Nvidia Explores Converting 130-Plus Telecom Sites Into Distributed AI Inference Network

Nvidia has proposed turning more than 130 old telephone buildings into a network of small AI data centers, The Information reported on November 5 [2]. The buildings belong to Brightspeed, a broadband provider owned by private equity firm Apollo Global Management. Together they have roughly 1 gigawatt of available power, enough, according to the report, for tens of billions of dollars of Nvidia's upcoming Vera Rubin systems, and the network could span as many as 20 states [2].

Nvidia first approached Apollo about the idea, then sent Brightspeed a written proposal on November 4 asking for more data and site visits, according to the report [2]. Retrofits with Vera Rubin racks could start in the first quarter of 2027, based on talks with suppliers [2]. The companies have not confirmed the plan.

Why old phone buildings

Telephone companies built central offices and switching buildings in towns for decades. They sit on dense power feeds, have fiber nearby and are already zoned for industrial equipment [2]. That makes them a possible shortcut around the slowest parts of building a new data center campus: power connections and permits.

The shortcut is partial. Retrofits would still need local approval for upgrades such as cooling, though fewer approvals than a new site [2]. Zac Smith, CEO of AI networking firm Datum, told The Information that older telecom buildings can unlock underused power but may cost more to run because of aging equipment [2].

The timing matters too. The Information reported that urgency grew after the November midterm elections, when Democratic candidates in states such as Texas called for moratoriums on new data centers [2]. The Decoder, citing Bloomberg, reported that voters rejected new data centers in 40 of 41 local ballot measures [2]. Reusing existing buildings means less new construction for communities to fight, though residents may still object to noise, cooling and power use at a site that is often in the center of town [2].

Inference, not training

Nvidia reportedly describes the plan as an inference network: sites that run models that are already trained, close to the people using them [2]. That fits the buildings. Training a frontier model takes tens of thousands of GPUs tightly linked in one hall; serving answers works from many smaller rooms [2].

The sites would be small by today's standards. Spread 1 gigawatt over more than 130 buildings and the average is under 8 megawatts each [2], while the largest U.S. data center projects range from 1.2 to 17 gigawatts per campus [1]. At that size the value comes from the number of sites, not the size of any one. Cooling is likely the main engineering cost, because old switching rooms were built for racks drawing a small fraction of what a modern AI rack draws. Network links matter more, and the buildings already sit on fiber routes [2].

The plan could also use technology from Groq, the inference-chip company whose technology Nvidia licensed last year, according to the report [2].

Apollo on both sides

Apollo's position is unusual. It controls the buildings through Brightspeed, which the report describes as nearing bankruptcy, and it recently agreed to buy tens of billions of dollars of Vera Rubin chips to lease to SpaceX [2]. An owner that both holds the real estate and finances the chips has an obvious reason to listen.

Big questions remain open: how many of the 130-plus buildings Nvidia actually wants, who would buy or lease the chips, and whether the possible users the report names, SpaceX, Anthropic and OpenAI, have been approached [2]. The 1 gigawatt is a portfolio total; how much power each site can actually use after upgrades is unclear [2].

What to watch

• A statement or filing from Nvidia, Apollo or Brightspeed
• Permit applications for cooling and electrical upgrades at Brightspeed sites
• Whether Groq technology appears in Nvidia's Vera Rubin inference products

Bottom line

On paper the idea makes sense: the bottleneck in AI has moved from chips to places to put them [2], and old telephone buildings already have power, fiber and zoning. But it is still a proposal. The number of sites, the customers and the cost of retrofitting aging buildings are unknown, and the first racks would not go in before 2027.