Conversion plan and available power
Nvidia is pursuing a plan to convert aging telephone exchange buildings in the United States into AI data centers and establish a distributed network across 20 states. The strategy aims to reduce the time, costs and permitting burdens associated with building new data centers while using the power and telecommunications infrastructure remaining at existing communications facilities.
The Information reported on November 5, local time, that Nvidia was discussing a plan with private equity firm Apollo Global Management to convert existing facilities belonging to Brightspeed, an Apollo-owned broadband telecommunications company, into AI data centers.
Brightspeed owns more than 130 buildings housing voice and data communications equipment across the United States. Nvidia determined that many of these facilities already receive enough power to operate the latest AI servers and that approximately 1 gigawatt (GW) of power could be available in total. That would be enough power to operate AI chip systems worth tens of billions of dollars.
Nvidia sent Brightspeed a business proposal and a request for additional information on November 4. It is also reported to have discussed with suppliers a plan to renovate some facilities and install next-generation Vera Rubin server racks starting in the first quarter of 2027. It is also reviewing the introduction of technology from AI chip startup Groq, for which it secured a technology license late last year.
Permitting and network operations
Even with existing facilities, however, equipment improvements such as chip-specific cooling systems would be needed, requiring some permits or approvals from local governments. Nevertheless, the process could be simpler than securing land, connecting to the grid and building new data centers.
The plan also responds to the possibility of increased regulation and community opposition to new data center construction following the US midterm elections. With some gubernatorial candidates in states including Texas calling for a moratorium on new data center construction, renovating existing telecommunications facilities could avoid some of the regulatory burdens associated with developing new sites.
Nvidia reportedly described the facilities as an "inference network." This means the focus would be on running models that have already been developed rather than training large AI models from scratch. Dedicated software for centrally managing servers distributed across multiple locations is also expected to be needed.
Cloud companies and AI developers including SpaceX, Anthropic and OpenAI have been mentioned as potential customers. These companies are seeking additional Nvidia servers to operate AI services but face shortages of chips and data center power. Whether Nvidia and Apollo are actually discussing leases with these companies has not been confirmed.
Commercial prospects and operating constraints
Apollo recently agreed to purchase Vera Rubin chip systems worth tens of billions of dollars to lease to SpaceX. This project could also combine data center assets, chip supply and leasing operations.
Reusing existing telecommunications facilities also has operational limitations. Jack Smith, CEO of AI networking startup Datum, cited the ability to use existing underutilized power as an advantage but noted that aging electrical equipment could make operating costs higher than those of newly built data centers.
The project can also be interpreted as a move by Nvidia to secure power and data center capacity in advance, rather than merely supplying AI chips. If grid connection delays or community opposition slow new data center construction, AI chips already secured could remain idle without power. A distributed network using existing telecommunications facilities would, in this interpretation, be an attempt to reduce such risks while accelerating AI infrastructure expansion.
Original Korean article: aitimes.com
English brief by Korea Wind Intelligence, koreawindintel.com
