Data centers are exploding with AI demand that is putting significant stress on the network connectivity infrastructure that ties them together. It’s a stress that network providers are racing to alleviate.

This stress is evident in the multitude of recent AI-related expansion and enhancement projects announced by firms operating up and down the networking stack. These moves are based on the need to both increase the capacity of networks running between AI-engorged centralized data centers and latency-minimizing Edge locations.

Jeetu Patel, Cisco’s chief product officer, recently told attendees at an analyst conference that three AI-related traffic waves will stress current network architectures and force a complete rethink of infrastructure from the data center out to the campus Edge.

Patel explained that the first wave is coming from so-called “chatbots,” where “I ask a question, I get an answer back,” adding that these AI platforms “have very spikey traffic patterns,” where “the utilization spikes up and then comes right back down.”

Patel noted that this pattern is being handled competently by current network architectures.

The next AI-driven traffic wave is coming from agentic AI, which is set to amplify those chatbot peaks into a sustained higher level of activity.

“The traffic patterns start to get much more sustained and persistent over time,” Patel noted of this model. “And so our current infrastructure is simply not built to go out and accommodate that level of traffic pattern.”

The third pattern is what Patel called “physical AI, where you’re going to need some more Edge-based computing and Edge-based networking that’s only going to compound the requirements.”

Patel explained that those last two traffic patterns are driving enhanced infrastructure requirements “both in campus branch as well as in data centers,” with the latter also needing to adhere to power limitations.

“And so what you're starting to see is re-architecting of data centers to accommodate for this new additional volume of usage,” Patel explained.

“And you're also starting to see re-architecting of campus branch networks because everything from WiFi to routing switching needs to get rethought.”

Dell’Oro Group echoed this concern in a recent report, which forecast increased spending on campus Ethernet switch gear “as enterprises invest in higher capacity networks.”

“The wide use of AI agents is expected to put new requirements on the [LAN],” Dell’Oro Group research director Siân Morgan wrote. “It is still too early to determine exactly how it will play out, but traffic patterns, volumes, and sensitivity to latency are likely to change – leaving room for product differentiation by campus switch vendors. Enterprises recognize the importance of a high-performance LAN, especially as they plan for the implementation of AI use cases.”

Stress on – or under – the street

This flood of need is forcing vendors to selectively attack the problem.

John Coster, manager for innovation, planning, and strategy at T-Mobile US, touched on this challenge during a panel discussion at the Yotta 2025 event in Las Vegas.

“Think about 125 miles (201 km) as a millisecond, and you’ve got like 30 miles (48 km) between towers, so think about 100 miles (161 km) from tower to tower from the fiber standpoint,” Coster said, adding in the roundtrip for data traveling over that fiber network. “We shoot for 10 milliseconds [of latency], and there are people that say with [augmented reality], virtual reality, you want to see less than three [seconds of latency].”

Coster said T-Mobile US works with application developers, such as the companies behind autonomous vehicles, to work through this latency deficit by fine-tuning the ability to process data within a vehicle and for what data needs to be sent back and forth over a fiber network.

“There are so many unknowns right now, it's just evolving and everyone's kind of trying to feel their way along,” Coster said. “How much do we invest in edge AI to handle the kinds of traffic when we don't know who the client is. It's a little bit of an inventive process.”

Changing the network architecture for AI

Lumen Technologies CTO Dave Ward explained the depth and challenge of this network architecture quandary in an interview with DCD last year.

“There are so many capacity constraints associated with constructing AI: power, literally power from the grid, and then where you can place your data center, can you get the GPUs?” Ward says. “This network capacity is a very scarce resource, in particular, where the data centers or AI data centers are being built.”

Analyst firm LightCounting, for instance, predicts growing AI use will result in a doubling of sales this year for Ethernet optical transceivers used in AI clusters.

“What this means for us as a connectivity partner is we fully plan on building an AI fabric between these locations and major data centers that are of the right power, size, and scale to host GPUs and the AI workloads and create a specialized connectivity fabric just for that purpose,” Ward says of this effort. “So much fiber and so many waves are required, and that has really become a different segment than some of the other cloud economic segments that are coming in, and we can construct just to that.”

Ward recently authored a Lumen white paper that called for a “reset of network capabilities,” contending that current systems are not fit to meet demands for the next generation of cloud.

That reset included the need to extend fiber and optical connectivity into areas where power exists, and data centers are planned. This will require work with hyperscalers and enterprises to target expansion into tier-two markets.

“This explosion of rural data center operator clusters only further exacerbates current architecture problems,” the white paper reads. “Instead of backhauling to a major metro to find a carrier-neutral facility interconnect, new local interconnection may be much more efficient. Lumen has identified dozens of new data center clusters scattered across the US that will require fiber, wave, and IP services.”

Dark density and powerful programmability

The Lumen white paper also suggests that more “purpose-built connectivity” is needed to truly meet the demands for AI workloads. Instead of connecting everything equally, Ward and Lumen’s idea for purpose-driven networks would see more dark fiber networks employed to help connect facilities.

Distributed data centers are of growing interest to hyperscalers, with solutions like Nvidia’s Spectrum-XGS looking to turn multiple, disparate data centers into unified “AI super-factories,” to borrow the chipmaker’s parlance.

Ward wrote: “While we expect that these factories will primarily communicate with one another through distributed and sharded training and reinforcement at an industrial scale, they will also interface with the already overburdened Cloud 1.0 architecture in major metropolitan areas to distribute inference and exchange data and models with partners."

In addition to dark fiber, Ward and the team at Lumen want networks supporting next-generation workloads to feature programmable underlays. This would allow operators to employ bandwidth from premises to the cloud and create fabrics as needed, with SD-WAN and secure access service edge (SASE) tunnels helping to keep workloads secure.

“Interconnected enterprises will most certainly transform with Cloud 2.0,” he added. “Both public and private internet networks must become faster – as well as more secure, distributed, and programmable – to meet new demands. … Most of the architectural foundations that defined Cloud 1.0 are obsolete and cannot support the requirements of the new cloud era.”

This need to stay ahead of demand is set to drive significant network investments over the next several years. This will include both the extension of those networks and in the technology to glean the most efficiency from those deployments, investments that will need to remain aligned with AI-fueled data center expansion.