Nvidia has approached Apollo about turning old telephone buildings into a network of small AI data centers, according to The Information. The sites belong to Brightspeed, the Apollo-backed telecom company, and could span as many as 20 states. The pitch landed this week, right after a midterm election in which data center moratoriums became a campaign issue.
This is still a proposal, not a deal. Details here come from secondary coverage of The Information’s report, and that coverage shows no public confirmation from the companies.
The Key Points
- The plan: Convert some of Brightspeed’s 130-plus telecom buildings into small AI data centers.
- The scale: Nvidia reportedly estimates the sites could offer roughly 1 gigawatt of total power.
- The status: Nvidia reportedly sent Brightspeed a written proposal on Wednesday and asked for more data and site visits.
- The motive: Coverage links the idea to growing local opposition to new data centers.
- The unknowns: It’s unclear who would buy or lease the chips, or how many buildings Nvidia wants to convert.
Why This Matters
For the past few years, the AI buildout has mostly meant bigger campuses. Nvidia’s own DSX blueprint, announced in October 2025, covers AI factories from 100 megawatts to multiple gigawatts. Its OpenAI partnership targets at least 10 gigawatts.
That approach needs huge sites, lots of new power, and local approval. The approval part is getting harder. Democratic candidates in states including Texas called for moratoriums on new data centers ahead of Tuesday’s vote, according to coverage of the report.
What stands out here is who’s making the move. Nvidia sells chips, so it doesn’t normally need to care where a data center goes. But if communities block new sites, there’s less room to install its hardware. Going after existing real estate suggests Nvidia now sees local politics as a threat to its own sales.
The Math Is Interesting
One gigawatt across 130-plus buildings works out to about 7 or 8 megawatts per site. That’s my rough math, and it assumes every building converts, which the report doesn’t say. Still, it shows the shape of the idea: lots of small sites instead of one giant campus.
Small, scattered sites wouldn’t suit training frontier models, which needs massive clusters packed tightly together. They make more sense for inference, the work of running models after they’re trained. Phone buildings also tend to sit close to the people they serve, and that can cut response times.
This fits a pitch Nvidia has made to telecom operators before. Its “AI grid” concept describes running inference across distributed network locations, including cell sites and mobile switching offices. The Brightspeed idea would be a much more concrete test of that thinking.
Don’t Assume It’s an Easy Fix
Old buildings don’t automatically solve the backlash. Some social posts have claimed these sites already have the power and zoning they need, but that framing doesn’t come from the report. The accessible coverage doesn’t say whether converted buildings would need new approvals.
There are practical questions too:
- Power and cooling: Phone equipment wasn’t designed for dense GPU racks, so retrofits may need serious upgrades.
- Local reaction: A familiar building may draw less attention than a new campus, but neighbors could still object to a noisier, more power-hungry one.
- Customers: Without a clear buyer or tenant for the chips, the economics remain an open question.
What to Watch Next
Markets are clearly nervous about data center regulation. Nvidia shares reportedly closed down 9%, their worst session since January 2025, and at least one report ties the drop to those fears. Treat that link as one outlet’s read, since big stock moves rarely have a single cause.
For practitioners, the immediate takeaway is about planning. If you’re mapping out inference capacity, regional and edge options may become more important as big projects face more local resistance.
The next signals are straightforward: site visits, a named customer, and whether other telecoms get the same call. Coverage suggests AT&T alone may have about 5,000 such buildings. If Brightspeed works, that’s a far bigger pool of real estate. More details are available in The Information’s original report.