Every token that moves, every model that trains, every inference that ships rides on infrastructure telcos own. They carry the traffic the AI boom depends on, and capture almost none of the money it makes.
There is a familiar shape to the AI gold rush. Nvidia sells the picks. The hyperscalers stake the claims and build the mines. Model labs and application companies pan the gold. And somewhere underneath all of it, quietly indispensable and quietly unpaid, sits the connectivity that lets any of it function. That layer is telecom, and in 2026 it is being treated almost exactly the way roads are treated: essential, used by everyone, and a line item nobody wants to fund.
The financial picture is stark. Worldwide telecom capex is projected to decline about 2 percent in 2026 and grow at roughly 1 percent a year through 2030. Set that against the historic surge of AI-infrastructure spending happening one layer up, in data centers and compute, and the contrast is almost comic. The traffic on the network is exploding. The investment in the network is flat to falling.
A gap that history has seen before
Carrier revenues are growing at roughly 2 percent a year. Capex is shrinking. Put those together and the divergence between what the AI economy demands of networks and what operators can monetize from them runs to about 46 percent. That number should make anyone who lived through 2001 uneasy, because the comparable divergence in the telecom-excess cycle that decade was around 32 percent. We are already past it.
The 2001 cycle is the right reference, but the conventional lesson drawn from it is the wrong one. Operators tend to remember it as a story about overbuilding, and conclude that the answer this time is restraint. That reading is incomplete and a little dangerous.
The mistake in 2001 was never too much fiber. It was capacity laid down with no business model sitting on top of it. Build ahead of monetization and you do not get growth. You get stranded assets.
The instinct that will backfire
Here is the contrarian read. Telecom's reflex, faced with irrelevance in the AI value chain, is to spend its way back in. More backhaul. More fiber. More data-center interconnect. The logic feels sound: the compute economy needs connectivity, so build the connectivity and the value will follow.
It will not, at least not automatically, and the math says so. With utilization low and no monetization layer in place, every incremental dollar of network capex risks becoming a stranded asset inside three to five years. The problem operators face is not too little network. It is that they have no business model on top of the network they already own. Pouring concrete onto a road you cannot toll does not make you a toll operator. It makes you a more leveraged owner of free roads.
The encouraging signal is small but real. IDC expects spend on telco operations and monetization systems to grow about 5.2 percent in 2026, reaching roughly $54 billion, as operators try to extract revenue from network capex already in the ground. That is the right direction. Note the proportion: a modest single-digit increase in the systems that turn networks into money, against a much larger installed base of networks that currently do not. The imbalance is the whole story.
Figure — Who captures the AI value
Layer of the AI value chain | Who captures the value | Telecom's position |
|---|---|---|
Chips / silicon | Nvidia and a few designers | None — buyer at best |
Data centers | Hyperscalers, specialist operators | Adjacent; mostly a tenant or supplier |
Models | Frontier labs | None |
Applications | Software vendors, enterprises | None |
Connectivity / telecom | Largely uncaptured | Owns the layer, monetizes least of it |
How to read it: Scan the middle column from top to bottom. Every layer has a clear winner until you reach the one telecom actually owns, where the value sits largely uncaptured. The bottom row is the entire strategic problem in a single line.
What this means for leaders
Treat monetization, not capacity, as the binding constraint. The reflex to fund more network should be the second decision, never the first. The prior question is what, exactly, you will charge for, and who pays. If a capex proposal cannot name its toll, it is a candidate stranded asset, regardless of how much demand the deck projects.
Build the toll booth before the road, or in parallel with it. The 5.2 percent growth in operations and monetization systems is the seed of a strategy, not the strategy itself. Differentiated connectivity for AI workloads, premium interconnect, and usage-based commercial models have to be designed and sold alongside the buildout, not bolted on after utilization disappoints.
Take the warning literally. If backhaul and data-center interconnect do not keep pace, the compute economy will simply grow around the networks operators built. That is not a metaphor. It is a description of how telecom becomes a regulated utility carrying someone else's gold, and the window to avoid it is measured in a few years, not a decade.
Roads do not get rich because traffic is heavy. They get rich when someone decides to charge for the lane. Telecom has spent two decades proving it can build anything the digital economy asks for. The unanswered question of 2026 is whether it can finally learn to charge for it, because the AI boom will not slow down to wait for the answer.
A BusinessInfomatics original. Figures from Dell'Oro Group, IDC, and the PwC 2026 telecom outlook.
