Nvidia is in talks with telecom provider Brightspeed to convert several of its buildings into small data centers, after pitching the idea to Brightspeed's parent company, Apollo Global Management, according to The Information.
The Information reported that Nvidia pitched the plan to Apollo executives, who asked it to submit written plans to the internet provider. Nvidia has since requested additional information, data and site visits to some of Brightspeed's existing infrastructure.
The proposal comes as larger facilities face delays and community opposition. More than 300 data center bans have been enacted across the United States, though they have so far delayed only three existing projects, and there are fears the bans will make new developments harder to start.
Many of Brightspeed's roughly 130 buildings already handle data and voice traffic and have the power infrastructure needed to run Nvidia servers. Sources say Nvidia estimates the sites could collectively deliver around 1 gigawatt of compute capacity, accommodating more than 2,200 Supermicro Vera Rubin NVL72 racks, each containing 72 Rubin GPUs and 36 Vera CPUs. The buildings would still need cooling systems and other supporting architecture, and local government permits would be required, though likely with less red tape and opposition than building from the ground up.
It remains unclear who would own the AI chips installed in the facilities. Apollo has already made significant bets on AI hardware, including a $40 billion deal with SpaceX that would give it around 360,000 Vera Rubin GPUs across 5,000 NVL72 racks, and the Brightspeed proposal could potentially help Apollo find a place for the Nvidia chips it ordered and get them running sooner than waiting years for a new data center.
Nvidia representatives have described the smaller sites working together as an "inference network" to help run existing AI models. OpenAI and Anthropic have also been considering similar setups of multiple small data centers. This is especially important as AI agents, which use five to ten times more tokens than humans, become more popular.
Analysis
Why This Matters
- Repurposing existing buildings could get thousands of Nvidia GPUs into service far faster than building new data centers from scratch, easing pressure on compute capacity for running AI models.
- The move is a direct response to community backlash and more than 300 local data center bans across the United States.
- An "inference network" of small sites could spread AI compute across many locations instead of concentrating it in giant campuses, changing where and how AI services are hosted.
Background
AI companies need vast computing capacity to train and run models, and large data centers have become harder to build amid local opposition, long approval processes and permitting delays. Nvidia is the dominant supplier of the AI chips that fill those centers. The reported talks with Brightspeed, a telecommunications provider owned by Apollo Global Management, would see existing telephone buildings reused as compact data centers, an approach the report says OpenAI and Anthropic have also been weighing.
Key Perspectives
Nvidia: describes the converted sites working together as an "inference network" for running existing AI models, and sees them as a faster alternative to waiting years for new data center approvals.
Voters and communities: overwhelmingly voted against data centers during the midterm elections, leading to more than 300 local bans and moratoriums across the United States.
What to Watch
- Whether Nvidia's requested site visits to Brightspeed infrastructure lead to a formal agreement and permit applications.
- Which party ends up owning the GPU racks installed in the buildings, a question the report leaves open.
- Whether the small-site approach becomes more common, given that OpenAI and Anthropic are reported to be considering similar setups.
