Nvidia is reportedly exploring a network of smaller AI data centres in former telephone facilities, with more than 130 Brightspeed properties under consideration. If pursued, the plan could put inference computing closer to customers, using existing sites rather than building every centre from scratch.
NVIDIA Watch analysis
What happened
TechRadar reports, citing The Information, that Nvidia contacted Apollo Global Management about converting Brightspeed properties across as many as 20 US states. The sites are said to offer about one gigawatt of electrical capacity collectively. Nvidia reportedly sent Brightspeed a proposal and sought information and inspections; the number of sites that might be converted, the ownership structure and eventual customers remain unknown.
The reported concept is an inference network, running models that have already been trained, rather than a large campus devoted to training them. TechRadar says renovations could begin as early as the first quarter of 2027 using systems based on Nvidia’s upcoming Vera Rubin platform. That is a reported possibility, not a confirmed schedule or construction plan.
Why it matters
A distributed network of inference sites could bring computing capacity nearer to the organisations and users that need it. Repurposing telecom buildings may also make use of existing power and property, though it does not make the practical work disappear: older facilities may need upgrades, and cooling, equipment and local approvals still matter.
The proposal’s scale is striking, but the commercial picture is still mostly blank. The report does not establish that any customer, including the companies it names as potential users, has agreed to rent capacity.
Our read
This is a consequential infrastructure possibility, not a deal ready for the ribbon-cutting. The interesting idea is Nvidia treating the existing telephone network’s property and power footprint as a potential platform for AI inference. The next test is whether that can become an operating network without the costs and constraints of older sites eating into the advantage.
What to watch
- Whether Nvidia, Apollo or Brightspeed confirms an agreement or gives its terms.
- How many sites are selected, and whether the reported one-gigawatt capacity is usable for the plan.
- Whether construction starts on the reported timetable, and which customers actually sign up.
Discussion spark: If existing telecom sites can host AI inference, should companies prioritise retrofitting them over building new data centres, even when the older facilities may be costlier to run?
Sources and evidence
- Nvidia finds another way to deploy its next-generation AI chips – TechRadar (8 November 2026, 23:15 UTC)
Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.