even though the report says the approval path is shorter than for new construction.
A proposal, not a confirmed conversion
Nvidia sounded out Apollo Global Management about converting a portfolio of aging telecom buildings into small AI data centers, according to the report.[1] Brightspeed, the Apollo-backed broadband provider, operates more than 130 sites originally built for voice and data traffic.[1] Nvidia submitted a written proposal to Brightspeed on Wednesday and also requested additional information and site visits, the report said.[1] The report does not establish how many buildings Nvidia wants to convert, so the size and scope of any deployment remain open questions.[1]
Inference capacity and a possible timeline
The Brightspeed portfolio is reported to have roughly 1 gigawatt of electrical capacity available for the proposed sites, potentially supporting tens of billions of dollars of Nvidia Vera Rubin servers.[1] Nvidia staff discussed with suppliers starting retrofits with Vera Rubin racks in the first quarter of 2027.[1] That discussion indicates a possible retrofit starting point, not a confirmed schedule for the overall project. The planned facilities would run already-trained models rather than host large-scale training clusters.[1] The design could also incorporate technology from Groq, whose technology Nvidia licensed last year.[1]
Existing power, with retrofit trade-offs
The report says existing telecom sites could offer inference capacity sooner than a new campus because their power capacity is already in place.[1] Repurposing buildings is intended to speed hardware deployment and avoid local opposition to new construction.[1] It would not remove the need for local permits: upgrades such as specialized cooling would still require approval, although the report says the retrofit approval path is shorter than for new construction.[1] Older telecom buildings can unlock underused power, but aging equipment can also make them more expensive to run.[1]
Buyers and market questions remain
The report does not identify who would buy or lease chips installed at the sites.[1] It names cloud providers and AI developers including SpaceX, Anthropic and OpenAI as possible users of the telecom-site capacity.[1] Apollo separately agreed to buy tens of billions of dollars of Vera Rubin chips to lease to SpaceX.[1] Brightspeed was reportedly nearing bankruptcy, adding financial context to the proposal without resolving its outcome.[1] The report also noted that regulatory fears around data-center buildout weighed on the sector; Nvidia shares closed down 9%, their worst session since January 2025.[1]
For teams planning AI infrastructure, the proposal highlights a possible route to inference capacity that uses existing power and buildings rather than a new campus. But the unknown site count and customer base, along with the need for local permits for upgrades such as specialized cooling, make it important to distinguish available capacity from capacity that is funded, approved and ready to use.
Why it matters
Bring this to your team — one session a day, built for your industry.
kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.
For teamsHow this developed
6 November 2026
Nvidia weighs converting Brightspeed telecom sites for AI inference