Nvidia Wants to Turn 130 Old Telephone Buildings Into AI Data Centres

Nvidia, the company whose chips power most of the world’s artificial intelligence, has a new plan to find space for its machines: turn more than 130 old telephone buildings across the United States into small AI data centres.

The proposal, first reported by The Information on 5 November 2026, would convert buildings owned by Brightspeed, a broadband provider owned by the private equity firm Apollo Global Management, into a network of small data centres spread across as many as 20 American states. Nvidia estimates the buildings could offer around 1 gigawatt of available power — roughly enough electricity to power hundreds of thousands of homes.

The plan is still only a proposal. No deal has been signed, no buildings have been converted, and Nvidia has not said which customers would use the sites. But the idea shows how far the world’s most valuable chipmaker is willing to go to solve one of the biggest problems facing the AI industry: finding enough power and space to run its chips.

What Nvidia has proposed

According to The Information, Nvidia approached Apollo Global Management about using buildings owned by Brightspeed, an Apollo-backed broadband provider that is reportedly nearing bankruptcy. Brightspeed operates more than 130 buildings that were originally built to manage voice calls and data traffic for local telephone networks.

Nvidia first raised the idea with Apollo executives, who asked the chipmaker to put its plans in writing. Nvidia then submitted a formal written proposal to Brightspeed on Wednesday, 4 November, and has asked for more technical information and visits to several of the sites.

The proposed network could span up to 20 states. Locations mentioned in the report include Dothan in Alabama, Killeen in Texas, and Mansfield in Ohio. It is not yet clear how many of the 130-plus buildings Nvidia actually wants to convert.

Nvidia staff have also discussed the plan with suppliers. According to the report, building work could begin in the first quarter of 2027, with the sites fitted out with Nvidia’s upcoming Vera Rubin server racks. The plan could also use technology from Groq, an AI chip startup whose technology Nvidia licensed last year.

For now, though, the proposal remains exploratory. Nvidia has requested information and site visits, but no final agreement, confirmed customer list, or approved conversion programme has been disclosed.

Why old telephone buildings?

To understand why Nvidia is looking at old telephone buildings, it helps to understand what a data centre needs. A data centre is a large building full of computers that run day and night. Those computers need three things: a lot of electricity, a fast internet connection, and permission to build.

Old telephone buildings, known as central offices, already have all three. They were built decades ago to connect phone calls, so they sit on strong power connections and fibre-optic cables. They are also already built, which means converting one into a data centre can be faster than starting from an empty field.

That speed matters. The AI industry is racing to build computing power, and new data centres can take years to plan, approve, and build. By reusing existing buildings, Nvidia hopes to get its chips working sooner.

The buildings also solve a power problem. Finding a gigawatt of electricity for a new site is very hard in many parts of the United States, because local power grids are already stretched. The Brightspeed buildings already have power connections in place, and Nvidia estimates the portfolio could offer around 1 gigawatt in total. That capacity could support tens of billions of dollars’ worth of Nvidia’s Vera Rubin servers, according to The Information.

There is a wider trend here too. Verizon’s chief executive has said the company’s old central offices are “power-ready, already permitted” as they are decommissioned. AT&T has around 5,000 such buildings. Nvidia’s 130 Brightspeed sites are only a small part of that total, but they come with something the others do not yet have: a concrete proposal and a power estimate.

The data-centre backlash behind the plan

The proposal has taken on added urgency because building new data centres has become politically difficult in the United States.

In the midterm elections held on Tuesday, 3 November, several winning candidates called for moratoriums — temporary bans — on new data centres. Texas was among the states where the issue featured in the campaign. Local communities have raised concerns about how much electricity and water giant data centres use, and about the noise and traffic they bring.

Reporting on the plan noted that more than 300 local bans on data-centre building have been passed across the United States, although they have so far delayed only a small number of existing projects. Even so, the political mood has made tech companies nervous, and investors have noticed.

Nvidia’s share price fell 9 percent on Thursday, 5 November, its worst day since January 2025, as part of a broader sell-off in chipmaker stocks driven by fears about data-centre regulation. The tech-focused Nasdaq index fell 5.5 percent the same day. (Read our earlier report: Wall Street Sell-Off Wipes Out Post-Election Rally as Nasdaq Falls Into Correction)

Retrofitting existing buildings would still need local permits for upgrades such as specialised cooling systems, but The Information noted it would involve fewer approvals than building a brand-new facility. That is the heart of Nvidia’s workaround: use buildings that already exist, in places that already have power, instead of fighting for permission to build giant new ones.

