What's happening?
According to The Information, NVIDIA is in talks with internet and telecommunications provider Brightspeed to convert its buildings into small data centers. The idea was first pitched to executives at Brightspeed's parent company, Apollo Global Management, which then asked NVIDIA to submit written plans.
NVIDIA has since requested additional information, data and site visits to some of Brightspeed's existing infrastructure. The company estimates these sites could collectively deliver around 1 gigawatt of compute capacity.
Why old telco buildings?
Brightspeed has around 130 buildings across the country. Many were already built to handle data and voice traffic and have the power infrastructure needed to run NVIDIA servers. However, the sites are not plug-and-play: cooling systems and supporting architecture would still need to be installed, and local government permits would be required.
Still, the approach is expected to face less red tape and opposition than building a data center from scratch.
Capacity and chips
According to the report, NVIDIA estimates the sites could accommodate more than 2,200 Supermicro Vera Rubin NVL72 racks, each containing 72 Rubin GPUs and 36 Vera CPUs.
Apollo Global Management has already made significant bets on AI hardware. Under a $40 billion deal with SpaceX, it would receive around 360,000 Vera Rubin GPUs across 5,000 NVL72 racks. It is unclear who will own the chips installed at Brightspeed facilities.
Smaller, distributed data centers
While large, homogeneous clusters in one location are preferable for training, NVIDIA representatives describe these smaller sites working together as an "inference network" to run existing AI models. This matters more as AI agents, which use five to ten times more tokens than humans, become more common.
OpenAI and Anthropic are also considering multiple smaller data centers. In the U.S., midterm votes on data centers mostly went in favor of restrictions, and more than 300 bans across the country are making new developments harder.
