Spending on data center switches deployed in AI back-end networks is forecast to surpass $100 billion by 2030 as demand for high-performance infrastructure amid the AI boom looks set to continue.

A Dell’Oro Group report suggests that back-end spending is being driven by expansion efforts across scale-up (adding more resources to an existing system to boost performance) and scale-out (cluster-to-cluster interconnectivity).

In addition to inside the data center, Dell’Oro Group’s forecast suggests the rise of scale-across architectures, in which geographically distributed data centers are interconnected to form a unified computing cluster, will further boost that drive. The likes of AWS’s Project Rainer or Microsoft’s Fairwater are among the increasing number of scale-across projects, with the concept expected to soar as a viable alternative to building massive facilities instead opting for multiple smaller sites to maximize available performance-per-watt.

“To meet this explosive demand for compute, we can no longer rely solely on scale-out networks to provide GPU-to-GPU connectivity across racks,” Dell'Oro Group VP Sameh Boujelbene explained. “This shift is driving the rise of scale-up fabrics that tightly couple GPUs and memory within a shared high-bandwidth environment, enabling the distributed reasoning.”

With network switch vendors set to benefit from increased sales, another winner of Dell’Oro’s projected back-end boom is Ethernet, with the standard set to emerge as the market’s “long-term winner.”

While the research firm notes that multiple technologies will co-exist across the stack, Ethernet looks set to continue its meteoric rise in the realm of high-performance backends.

Ethernet was the most populous interconnection standard in the supercomputing world in 2025, while HPE’s "Ethernet with a twist" Slingshot offering found in many of the top-performing systems.

Dell’Oro Group data published last year projected Ethernet to drive nearly $80 billion in data center switch sales over the next five years.

“Scale-up compute fabrics have been historically dominated by proprietary fabrics such as NVLink, but just as Ethernet prevailed over InfiniBand in large-scale scale-out environments, we are now seeing alternative technologies such as UALink and Ethernet gain momentum in scale-up compute fabrics, Boujelbene added. “While we predict a strong adoption of UALink, we expect Ethernet to emerge as the long-term winner across both scale-up and scale-out architectures.”

Even Nvidia, long wedded to InfiniBand, has seen its Ethernet-based Spectrum-X offerings go from a standard start to “roughly on par” with its rival protocol products.

The tech giant was name-dropped in Dell’Oro Group’s report, claiming Nvidia would be spearheading the adoption of co-packaged optics (CPO) to create silicon photonics switches. CPO iterations of both its Spectrum-X and InfiniBand-based Quantum-X fabrics are due to drop this year, with engineers from the chip giant opining to SDxCentral that they would offer 3.5-times more power efficiency, 63-times greater signal integrity, and 10-times better network resiliency at scale.

Joining Nvidia in the nascent CPO space are Broadcom, which has been working on co-packaged optical switches since 2021, with ongoing efforts aiming at 800 Gb/s (800G) and 1.6 Tb/s (1.6T) offerings. The vendor revealed to the world late last year that Meta had been testing its photonic offerings, resulting in one million cumulative 400G equivalent port device hours without a single link flap.

But for all its efforts, Broadcom CEO Hock Tan has erred on the side of caution when it comes to CPOs, telling investors in mid–December: “I could see a point in time, in the future, when silicon photonics matters as the only way to do it. We're not quite there yet, but we have the technology, and we continue to develop the technology.”