Nvidia reportedly wants to turn more than 130 old telephone buildings into a network of small AI data centers. On November 5, 2026, The Information reported that Nvidia sent a written proposal to Brightspeed, a broadband company owned by the private equity firm Apollo, and asked for more data and site visits. The buildings total about 1 gigawatt of available power, enough for tens of billions of dollars of Nvidia's Vera Rubin systems, and the network could reach as many as 20 states.
As of this writing we found no statement from Nvidia, Apollo or Brightspeed. The original report is behind a paywall. We read its details through a Yahoo Finance summary that cites it. So this post is a sourced account of a report, with the gaps marked.
What is reported
| Detail | Reported | Confirmed by the companies? |
|---|---|---|
| Nvidia approached Apollo about a network of small data centers | Yes, per The Information | Not in coverage we found |
| Written proposal to Brightspeed on Wednesday, November 4 | Yes | No |
| More than 130 Brightspeed buildings, originally for voice and data traffic | Yes | No |
| Roughly 1 GW of available power across the portfolio | Yes | No |
| Could support tens of billions of dollars of Vera Rubin systems | Yes | No |
| Network could span up to 20 states | Yes | No |
| Retrofits with Vera Rubin racks possibly starting in Q1 2027 | Yes, from talks with suppliers | No |
| Plan could use technology from Groq, licensed by Nvidia last year | Yes | No |
| How many buildings Nvidia wants | Unclear in the report | n/a |
| Who would buy or lease the chips | Unclear in the report | n/a |
Why old phone buildings are attractive
Telephone companies built central offices and switching buildings across towns for decades. They sit on dense power feeds, have fiber nearby and are already zoned for industrial equipment. In the reporting, Nvidia sees that as a shortcut: the sites have existing power connections and zoning, so a retrofit could skip some of the slow steps of building a new campus.
The permits do not vanish. The report says retrofits would still need local approval for upgrades such as cooling. They would simply need fewer approvals than a new-build site. Zac Smith, CEO of AI networking firm Datum, told The Information that older telecom buildings can unlock underused power but may cost more to run because of aging equipment.
Inference, not training
Nvidia reportedly describes the plan as an inference network: sites that run models that are already trained. That fits the buildings. A 130-building portfolio of a few megawatts each is a poor place to train a frontier model, which needs tens of thousands of GPUs tightly linked in one hall. It is a good place to serve answers close to users, where delay matters more than scale.
One detail points the same way: the plan could include technology from Groq, the inference-chip company whose technology Nvidia licensed last year, according to the report. We could not verify that link from a second source.
For Nvidia's wider data center software stack, see our post on Nvidia's AI factory operations agent.
The timing: a workaround after the midterms?
The report says the urgency grew after the midterm elections, when Democratic candidates in states such as Texas called for moratoriums on new data centers. We covered the vote's fallout in our post on AI stocks falling on data center fears, and the campaign context in data center backlash and the midterms.
The Decoder also reported, citing Bloomberg, that voters rejected new data centers in 40 of 41 local ballot measures, and that OpenAI moved the same day to publish voluntary community commitments; see OpenAI's community compact. A retrofit of existing buildings is a different route to the same goal: more capacity, less local fight.
Yahoo Finance's summary of the report also says Nvidia's stock closed down 9% on Thursday, its worst day since January 2025. The article does not link the drop to this story, and neither do we.
Who owns the buildings, and why that matters
Brightspeed is an Apollo-backed broadband provider. What matters here is the ownership chain. The buildings belong to a company that Apollo controls, and the report says that company is nearing bankruptcy. Nvidia's written request went to Brightspeed, while the first approach went to Apollo. That means the decision likely sits with a private equity owner that is also, according to the same report, buying Vera Rubin chips to lease to SpaceX. An owner who both holds the real estate and finances the chips has an obvious reason to listen.
This is also why the story is hard to confirm. Private companies do not file press releases for a proposal that may never be signed. The first public sign would likely be a permit request, a local news story about a site visit, or a regulatory filing by one of the parties.
How a small inference site differs from a campus
The giant campuses in the news, such as Google's planned build in Wyoming that we covered in our Cheyenne report, are sized in hundreds of megawatts and need new power plants or fuel cells. A telephone building is a different scale. Spread the reported 1 gigawatt over more than 130 sites and the average is under 8 megawatts per building. That is our division of the reported totals, not a figure from the report, and the real spread between sites could be wide.
At that size, three design facts follow:
- Fewer racks per site. A few megawatts holds a modest number of rack-scale systems. The value comes from the number of sites, not the size of any one.
- Cooling is the hard part. Old switching rooms were built for racks that drew a small fraction of what a modern AI rack draws. The report's note that retrofits still need cooling upgrades points at the main engineering cost.
- Network matters more. Each site needs a fast link to users and to other sites. Telephone buildings already sit on fiber routes, which is part of their appeal.
None of this makes the plan easy. It does explain why Nvidia would frame it as inference: serving answers from many places is a workload that tolerates small sites, and it is the part of AI demand that grows every time a person uses a chatbot or an agent.
What it means for people who build with AI
For most developers nothing changes this month. Capacity that comes online in 2027 would show up as more supply, which could ease limits and queues for hosted models, but it will not change a price list by itself. If you run latency-sensitive apps, such as voice agents, a nearby inference site could cut round-trip time, but only if your model provider chooses to place workloads there.
For anyone tracking AI infrastructure risk, the useful lesson is that the bottleneck has moved from chips to places to put them. The same pressure showed up in our report on how the market reacted to data center fears, where investors priced in slower buildouts.
What is missing
- Company statements. None found from Nvidia, Apollo or Brightspeed.
- The number of sites. The 130-plus figure is Brightspeed's portfolio, not Nvidia's target.
- Who pays and who uses it. The report names SpaceX, Anthropic and OpenAI as possible users, but says nothing on whether they were approached. Apollo recently agreed to buy tens of billions of dollars of Vera Rubin chips to lease to SpaceX, so Apollo is already a chip-leasing player.
- Brightspeed's finances. It is described as reportedly nearing bankruptcy. A distressed owner might welcome a buyer for its real estate, but the terms are unknown.
- Power in practice. "About 1 GW of available power" is a portfolio total. It does not say how much is usable per site, or how much grid work each retrofit needs.
Why it matters
Supply is tied to power and permits. We tracked the same pressure in Microsoft's tighter oversight of neocloud projects, where delays and redesigns pushed back revenue targets. Chips are only useful if there is somewhere to plug them in.
It shifts where AI runs. Spreading small inference sites across 20 states puts compute nearer to users. For developers, that could cut response delay for apps that serve many local users. It does not change the price per token on its own.
It tests the backlash. Retrofits quiet the "new giant campus" objection, but residents may still object to noise, cooling and power use at a site in their town. The old buildings are often in the center of communities, not in remote fields.
What to watch
- A statement or filing from Nvidia, Apollo or Brightspeed.
- Local permit filings for cooling and electrical upgrades at Brightspeed sites.
- Any named customers for the network.
- Whether the Groq technology appears in Nvidia's Vera Rubin inference products.
Bottom line
A single paywalled report, repeated by finance outlets, says Nvidia has made a written proposal to use Brightspeed's old telephone buildings as small AI inference sites. The numbers are large: 130-plus buildings, about 1 GW, up to 20 states. The commitments are not: no company has confirmed it, the target count is unknown, and the first racks are only discussed for 2027. We will update this post if a company responds.