What is an inference network?

Nvidia has described the proposed Brightspeed sites as an “inference network”. That phrase is worth explaining, because it tells us what these data centres would actually do.

AI work comes in two stages. The first stage is training: teaching an AI model by showing it enormous amounts of data. Training needs huge, concentrated computing power, which is why companies build giant data-centre campuses.

The second stage is inference: running the trained model so that real people can use it. Every time someone asks a chatbot a question, generates an image, or uses an AI assistant, that is inference. Inference can be split across many smaller sites, because each request is small and can be handled close to the user.

That is what Nvidia is proposing: a network of small sites that run already-trained AI models, rather than one giant campus that trains new ones. Smaller sites spread across 20 states could serve users across the country while tapping into pockets of electricity that would be too small for a giant training centre on their own.

Other companies are thinking along similar lines. OpenAI and Anthropic have also been considering networks of multiple small data centres, according to The Information. Anthropic has cited server shortages as one of its top business risks and has committed hundreds of billions of dollars to computing leases.

The money and the machines

The sums involved in this plan are large, even by the standards of the AI boom.

The servers Nvidia wants to install are its Vera Rubin systems, the next generation of its AI hardware after the Blackwell chips that dominate the market today. The Brightspeed sites could hold tens of billions of dollars’ worth of these servers, according to The Information. Tom’s Hardware, reporting on the same plan, said the sites could accommodate more than 2,200 Rubin GPU racks.

Apollo’s role goes beyond owning Brightspeed. The investment firm recently agreed to buy tens of billions of dollars’ worth of Vera Rubin chips, which it plans to lease to SpaceX. That makes Apollo one of the largest buyers of AI servers anywhere — but it does not mean SpaceX will use the Brightspeed buildings. The Information said potential customers could include SpaceX, Anthropic, and OpenAI, all of which face server shortages and power constraints, but it is not clear whether Nvidia or Apollo has approached them about this specific project.

It is also not clear who would ultimately buy or lease the chips for these facilities. That is one of the biggest unanswered questions hanging over the proposal.

Brightspeed itself is in a difficult financial position. People familiar with the effort told The Information the company is reportedly nearing bankruptcy, and reports say it may seek bankruptcy protection in December. The company was formed in 2022 after Apollo acquired parts of Lumen Technologies’ local telephone and broadband businesses for $7.5 billion.

Nvidia’s own week has been turbulent. Its shares fell sharply on Thursday, but rose more than 4 percent in pre-market trading on Friday, 6 November, as investors reacted to the Brightspeed report. The company is also dealing with a separate problem in China, where around 600,000 of its advanced H200 AI chips are stuck in Hong Kong awaiting import approval from Beijing. (Read more: 600,000 Nvidia H200 AI Chips Stuck in Hong Kong as China Holds Back Import Approval)

What could go wrong

The plan sounds neat on paper, but there are real obstacles.

First, old telephone buildings were not designed to house modern AI servers. They may need specialised cooling systems and electrical upgrades, which cost money and take time. Zac Smith, chief executive of AI networking firm Datum, told The Information that older telecom buildings can unlock underused electricity, but they can sometimes be more expensive to run because of ageing electrical equipment.

Second, converting the buildings would still need local government permits for the upgrades, even if fewer than a new build. Local opposition to data centres is about power, water, and noise — and a converted telephone building uses power and water too.

Third, the numbers are still uncertain. Nvidia’s 1-gigawatt estimate is just that — an estimate. Nobody has confirmed how much power each building can actually deliver once it is fitted with modern servers, or how many of the 130 buildings are suitable at all. The Information also said it is unclear how many buildings Nvidia wants to retrofit.

Fourth, there is the customer question. A network of data centres is only useful if someone pays to use it. Nvidia has not named a single customer for the Brightspeed sites, and the companies mentioned as possible users — SpaceX, Anthropic, OpenAI — have not been confirmed as being involved.

Finally, there is Brightspeed’s own future. If the company files for bankruptcy protection, as reports suggest it might, that could complicate any deal, even though the buildings themselves would still exist.

