The Third Tier: Scale-Across Like Kriss Kross

Key Takeaways

• Scale‑across links AI clusters across buildings over 500 m
• Coherent optics and disaggregated WDM make long‑distance fabric possible
• NVIDIA Spectrum‑XGS claims 1.9× NCCL speed at 10 km
• Cisco, Broadcom, Ciena, Nokia all announced dedicated scale‑across gear
• Hyperscalers already run multi‑site training, driving market demand

Pulse Analysis

The rapid growth of foundation models has exposed a hard limit: a single data‑center can no longer house all the GPUs needed for frontier training. "Scale‑across"—the third tier introduced by NVIDIA—extends the high‑bandwidth, low‑latency fabric beyond the rack and hall, reaching distances of 500 meters to hundreds of kilometers. By treating separate buildings as a unified compute pool, organizations can aggregate more accelerators without the costly overhead of duplicating power and cooling infrastructure.

At the heart of scale‑across is coherent optical technology. Unlike traditional Ethernet, coherent pluggables embed digital signal processing within the transceiver, allowing disaggregated wavelength‑division multiplexing (WDM) to deliver 200 Gb/s‑plus links over 20 km with sub‑microsecond latency. Vendors have raced to commercialize this stack: NVIDIA’s Spectrum‑XGS adds distance‑aware congestion control; Broadcom leverages its Jericho/StrataDNX silicon for deep‑buffer routing; Cisco’s P200 "Dark Pyramid" silicon powers the 8223 routers announced in late 2025. Early benchmarks show up to 1.9× improvement in NCCL all‑reduce, a critical operation for distributed gradient synchronization.

For hyperscalers, scale‑across is already a strategic imperative. Google’s Gemini Ultra, Microsoft’s AI WAN, OpenAI’s multi‑cluster GPT‑4.5 training, and AWS’s Project Rainier all rely on cross‑site connectivity to reach petaflop‑scale compute. The market response is evident: Dell’Oro reports a 40 % YoY surge in disaggregated WDM shipments in 2025, outpacing integrated gear. As AI workloads continue to outgrow single‑site capacity, scale‑across will become a foundational layer, prompting data‑center operators to redesign campuses and creating a lucrative, long‑term revenue stream for optics manufacturers.