16
Verizon, AT&T, and SK Telecom are betting huge on AI infrastructure. However expertise advisory group Omdia says the headline numbers are ceilings, and warns income will path capability development for years.
In sum – what to know:
Unproven scale – Verizon’s $1bn Google DCI settlement is signed and reside; Schulman’s declare of a multi-billion-dollar “development vector” from 2027 continues to be a forecast.
Small contracts – Hyperscalers spec offers at their eventual ceiling (typically 400 waves of capability) however signal carriers up for a fraction in the beginning (as few as six waves).
Visitors ≠ income – Capability demand from AI is “assured” to rise, Omdia says, however income per bit retains falling, highlighting a structural distinction within the two development curves.
Verizon chief Dan Schulman lately advised buyers that AI infrastructure offers will “stack up” in 2027. The provider’s $1 billion dark-fiber cope with Google, signed in late July, is, he believes, simply the beginning. Different offers value “a number of billions” will comply with by year-end, forming what he calls “long-duration, high-quality contracted income streams” that stack up into a brand new development vector from 2027, earlier than increasing considerably within the subsequent 5 to 10 years.
“This can be a very totally different income development profile than Verizon has had in a really very long time,” he mentioned.
As mentioned this week, telcos are more and more positioning themselves as infrastructure suppliers for the AI economic system. The Google deal is proof of that shift. Expertise advisory group Omdia’s Telco B2B AI Monetization Index places complete telco AI income at $4 billion globally in 2025, rising at a CAGR of 65 % to 2030. However how briskly, and the way far, that shift scales continues to be an open query, and the numbers counsel it’s earlier days than the framing implies.
Brian Washburn, chief analyst for telco B2B options at Omdia, has spent three years modelling AI community site visitors. He has recognised a sample behind headline offers like these: hyperscalers spec out contracts at their eventual ceiling, however signal carriers up in the beginning for a fraction of it. “Plenty of these multi-billion-dollar contracts have potential, and the expansion curve may be very promising,” Washburn mentioned. “Nevertheless it all begins form of small.”
A “wave”, brief for wavelength, is a single channel of sunshine on a fiber-optic cable, each able to carrying a set quantity of information; carriers gentle up extra waves as demand grows, with out laying new cable. A hyperscaler may spec a deal at 400 waves of 400 gigabits every, framing that because the ceiling for a 20-year settlement. The preliminary signed contract, nonetheless, may cowl six waves, which is round 1.5 % of the acknowledged capability.
Fiber-glut parallel
Washburn attracts a direct parallel between the present overbuild and the fiber glut within the Nineteen Nineties, the place an excessive amount of funding capital was chasing after too few business income {dollars}, the dynamic that led to fiber routes being turned off, and finally led to consolidation of CLECs like Stage 3 and Zayo. Whereas he expects one thing comparable in AI infrastructure, given there are “too many gamers and an excessive amount of fragmentation,” he sees a vital distinction for carriers. “We’re the secondary beneficiary,” he mentioned.
If the AI infrastructure layer consolidation turns into a large number, the underlying demand for connectivity (supporting GPUs as a service and AI platform necessities, for instance) doesn’t disappear, it simply will get served by whoever is left.
Thankfully, telcos are being cautious about separating out alternatives with neoclouds and hyperscalers which might be actual and cogent, and people which might be good concepts on paper and chasing funding. That self-discipline to work on smaller and extra pragmatic offers, somewhat than larger speculative ones, Washburn mentioned, is partly compelled, as telcos don’t have entry to the near-unlimited capital of AI gamers. One instance of this strategy is the pre-purchase of optical transport tools to supply wavelengths with brief provisioning home windows; a calculated, incremental dedication somewhat than a speculative land-grab.
AT&T’s community redesign is a living proof. The provider is intentionally re-engineering for upstream capability somewhat than the downstream speeds carriers have marketed since 4G, betting that agentic AI site visitors (akin to drones, autonomous automobiles, robotics) will push massive volumes of latency-sensitive knowledge from the sting again to the community. Just like the darkish fiber buildout, it exhibits how telcos are constructing out infrastructure forward of confirmed demand.
SK Telecom is making an identical guess. The provider has arrange a devoted subsidiary, SK Hyper, to construct as much as 15GW of AI knowledge centre capability, backed by as much as $540 million in dedicated capital by to 2030. However even SK Telecom is hedging the timeline, with the primary 5GW section not due till 2029, and the total 15GW goal not due till 2035, “topic to market demand and different components.”
An upside caveat
Telcos could not want the AI infrastructure story to play out completely as a way to profit. Capability demand from AI site visitors is “assured” to maintain rising, he mentioned. Income is a distinct matter: “The quantity of revenues that one can extract from that may proceed to rise, however at a far slower tempo than the demand for capability,” he mentioned, as a result of because the expertise retains advancing, “the value per bit retains happening.” Rising site visitors and rising income, in different phrases, usually are not on the identical curve, and the hole between them is structural.
Washburn factors to Stage 3’s nationwide community, the place conduit laid within the Nineteen Nineties is now greater than 25 years previous. The operator is now placing contemporary fiber into empty house in that very same conduit, having run low on capability in its authentic fiber pairs, to increase the asset’s working life by one other 25 years. A 50-year planning horizon on a single piece of infrastructure, he mentioned, is the form of timescale most industries can not think about. By comparability, Schulman’s AI infrastructure narrative includes a 2027 horizon, which is a fraction of the timescale telcos are literally constructing for.
This mismatch could also be much less a contradiction than the important thing differentiator. Whether or not “a number of billions” from a handful of offers compounds into the decade-long development or settles into one thing steadier and smaller stays to be seen. Both means, carriers are betting the pipes stay a helpful asset, no matter which model of the story seems to be true.