NVIDIA is pursuing an alternative route to bypass regulatory resistance against building new data centers—by directly retrofitting existing telecom company buildings.
According to a November 5 report by tech media outlet The Information, two people with direct knowledge of the matter said NVIDIA has approached private equity giant Apollo Global Management with a plan to build a distributed network of small AI data centers across as many as 20 U.S. states, relying on the existing buildings of a struggling telecom company. This week's gubernatorial election results have put new data center construction plans in multiple states at risk of potential moratoriums, and retrofitting existing buildings could help NVIDIA and its customers avoid some upcoming regulatory restrictions.
The report noted that people familiar with the matter said NVIDIA has already identified its target assets: more than 130 buildings owned by Apollo-backed broadband internet service provider Brightspeed contain approximately 1 gigawatt of usable power in total, enough to support tens of billions of dollars’ worth of NVIDIA Vera Rubin chip systems. NVIDIA insiders have characterized the proposed network as an “inference network,” meaning its primary purpose would be running existing AI models rather than training new ones.
Potential customers for this plan include cloud service providers and AI developers such as SpaceX, Anthropic, and OpenAI. These companies, under the dual pressure of tight server supply and power resource shortages, are urgently seeking more NVIDIA server compute capacity.
Target Assets: Brightspeed's Buildings and Power
The core target of this retrofit project is Brightspeed's existing telecom facilities. Brightspeed was founded in 2022, when Apollo acquired telecom company Lumen Technologies' local telephone business and broadband assets for $7.5 billion and formed the company.
According to the report, people familiar with the matter revealed that Brightspeed currently operates more than 130 buildings, which house the infrastructure for managing voice and data traffic, and many of the buildings are already equipped with sufficient power infrastructure to run NVIDIA's latest servers. NVIDIA estimates that Brightspeed's geographically dispersed buildings collectively have approximately 1 gigawatt of usable power—a scale that could support billions to tens of billions of dollars’ worth of Vera Rubin chip systems.
However, the company is reportedly on the brink of bankruptcy, which adds a sense of urgency to the disposal of these assets. People familiar with the matter said it was in recent weeks, as bankruptcy expectations for Brightspeed intensified, that NVIDIA accelerated its plans to repurpose the company's assets.
Progress: Written Proposal and Site Visits
In terms of how the plan is being advanced, NVIDIA first pitched the concept to Apollo executives, who then asked NVIDIA to submit the proposal in writing to Brightspeed. According to the report, a third party familiar with the project said NVIDIA formally submitted its proposal to Brightspeed on Wednesday, along with requests for supplementary information, data, and on-site inspections of multiple locations.
NVIDIA employees have already been in communication with suppliers about retrofitting some Brightspeed buildings, with plans to begin deploying Vera Rubin server racks as early as the first quarter of 2027. Two people with direct knowledge also said the plan may incorporate technology from Groq—an AI chip startup whose technology NVIDIA licensed late last year.
On the engineering side, upgrades such as installing dedicated cooling systems for the chips may still require local government permits or approvals, but the process needed would be significantly less than building new facilities from scratch.
Regulatory Context: Gubernatorial Elections and New Data Center Rules
The strategic value of this project has become increasingly prominent after Tuesday's midterm elections. Newly elected governors in multiple states had previously called for moratoriums on new data center construction and pushed for related regulatory legislation, including in Texas, which is critical to NVIDIA. The model of retrofitting existing buildings could help companies avoid some of the regulatory red lines about to take effect.
People familiar with the matter noted that Brightspeed's buildings are mostly located in small and medium-sized cities, including Dothan, Alabama; Killeen, Texas; and Mansfield, Ohio. These types of dispersed facilities differ from large centralized supercomputing centers in their regulatory classification, which helps reduce policy friction.
Market Demand: Compute Hunger from Anthropic and Others
What is driving this plan forward is not only regulatory pressure but also robust market demand. Zac Smith, CEO of AI networking startup Datum, said retrofitting old buildings can unlock previously idle power resources, although such facilities sometimes have higher operating costs due to aging electrical equipment. He said that over the past few months, rapidly rising demand from cloud service providers and AI developers has attracted many to revisit the idea of repurposing “dilapidated 1980s-era telecom data centers.”
Anthropic is particularly noteworthy. Smith revealed that Anthropic has recently been frequently asking him where to find 5 to 10 megawatts of power capacity in old buildings. Anthropic has acknowledged to potential investors that server shortages are one of the top risks facing its business, and has accelerated commitments to lease facilities or servers from cloud service providers and data center developers, worth as much as hundreds of billions of dollars.
As for potential buyers, it remains unclear who will ultimately purchase the NVIDIA chips designated for Brightspeed's facilities. Notably, Apollo recently agreed to purchase tens of billions of dollars’ worth of Vera Rubin chips and lease them to SpaceX, making it one of the world's largest buyers of AI server chips. Whether NVIDIA or Apollo has already approached the aforementioned potential customer groups remains unclear.
Comments