Bandwidth bought for knowledge middle connectivity surged by practically 330% between 2020 and 2024, a brand new report reveals, pushed by hyperscale enlargement and AI developments.
Among the many most notable findings of Zayo’s inaugural Bandwidth Report (registration required) is the outsized function of hyperscale knowledge middle operators and carriers. In 2024, simply 10 consumers – predominantly in these two segments – accounted for practically 62% of all bandwidth purchases.
The report attracts on buying knowledge from over 1,800 Zayo clients and a survey of 16 enterprise decision-makers.
Between 2020 and 2024, hyperscalers had been liable for 57% of Zayo’s metro darkish fiber installations and 41% of all large-scale bandwidth offers exceeding one terabit. Whole bandwidth purchases greater than doubled throughout the interval, rising by practically 133% to succeed in 42.4 terabits.
Metro darkish fiber is the availability of ‘unlit’ optical fiber infrastructure in city areas. Demand right here has adopted a equally steep curve. From 2023 to 2024, metro darkish fiber purchases rose by 268%, whereas long-haul darkish fiber purchases grew by 53%.
In the identical timeframe, the entire metro fiber rely rose greater than 600%, together with a 2,300% enhance within the largest single metro deal by fiber rely. The biggest long-haul darkish fiber deal in 2024 was an 864-fiber rely buy by a hyperscaler.
AI Drivers
AI has emerged as the first catalyst for these networking developments, with inference and coaching workloads inserting extraordinary stress on bandwidth necessities. Zayo reported greater than $1 billion in AI-related long-haul community offers in 2024, with one other $3 billion within the pipeline.
Chaz Kramer, Zayo’s vice chairman of fiber and transport, informed DCN {that a} major problem for knowledge middle operators is securing geographic places with ample, accessible energy to help the facility calls for of AI workloads.
“Most of the extra most well-liked areas are already power-constrained, forcing operators to discover secondary markets or spend money on customized energy infrastructure,” Kramer mentioned.
Hyperscalers are more and more shifting away from leased bandwidth providers and towards proudly owning darkish fiber to regulate scale, latency, and value.
By buying bulk fiber, they keep “reserves” in key markets, enabling speedy capability upgrades with out ready for supplier provisioning. Some are additionally deploying parallel long-haul and metro networks to create fault-tolerant pathways between AI coaching clusters.
“This pattern displays a broader transfer towards vertically built-in community management to align infrastructure with compute technique,” Kramer says.
Reshaping Community Topologies
Jimmy Yu, who tracks optical transport gear marketplace for knowledge middle interconnect at Dell’Oro Group, says the surge is reshaping community topologies and amplifying demand for high-capacity routes in each conventional and rising knowledge middle markets.
He notes that AI demand acts as each a disruptor and an accelerator, with wavelength and darkish fiber necessities being formed not solely by the dimensions of AI fashions and coaching knowledge however by the shift towards distributed infrastructure methods, hybrid connectivity, and edge buildouts.
As hyperscalers ramp up AI deployment, hovering energy necessities and rising bandwidth calls for are reshaping the place and the way knowledge facilities are being constructed.
In keeping with Yu, the fee and scale of AI workloads are creating important constraints—and forcing operators to rethink their infrastructure methods.
“There’s a shortage of electrical energy, actually,” Yu says. “Since there’s a constraint in energy, you’ll be able to’t construct actually giant knowledge facilities, like mega knowledge facilities with AI, as a result of the quantity of energy consumption for Nvidia GPUs is simply a lot larger than what they had been doing earlier than.”
The steep vitality calls for per rack are driving corporations to construct exterior conventional metro hubs, the place electrical infrastructure is usually maxed out.
“They appear to different components of the globe and in america the place there may be electrical energy or surplus electrical energy, the place there may be house to construct a knowledge middle,” Yu says. “However meaning you now want to attach it, as a result of a knowledge middle standing alone is ineffective.”
Yu emphasizes that connectivity is a crucial enabler of the AI infrastructure buildout.
“You may fill [a data center] with all of the servers on the earth, however and not using a good knowledge middle interconnect, it’s ineffective infrastructure,” he says. “It’s worthwhile to get fiber to all these knowledge facilities.”
The geographic distribution of knowledge facilities is altering quickly consequently.
“That’s precisely what’s going on,” Yu says, referencing knowledge that confirmed outsized progress in cities like Memphis, Tennessee, the place demand for long-haul and metro bandwidth grew from 0.3 terabits in 2023 to 13.2 terabits in 2024.
In Memphis, demand for long-haul and metro wavelength connectivity grew by 4,300% from 2023 to 2024, whereas Salt Lake Metropolis noticed a 348% enhance, pushed largely by hyperscaler curiosity in inexpensive land and energy.
“With fiber optic cabling, you’ll be able to restrict the affect of latency, particularly if the routes are extra direct,” Yu provides. “When you’ve got your individual darkish fiber and it’s a direct path to the purpose you should get to, it doesn’t matter if the info middle is farther away.”
Yu sees the present section as only the start, noting that it’s nonetheless the early days of AI.
“For at the very least the following two to a few years, simply to construct out the AI infrastructure and join all these knowledge facilities, that’s going to take up a variety of bandwidth,” he says.
