11
Connectivity, not just compute, will determine AI’s infrastructure future, argues former AT&T president Glenn Lurie. Drawing on the iPhone’s launch, he says operators, hyperscalers and technology companies must build together before infrastructure bottlenecks dictate the outcome.
In 2007, we launched the first iPhone at AT&T on a 2G network. A year later came the 3G version and the App Store. What made that shift possible wasn’t technology alone. Two industries that had spent years wary of each other had to figure out how to build together. We opened parts of the network we had never opened before. Neither company could have created what came next on its own.
Once we stopped trying to, an entire application economy emerged on top of it.
I have watched the AI buildout with that experience in mind, and I see the same pattern taking shape. Only this time, we are not handling it nearly as well. Connectivity has quietly become the deciding half of the AI story, because everything above it stops working when it falls short. People who came up in wireless understand this without being told.
They spent their careers learning that coverage, capacity and latency are the entire product. A network failing at two in the morning costs you a customer regardless of the roadmap. That same infrastructure mindset hasn’t always carried into the data center buildout. Connectivity can arrive late in those conversations, treated as a sourcing exercise once the site is chosen and the power is negotiated.
The data center buildout is already showing us what happens when infrastructure cannot keep pace with ambition. Nearly 40% of the data center projects slated to open this year are expected to miss their dates by three months or more, held up by labor, power, equipment and permits, while North American vacancy sits near 1%. Every delayed megawatt has a tenant waiting.
Research from the Fiber Broadband Association put the requirement at roughly 214 million additional fiber miles in the United States by 2029, about 2.3 times what is in the ground today. Fiber at that scale means easements, permits, railroad crossings and crews already committed elsewhere. None of it moves faster simply because there is more capital behind it.
The money says the same thing from the other direction. The four largest hyperscalers are on track to spend roughly $760 billion on capital projects this year, against about $413 billion in 2025. Worldwide telecom capex, meanwhile, is forecast to decline about 2% over the same period. U.S. carriers will still put north of $40 billion into their networks this year. Having run one of those businesses, I know how hard those dollars are to earn back.
The industry making the largest bet on AI is depending on capacity it does not fund, does not control and has almost no hand in planning.
Which brings me to what the industry owes itself. A real orchestration layer. Workloads need to move across networks, across owners and out to the edge based on what the application requires at that moment, without a developer or customer ever choosing a network by hand.
The end user shouldn’t have to know or care whether that intelligence traveled over fiber, wireless or satellite. The software that finally makes that invisible will be worth a great deal, but it only works if the interfaces underneath are agreed on first. No single company owns enough of the path to set them alone. Operators, cloud providers, chip companies and power providers will have to build those interfaces together.
There is more of that in place than people realize.
Under the GSMA’s Open Gateway, 86 operator groups representing more than 300 networks and roughly 80% of global mobile connections have aligned on a common set of network APIs, giving software a standardized way to ask a network for the performance it needs. Getting that many competitors to agree on anything is hard, and the operators did it anyway. Very little AI software has been written to take advantage of it, and that is a gap worth closing.
That orchestration layer also changes what happens inside the network. Agentic systems can increasingly detect a problem, diagnose it across domains and act before a customer ever notices. The same intelligence can increasingly resolve customer problems on the device rather than routing them through a call center. These are the beginnings of the self-healing networks the industry has talked about for years, but they only work if the connectivity underneath them is built to operate as one system.
The lesson I carried out of 2008 is that technology tends to arrive roughly when the engineers say it will. The partnerships tend to arrive later, and those partnerships ultimately decide how a technology era plays out.
Compute companies, network operators, satellite providers and power companies are going to depend on one another whether they plan for it or not. The winners will be the ones that decide to build together before the infrastructure forces them to.
