AT&T is building its network for AI traffic that doesn't exist yet

AT&T is no longer building its network to push video down to phones. It is building one for the drones, robots, and AR glasses that will push data back upstream, and the company's July 22, 2026 earnings call turned that inversion into a stated capital priority.

"This is the agentic AI wave," CEO John Stankey said on the Q2 2026 call, naming autonomous vehicles, robotics, and augmented-reality glasses as the upstream drivers that will reshape what a carrier's network has to do. Stankey's framing positions AT&T as a network for machines that sense, decide, and act on their own, not the human viewers that 4G and 5G were designed to serve.

AT&T operates roughly 5,000 central offices and 75,000 cell sites, with fiber running directly to about 600 U.S. data centers. The carrier has rolled out 400-gigabit wavelength service to roughly 40 metro markets and around 130 interconnection nodes, the on-ramps where large enterprises, cloud regions, and AI training clusters meet the long-haul fiber. That gives AT&T dedicated 400 Gbps pipes between data centers, cloud zones, AI compute clusters, and big enterprise sites. The infrastructure is not new in absolute terms. The purpose has changed: capacity that runs toward the network from the device, instead of the other way around.

That flips the optimization AT&T has spent two decades selling. The carrier is now tuning its mid-band spectrum from EchoStar and its 600 MHz low-band for deep indoor uplink, not for headline download speeds. Mid-band carries the bulk of the traffic, while 600 MHz punches signals into basements, factory floors, and the places where a delivery robot or a docked drone has to send high-resolution video and telemetry without losing a packet. The same antennas that once had to be evaluated on how fast a phone could pull a Netflix stream are now being graded on how steadily a robot arm can stream a multi-megapixel scene to a model in a regional data center.

AT&T also claims that AI-driven cell-site "sleep" modes, where radios power down during low-traffic windows and wake on demand, are cutting site energy use 20 to 30 percent. The figure is AT&T's own. The carrier's energy management page does not break out which sites are running the software or how the savings scale as more radios cycle. Twenty-to-thirty percent is the kind of number that travels well in a press release and reads less well under a third-party meter. Cell-site sleep only matters at scale if a meaningful share of the 75,000 sites are running the software, and AT&T has not said.

Stankey cited Cisco modeling that puts consumer traffic at 6.6 times today's level by 2035, with AI inference accounting for roughly a quarter of the total. Industry studies that AT&T did not name project AI agents scaling from tens of billions today to trillions by 2036, and daily global bandwidth from about 100 exabytes to around 8,100 exabytes over the same period, a compound annual growth rate north of 50 percent. RCR Wireless is the conduit for those numbers in the source set; the underlying studies are not in the bundle, and the projections are framing AT&T is using, not facts it has measured.

The Q2 2026 numbers tell you the company is putting real money behind the bet even if the demand is hypothetical. Consolidated revenue came in at $31.6 billion, with record combined fiber and fixed wireless net additions and the strongest consumer postpaid wireless account growth in more than three years. AT&T also said it is accelerating share repurchases to roughly $10 billion for the full year, a capital-return signal that the operating business can fund the upstream rebalancing without leaning on debt. The full call transcript shows the same message going to both consumer and business analysts: the network is being designed for the next decade's traffic shape, not the last one's.

The open question is whether the demand shape Stankey is betting on actually arrives. The trillion-agent projection is the right kind of number to get a board to fund mid-band spectrum and 400G buildouts, and the wrong kind of number to bet a decade of capex on without independent confirmation. AI-RAN trials — running AI workloads inside the radio access network itself rather than only on top of it — are still early across the industry, and the cell-site-sleep savings only matter at scale if a meaningful share of the 75,000 sites are running the software.

AT&T is, in effect, building roads for traffic it hopes will come. Whether those roads are wide enough is the test the next four quarters of net adds, energy use, and 400G attach rates will start to answer.