Capital expenditures on data center construction and expansion continue at a torrid pace in 2026. Following Q1 earnings, most analysts raised their hyperscaler capex spending projections for the year – exceeding the 60+ percent growth rate we witnessed in each of the past two years..
Global data center capital expenditures across hyperscalers, neocloud service providers, and sovereign AI projects are now projected to exceed $1.7 trillion by 2030. In parallel, the transition to so-called ‘AI factories’ underpinned by accelerated compute processors (XPUs) is increasing connectivity demands between data centers..
A specialized form of data center interconnect (DCI), scale-across networking, has emerged to enable AI and cloud providers to utilize optical connectivity to scale XPU clusters and backend networks beyond the physical limits and power constraints of a single data center.. Building and training advanced large language models often requires months of processing time, despite access to clusters containing hundreds of thousands of XPUs.
The race to increase computational parallelism and reduce training times is in full view, with multiple data center projects (including xAI’s announced Colossus expansion into a third data center building in the Memphis area) targeting more than one million XPUs per cluster.. At the same time, global data center investment is accelerating the buildout of modular, distributed facilities and upgrades to existing sites, to support inferencing workloads to deliver rapid responses to enterprise and consumer queries.
While AI training workloads have previously dominated data center demands, inference workloads will overtake AI training as early as 2026.. For both traditional DCI and backend scale-across networking, high-speed optical connectivity is a foundational part of the AI fabric, linking data centers and accelerated computing across the AI grid.
To cost-effectively meet the needs of data center operators and the wholesale transport providers that support them, the industry needs continued innovation and integration to scale capacity, reduce space and power per bit, and optimize costs.. Three innovations are central to this next phase of DCI and scale-across networking: coherent pluggables, multi-scale open optical line systems, and AI-enabled network automation..
To cost-effectively meet the needs of data center operators and the wholesale transport providers that support them, the industry needs continued innovation and integration to scale capacity, reduce space and power per bit, and optimize costs.. Coherent pluggables. Coherent pluggables have evolved from a metro DCI option to a mainstream building block, for both traditional DCI and emerging scale-across architectures..
Their value lies in the ability to integrate high-capacity optical connectivity directly into routers, switches, and compact transport systems while reducing power consumption, footprint, and overall cost per bit. That model has already reshaped metro and regional DCI, and the transition from 400G ZR/ZR+ to 800G ZR/ZR+ is now extending pluggable economics and operational simplicity to higher-capacity AI and cloud interconnect applications..
Industry analysts increasingly view pluggables as a primary engine of bandwidth growth, with 800G rollouts accelerating in 2026 as distributed data centers and AI clusters drive demand for more capacity between sites.. Optical line systems. Optical line systems (OLS) are the on-ramps and refueling stations for optical networks — aggregating and amplifying wavelengths on fiber.
With increased integration and miniaturization, one can now multiplex/demultiplex up to 64 wavelengths onto a fiber pair with a single compact module, enabling 50-75 Tb/s of transmission capacity per fiber pair.. Network operators can choose between cost- and power-optimized fixed-filter designs and flexible, programmable ones that support varying data rates and multiple generations of pluggable and embedded optical engines..
In parallel with multiplexing/demultiplexing DWDM wavelengths at terminal sites, integration and innovation are delivering compact in-line amplifiers (ILAs) to support multiple fiber pairs with reduced power and footprint. Optical signals over single-mode fiber attenuate approximately 0.2 dB per kilometer and must be amplified every 50- 100 km (~30 – ~60 miles)..
ILAs are deployed along the fiber path to provide this amplification. Today’s ILAs amplify a single fiber or fiber pair per module. But with the largest AI and cloud providers driving petabits of interconnect capacity across hundreds of fiber pairs for applications like scale-across networking, this approach would quickly exhaust the space and power constraints of today’s small amplifier huts.
Emerging compact, high-density ILAs support multiple fiber pairs per module and enable amplification of up to 160 fiber pairs per seven-foot (2m) rack – or an 8x density improvement over the prior generation.. Network automation. As optical connectivity underpins the AI grid for both training and inference, automation is becoming an architectural imperative, not just an operational efficiency tool.
Network automation begins with visibility and programmability through open APIs, streaming telemetry, and standardized data models that expose real-time network conditions and key performance indicators to orchestration and service assurance systems.. In turn, network automation helps operators improve resource utilization, optimize performance, support new service revenue, and connect and adjust transmission capacity with orchestrated workloads across distributed accelerated computing locations.
Network automation also strengthens resiliency by identifying anomalies and fiber impairments early, before they propagate across and impact the compute domain.. Delivering the next phase. As AI investment continues to accelerate, the networks connecting data centers and distributed, accelerated compute resources are more critical than ever.
Meeting the needs of both massive training clusters and distributed inference workloads will require optical architectures that are not only higher capacity but also more compact, power-efficient, and automated.. That is why coherent pluggables, multi-scale optical line systems, and open, telemetry-driven network automation are emerging as foundational building blocks to cost-effectively deliver the next phase of data center interconnect and scale-across networking..
More in Networking. 07 Aug 2026
