Nvidia eyes 130+ telecom buildings for AI data centers to deliver 1 gigawatt compute capacity

Nvidia Pitches Telecom Building Conversion to Apollo Global Management

Nvidia has approached Apollo Global Management with a proposal to repurpose telecom buildings into AI data centers, specifically targeting properties owned by Brightspeed, Apollo's broadband operator subsidiary1

1

. Following initial discussions with Apollo executives, Brightspeed received a written proposal from Nvidia and has been asked to provide additional information, data, and site visits to evaluate the feasibility of converting existing infrastructure2

2

. The telecommunications company operates around 130 buildings spread across approximately 20 states, many already equipped to handle data and voice traffic with the necessary power infrastructure to support Nvidia servers3

3

. This strategic move comes as Brightspeed reportedly faces financial difficulties and is close to bankruptcy, making the timing potentially advantageous for both parties2

2

.

Brightspeed Telecom Sites Could Deliver 1 Gigawatt AI Compute Capacity

The collective electrical capacity of Brightspeed's properties totals approximately 1 gigawatt, which Nvidia estimates could accommodate more than 2,200 Supermicro Vera Rubin NVL72 racks1

1

. Each rack would contain 72 Rubin GPUs and 36 Vera CPUs, representing substantial AI infrastructure deployment using next-generation AI chips4

4

. While these Brightspeed telecom sites already possess the power capacity to run Nvidia servers, they're not plug-and-play solutions. Significant upgrades remain necessary, including installing cooling systems, power conditioning equipment, and other supporting architecture1

1

. These modifications would still require permits and approval from local governments, though the process would likely encounter less red tape and community opposition compared to constructing data centers from scratch1

1

. Datum Chief Executive Zac Smith noted that older telecom buildings can "unlock underutilized power," though he cautioned that aging equipment may increase operating costs and require additional maintenance and replacement work3

3

.

Distributed AI Inference Network Strategy Emerges

Nvidia representatives have described these smaller facilities working together as a distributed AI inference network rather than massive training clusters1

1

. The facilities would be designed primarily to run AI models after their initial training, using systems based on the Vera Rubin platform, rather than operating as large-scale training campuses3

3

. This distributed setup would handle live AI requests closer to users, offering more flexibility than relying on a few delayed hyperscale facilities4

4

. The approach becomes particularly important as AI agents, which consume five to ten times more tokens than humans, gain popularity1

1

. Potential users could include OpenAI, Anthropic, and SpaceX, though no confirmed discussions about this specific network have been established3

3

. Having multiple small data centers has been under consideration by companies like OpenAI and Anthropic, especially as larger facilities face delays and community backlash1

1

.

Apollo's AI Hardware Investment Creates Strategic Alignment

Apollo Global Management has already committed to significant AI infrastructure investments, including a $40 billion deal with SpaceX that would provide around 360,000 Vera Rubin GPUs across 5,000 NVL72 racks1

1

. While ownership of AI chips installed in Brightspeed facilities remains unclear, the proposal could help Apollo find locations for ordered Nvidia hardware and get systems operational quickly1

1

. This represents a faster solution than waiting years for new data center approvals while facing community resistance even after receiving government authorization1

1

. Nvidia has reportedly discussed beginning renovations with suppliers during the first quarter of 20272

2

. However, key commercial details remain unresolved, including how many properties would ultimately be converted and which customers would lease AI compute capacity3

3

.

Data Center Opposition Drives Alternative Infrastructure Approaches

The strategy to repurpose telecom buildings gains urgency amid growing political resistance to new data center construction. Americans overwhelmingly voted against data centers during midterm elections, resulting in more than 300 data center bans enacted across the United States1

1

. While these bans have only delayed three existing projects so far, concerns persist that moratoriums will make starting new developments increasingly difficult and challenge AI companies' ability to secure necessary AI compute capacity1

1

. Political calls for temporary restrictions have emerged from Democratic candidates in states including Texas2

2

. Retrofitting existing properties may face fewer planning obstacles than entirely new developments, though local approvals would still be required for cooling equipment and other upgrades2

2

. The proposal to repurpose telecom buildings has received renewed attention as companies seek alternatives to traditional AI infrastructure development1

1

. Nvidia shares fell more than 9% in Thursday trading, their largest daily decline since January 2025, amid broader chip stock retreat as investors assessed regulatory uncertainty surrounding computing developments3

3

.