Nvidia Eyes Brightspeed's Central Offices for 1GW GPU Inference Network

Nvidia has pitched Apollo Global Management on deploying up to 1GW of GPU compute capacity across Brightspeed's roughly 130 central offices in the US, according to a report by The Information.

The plan would turn the telephone operator's legacy exchange buildings into edge compute nodes for a new inference network. The two companies are reportedly sharing information and arranging site visits. Nothing is signed — but the scale of the number alone tells you this isn't a pilot dressed up as a press release.

The sites would also include hardware from Groq, the AI chip startup Nvidia holds a major technology license with. That detail matters: an inference network built on a mix of Nvidia GPUs and Groq's silicon suggests Nvidia is thinking about capacity first and exclusivity second.

The first capacity at retrofitted buildings could go live as soon as Q1 2027, The Information reports. Getting there won't be simple. The buildings would need significant power and cooling infrastructure upgrades, and possibly permit changes, before a single rack draws current.

Why central offices?

Brightspeed was founded in 2021, after Apollo Global acquired Lumen's incumbent local exchange carrier business across 20 states and started converting the copper network to fiber. Fiber needs fewer connection hubs than copper, which leaves a footprint of increasingly idle exchange buildings.

Operators globally face the same question as they decommission copper. Some walk away entirely, leaving inner-city sites to be redeveloped as housing. Others, including Telefónica and Ziply, are upgrading and repurposing their buildings for edge colocation.

Central offices have real advantages: high redundancy and low-latency locations. They also have real problems. The facilities are extremely old, very low-density, and were designed for DC rather than AC power — so accommodating modern IT gear takes substantial retrofit work.

There's a wrinkle on ownership, too. Many operators have already sold their central offices in sale-and-leaseback deals to fund network growth. Brightspeed itself has been looking to sell a number of its data center sites in leaseback deals across the US, having recently put facilities on the market in Arkansas, Virginia, Missouri, and Ohio. If the buildings change hands, Nvidia's counterparty may not stay the same.

Why now — and why these two?

The timing is awkward for both companies. Brightspeed recently appointed a new CEO amid reports the firm is close to bankruptcy. A 1GW inference anchor tenant would be a lifeline for a company whose copper-era assets are losing relevance by the quarter.

Nvidia, meanwhile, carries its own baggage. CEO Jensen Huang has been embroiled in an affair scandal involving a potential Chinese spy — a distraction, if nothing more, for a company negotiating infrastructure deals that could take years to mature.

What happens next?

The reported sequence is straightforward: information sharing, site visits, then presumably a decision on which buildings make the cut and who pays for the power and cooling retrofits. The Q1 2027 target gives both sides roughly two years to turn drafty telephone exchanges into GPU halls.

For the wider industry, the deal — if it happens — would be the clearest signal yet that telco real estate, long treated as a copper-era liability, has a second life as AI inference infrastructure.