What happens next

The next steps are practical rather than dramatic. Nvidia has asked Brightspeed for more data and site visits. Engineers will need to inspect the buildings, check their power connections, and work out what upgrades each one would need. Suppliers have discussed starting retrofits in the first quarter of 2027, but no construction schedule has been confirmed.

Whether the plan goes ahead will depend on what those inspections find, whether the economics work, and whether customers sign up. If it does go ahead, it could become a template for the whole industry: instead of building ever-bigger data-centre campuses, tech companies could spread their computing across hundreds of small, converted buildings that already have power and fibre.

That would be a significant shift. For years, the AI race has been about building bigger — bigger chips, bigger data centres, bigger power deals. Nvidia’s Brightspeed proposal suggests the next phase of the race might be about building smarter, by reusing what already exists.

For now, though, it is only a proposal. The old telephone buildings are still telephone buildings, and the AI data centres inside them exist only on paper.

Frequently asked questions

What is Brightspeed?

Brightspeed is a broadband provider owned by the private equity firm Apollo Global Management. It was formed in 2022 after Apollo acquired parts of Lumen Technologies’ local telephone and broadband businesses for $7.5 billion. It operates more than 130 buildings originally used to manage voice calls and data traffic.

How many buildings could Nvidia convert?

More than 130 Brightspeed buildings have been discussed, across as many as 20 US states. But it is not clear how many Nvidia actually wants to retrofit, and some buildings may turn out to be unsuitable.

What is a gigawatt?

A gigawatt is a unit of power equal to one billion watts. Nvidia estimates the Brightspeed buildings could offer around 1 gigawatt of available power in total — roughly enough electricity to power hundreds of thousands of homes.

What are Vera Rubin servers?

Vera Rubin is Nvidia’s next generation of AI server hardware, following its Blackwell chips. The servers are designed to run and train advanced AI models, and they are some of the most powerful — and most expensive — computing machines in the world.

What is an inference network?

An inference network runs AI models that have already been trained, so that people can use them — for example, when you ask a chatbot a question. Inference can be spread across many small sites, unlike training, which usually needs one giant data centre.

Has a deal been signed?

No. Nvidia has submitted a written proposal to Brightspeed and asked for more information and site visits. No final agreement, confirmed customer list, or approved conversion programme has been disclosed.

Conclusion

Nvidia’s plan to turn 130 old telephone buildings into AI data centres is a clever answer to a real problem: the AI industry needs more computing power, but building giant new data centres has become slower, more expensive, and more politically difficult.

By reusing buildings that already have power and fibre connections, Nvidia could deploy its chips faster and sidestep some of the opposition facing new construction. The idea of an “inference network” — many small sites instead of one giant campus — could also change how AI computing is built.

But the plan is still only a proposal, and plenty could stop it. The buildings may need costly upgrades, customers have not signed up, and Brightspeed itself is reportedly close to bankruptcy. Whether the old telephone exchanges of America become the data centres of the AI age will be decided by engineers, accountants, and local planning officials — not by press releases.

What is certain is that the race for AI computing power is no longer just about building bigger. It is now also about finding power wherever it already exists — even in buildings that were built to carry phone calls, not artificial intelligence.

Sources

• The Information, via Stocktwits: “Nvidia Eyes Apollo’s Struggling Telecom Sites To Sidestep Data Center Moratoriums: Report” — https://stocktwits.com/news-articles/markets/equity/nvidia-eyes-apollo-s-struggling-telecom-sites-to-sidestep-data-center-moratoriums-report/cNG1vn9R94o
• TechTimes UK: “Nvidia Is Looking at Old Telephone Buildings for AI Data Centres as Demand for Computing Power Grows” — https://www.techtimes.co.uk/nvidia-brightspeed-telecom-buildings-ai-data-centres-1809763
• Tom’s Hardware, via Zot News: “Nvidia reportedly in talks to repurpose Brightspeed buildings as tiny data centers” — https://zot.news/article/nvidia-reportedly-talks-repurpose-brightspeed-buildings-tiny-rx3pj
• FourWeekMBA: “Nvidia’s Phone-Building Workaround: AT&T Has About 5,000” — https://fourweekmba.com/ai-nvidias-phone-building-workaround-att-has-about-5000/

This article is for general information and education only. It is not legal or financial advice.