Nvidia Eyes Old Telecom Buildings for Distributed AI Network

Nvidia’s pitch targets 130-plus Brightspeed sites across 20 states, with Vera Rubin server racks potentially arriving as early as early 2027

Annemarije de Boer Avatar
Annemarije de Boer Avatar

By

Image: Deposit Photos

Key Takeaways

Key Takeaways

  • Nvidia pitched Apollo Global to convert 130-plus Brightspeed telecom buildings into distributed AI inference sites.
  • Existing telecom infrastructure offers roughly 1 gigawatt of combined electrical capacity across up to 20 states.
  • Ownership, financing, and site readiness remain unresolved, with no confirmed agreement among the three parties.

Communities have pushed back on new data center construction over power consumption, land use, water demands, and noise. According to reporting from The Information, Nvidia has approached Apollo Global Management about converting more than 130 legacy telephone and telecom buildings owned by Brightspeed, an Apollo-backed broadband provider, into a network of smaller AI facilities.

The proposed footprint could span as many as 20 U.S. states.

Running AI, Not Training It

The pitch centers on inference, the part of AI that most people actually interact with every day.

Inference means running an already-trained model to handle user requests, not the computationally intensive work of building a model from the ground up. Nvidia has described the concept as an “inference network”: a geographically distributed system for running trained AI models closer to where demand exists.

The company reportedly submitted a written proposal to Brightspeed requesting facility information and site visits. Per reporting from The Information, Nvidia staff discussed equipping selected buildings with Vera Rubin server racks as early as the first quarter of 2027.

Why These Buildings

Legacy phone facilities carry something newer greenfield sites cannot: existing infrastructure.

Central offices and telecom buildings were engineered for reliability and continuous connectivity. Many already carry electrical capacity, fiber runs, and communications infrastructure that could reduce the need for new grid connections and potentially ease some permitting steps compared to purpose-built campuses, though any advantage would vary by site and jurisdiction.

The aggregate reported electrical capacity across the Brightspeed portfolio is approximately 1 gigawatt. That figure represents combined potential across many sites, not capacity confirmed as deployable today. To put it plainly: a building with existing power feeds is not the same as a building ready for high-density AI servers.

What Still Has to Happen

Existing power is not the same as usable power, and the list of required work is long.

Each building would require site-level validation. That includes cooling systems capable of handling high-density AI servers, upgraded electrical distribution, fire protection, backup generation, floor-load assessments, and regulatory sign-off.

Reusing existing buildings may limit some zoning hurdles, but it does not eliminate the approval process. Permits and upgrades would still be required at converted sites, with the scope depending on each building’s condition, local code, and proposed equipment.

What Remains Unresolved

The commercial structure of the deal is unclear, and no agreement has been confirmed.

Reporting does not establish who would own or lease the equipment inside these buildings. It is also unclear which customers would use the capacity or how many of the 130-plus sites would actually be converted.

How the project would be financed, and whether Brightspeed has formally agreed to proceed, remains unanswered. Nvidia, Apollo, and Brightspeed have not publicly confirmed the proposal.

The Bigger Stakes

The outcome could reshape how the industry approaches AI infrastructure, or confirm that purpose-built campuses remain the only practical path.

If the retrofit model proves workable, it could encourage conversions of former central offices, industrial facilities, and other underused buildings into regional AI nodes. If site-specific costs prove unworkable, operators may find that large greenfield campuses remain more practical despite the community friction they generate around AI data centers.

That answer depends on details that are not yet public.

Share this

At Gadget Review, our guides, reviews, and news are driven by thorough human expertise and use our Trust Rating system and the True Score. AI assists in refining our editorial process, ensuring that every article is engaging, clear and succinct. See how we write our content here →