Nvidia Plans to Convert Old Telephone Buildings into AI Inference Networks, Bypassing Data Center Regulations
nashnova research
Nvidia has submitted a written proposal to Apollo Global Management to convert 130-plus legacy telecom buildings owned by Apollo's Brightspeed into small AI data centers across up to 20 states — a first attempt to solve AI's power and permitting bottleneck through existing real estate and distributed deployment.
Why is Nvidia eyeing old telephone buildings?
Multiple U.S. states are tightening regulation on new data-center construction. Newly elected governors — including in Texas, a state critical to Nvidia — have called for moratoriums or stricter oversight.
This means → the traditional path of building large greenfield data centers faces ballooning approval timelines and political risk. Nvidia needs an alternative that sidesteps the front door.
Retrofitting existing buildings still requires some local permits, but far fewer than new construction. In plain terms = renovating an old building is much easier to get approved than breaking ground on a new one.
How much compute can these old buildings actually deliver?
Apollo's broadband operator Brightspeed owns over 130 buildings originally used to manage voice and data traffic, spread across mid-size cities like Dothan, Alabama; Killeen, Texas; and Mansfield, Ohio.
Nvidia estimates the buildings' combined usable power at roughly 1 gigawatt — equivalent to a nuclear power plant's output — enough to run tens of billions of dollars' worth of Vera Rubin chip systems.
This reflects a hidden advantage of legacy telecom facilities: the electrical infrastructure built decades ago for phone and data traffic means many buildings already have enough on-site power to run Nvidia's latest servers without major grid upgrades.
How far along is the plan?
Nvidia initially pitched the idea verbally to Apollo executives. Apollo asked for a written proposal; Nvidia formally submitted it on Wednesday, requesting access to more site data and on-site inspections.
Nvidia's internal team has already engaged suppliers on retrofitting select Brightspeed buildings, targeting deployment of Vera Rubin server racks as early as Q1 2027.
The plan may also incorporate Groq's AI chip technology — Nvidia licensed Groq's tech late last year. This means → the network is not necessarily locked to Nvidia's own chips; the architecture leaves room for compatibility.
What is an "inference network," and who would use it?
Nvidia is positioning these facilities as an "inference network" — purpose-built to run already-trained AI models and handle user requests, not to train new models.
Potential customers include SpaceX, Anthropic, and OpenAI, all facing a twin squeeze of tight server supply and power shortages at new facilities. It is not yet clear whether Nvidia or Apollo has approached these companies.
Dedicated software would let the operator manage geographically scattered servers as a unified network. Put simply = physically they are small buildings dotted across states, but the software layer treats them as one grid.
Can this actually happen — and what could stop it?
Brightspeed was formed in 2022 after Apollo acquired Lumen Technologies' local phone and broadband assets for $7.5 billion. The company is now reportedly near bankruptcy — how Apollo chooses to dispose of these assets will directly determine whether the plan moves forward.
Datum CEO Zac Smith, who operates AI network infrastructure, noted that older buildings can have higher operating costs than new facilities due to aging electrical equipment. This means → "having existing power" and "being economical" are two different things; retrofit costs are a real obstacle.
Whether the plan advances ultimately hinges on two factors: Apollo's asset-disposition decision, and whether Nvidia can lock in enough end customers to absorb the compute capacity — a critical test of the "inference network" business case.
市场有风险，内容仅供研究参考，不构成投资建议。