Data centre interconnect (DCI) is one of the optical communications industry's growth stories, with recent market intelligence from Dell’Oro Group suggesting global DCI grew by 40% year-on-year in the first quarter of 2026.
But Tim Doiron, Vice President of Solution Marketing at Nokia, believes it would be a mistake to view that growth as the result of a single trend. Instead in an exclusive article for Fibre Systems, he identifies three distinct forces reshaping DCI: traditional connectivity between data centres, the emergence of scale-across networking for massive AI training clusters, and the growing requirement to distribute AI inference closer to end users.
The most dramatic change may be the emergence of scale-across networking. For Doiron, this represents a big shift in how hyperscalers can build AI infrastructure. Rather than keeping accelerated compute inside the physical boundaries of one facility, scale-across networking enables GPU and TPU clusters to extend across multiple data centres.
The ambition is considerable: hundreds of thousands of GPUs at one site could be connected to hundreds of thousands more at another location, potentially hundreds of kilometres away, effectively creating a single synchronous super-cluster. “The goal of scale-across networking is to enable massive computational super-clusters that physically span multiple facilities and power grids for AI training,” explains Doiron.
The motivation is equally compelling. As AI models become larger, the power, cooling and physical-space constraints of individual data centres are becoming increasingly difficult to overcome. For hyperscalers, reducing the time required to train frontier AI models can have significant commercial value. Scale-across therefore turns DCI into something much more fundamental: a mechanism for scaling compute itself.
Yet Doiron points to another constraint that could prove just as significant: the limits of optical fibre. “Historically, we could reliably expect a 30% capacity improvement with each new DSP cycle,” says Doiron. “Today, that generational improvement has slowed to around 10%.”
The consequence is a change in how network capacity must scale, creating another infrastructure challenge. Thousands of fibre pairs have to be managed and amplified along routes where inline amplifier sites are themselves constrained by space and power. And adding capacity is only part of the story. The diversity of AI infrastructure means that different DCI applications require different optical architectures.
Another emerging requirement is distributed inferencing. Nokia Bell Labs analysis indicates that approximately 21% of AI inferencing workloads could run within telecom provider facilities by 2030. For Doiron, this creates an opportunity for operators to exploit their existing central offices and edge locations to host accelerated compute closer to users.
“Telecom operators have an edge footprint that hyperscalers and neoclouds simply don’t,” he says. The result could be a distributed AI infrastructure in which telecom networks become more than a connectivity layer – instead becoming part of the platform on which real-time AI services operate.
Traditional DCI, scale-across and distributed inferencing may have different requirements, but Doiron sees them as part of the same fundamental transformation. DCI is evolving from a transport necessity into a strategic component of AI infrastructure. As compute becomes larger, more distributed and more demanding, the optical network must evolve with it.
“The optical technologies required to serve all three DCI use-cases ... represent a fundamental rewriting of the network design playbook,” Doiron concludes.
To read more about Nokia’s exclusive analysis – and a host of other related material – download our eBook on scaling optical infrastructure for AI. Explore campus DCI innovation, high-fibre-count cable strategies, deployment acceleration, and testing at scale with Nokia, US Conec, and VIAVI.