As artificial intelligence demands more computing power than any single data center can supply, hyperscalers are building networks of campus data centers 2-20 kilometers apart linked by fiber-optic cables—a pattern called 'scale-across.' This requires coherent optics, a technology originally designed for expensive long-haul telecom networks, but vendors are now offering cheaper variants tailored to AI use, making large-scale distributed AI infrastructure economically viable.
What happened
Hyperscalers are spreading GPU clusters across campus data centers 2-20 kilometers apart and linking them with high-speed optical networks—a model called 'scale-across'—to overcome single-data-center limits in power, space, and cooling. As AI interconnect bandwidth rises to 800G and 1.6T, coherent optics are replacing older transmission technology, but vendors are now rolling out cheaper variants (coherent lite and multi-rail designs) to make large-scale deployments affordable.
Why it matters
Single data centers face hard caps on power supply, floor space, and cooling capacity, blocking the expansion hyperscalers need for AI training and inference at scale. Scale-across solves this by distributing workloads across multiple sites linked by fiber. Coherent optics—the optical technology that enables this—were originally built for expensive long-haul networks; the new cost-cutting designs make coherent feasible for AI infrastructure, removing a major barrier to hyperscaler expansion.
What to watch
Optical equipment vendors are competing to offer the most cost-effective coherent solutions for AI data centers through innovations like coherent lite (which simplifies digital signal processors and optical components) and multi-rail systems (which share cabinets, power, and cooling across multiple optical rails to cut repeater costs). The supply chain for these components and systems will be a key bottleneck as hyperscalers ramp scale-across deployments.
Hyperscalers face a hard ceiling when trying to expand AI training and inference within a single data center. Power supply, floor space, and cooling capacity all hit limits simultaneously, creating a bottleneck for the GPU-intensive workloads that modern AI requires. To break through, companies are adopting a distributed architecture: spreading GPU clusters across multiple campus data centers located 2-20 kilometers apart, then binding them together with high-speed optical fiber networks. This model, called 'scale-across,' lets dispersed computing resources operate as a unified system for AI training and inference.
The optical technology underpinning scale-across has undergone a critical shift. When AI interconnect bandwidth was lower, traditional intensity modulation/direct detection (IMDD) technology sufficed for short- and medium-distance links. But as bandwidth climbs to 800G and 1.6T, IMDD becomes insufficient. Coherent optics—a technology that encodes more data per signal and achieves longer distances at those speeds—is now the mainstream choice. However, coherent optics were originally designed for long-haul telecom backbone networks, where they command premium prices because they use expensive digital signal processors (DSPs) and precision-engineered optical components. Applying off-the-shelf coherent gear to scale-across deployments would drive costs up prohibitively.
To solve this, optical equipment vendors are deploying cost-optimized coherent variants. Coherent lite simplifies the DSP and optical component specifications, reducing module cost while retaining the bandwidth and distance needed for AI interconnects. Multi-rail architecture takes a different approach: it packs multiple optical transmission systems into shared cabinets, with shared power and cooling infrastructure, and consolidates repeater facilities—cutting the total system cost of deploying coherent optics at scale. These innovations are purpose-built for the hyperscaler use case, making coherent optics economically viable for large-scale AI data center rollouts and removing a major constraint on scale-across adoption.
The shift toward scale-across reflects a fundamental constraint hitting hyperscalers as AI workloads grow: no single data center can provide enough power, physical space, and cooling to train and run the largest AI models. Rather than build bigger single facilities, the industry is pivoting to a distributed model where GPU clusters at multiple sites work together seamlessly via high-speed fiber links. This requires optics technology that can push bandwidth to 800G and 1.6T while spanning kilometers between campuses.
Coherent optics are essential for this job—they can handle the bandwidth and distance that older intensity modulation/direct detection (IMDD) technology cannot—but they were originally engineered for expensive long-haul telecom networks, making them too costly for hyperscaler playbooks. The introduction of cost-reducing variants like coherent lite (simplified processors and components) and multi-rail architectures (shared infrastructure across multiple optical systems) narrows that cost gap. This innovation cycle directly enables the business model: as costs fall, scale-across becomes economically competitive, and hyperscalers can justify the capital and operational expense of managing distributed AI clusters.
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytime
Ask AI anything about this article. Q&As are published on this page for other readers too.
The AI news that matters, in one minute each morning.
Sign up free