Hyperscalers want 800G
- Cisco's chief architect described 'scale-across' networking connecting many GPUs, which drives new bandwidth patterns. - He estimated these designs need roughly 14× the bandwidth of traditional DCI and demand 800G pluggables plus deep-buffered switches. - That bandwidth profile creates multi‑billion dollar opportunities for optics and switch vendors selling high-speed interconnects (x.com).
The new bottleneck in artificial intelligence is moving data between data centers, and Cisco says that job now points hyperscalers to 800-gigabit links. (lightreading.com) In Cisco’s framing, “scale across” means linking GPU clusters spread across multiple facilities when one campus runs out of power or floor space. Bill Gartner, Cisco’s senior vice president and general manager of optical systems and optics, told Light Reading that this inter-data-center traffic needs about 14 times the bandwidth of a traditional data center interconnect. (lightreading.com) Cisco made that pitch public on October 8, 2025, when it launched the 8223 routing system and Silicon One P200 chip for distributed AI workloads. The company said the fixed router delivers 51.2 terabits per second and is shipping to initial hyperscaler customers. (prnewswire.com) The plumbing matters because AI clusters no longer fit neatly inside one building. Cisco said on February 10, 2026 that its newer G300 silicon and updated Nexus 9000 and Cisco 8000 systems are aimed at “gigawatt-scale” AI clusters, with 102.4 terabits per second of switching capacity and support for high-density optics. (newsroom.cisco.com) An 800G optical module is the laser-powered plug that turns electrical signals from a switch into light on fiber. Cisco’s 2025 Cisco Live optics deck said 800-gigabit ports were set to become the majority of switch ports in AI back-end networks in 2025, with 1.6-terabit ports taking over by 2027, citing Dell’Oro Group research. (ciscolive.com) That speed shift is already reshaping vendor forecasts. Dell’Oro said in January 2025 that switch sales in AI back-end networks are expected to exceed $100 billion over five years, with most deployed ports moving from 800G in 2025 to 1.6T in 2027 and 3.2T in 2030. (prnewswire.com) Cisco is not the only company chasing that spend. Next Platform reported in February 2026 that Cisco’s new systems are part of a broader fight with Nvidia and Broadcom over Ethernet-based AI fabrics, while Dell’Oro has said Ethernet switch sales in AI back-end networks more than tripled in the fourth quarter of 2025. (nextplatform.com, ainvest.com) Cisco is also arguing for deep-buffered switches, which hold more packets in memory when traffic arrives in bursts from synchronized GPU jobs. Fierce Network reported in October 2025 that Cisco’s scale-across design leans on deep buffering, a choice that could split the market from lower-buffer designs favored elsewhere. (fierce-network.com) The thread running through all of it is simple: bigger AI clusters are turning data-center interconnect from a carrier niche into core compute infrastructure. If hyperscalers keep stretching clusters across campuses and regions, 800G optics and the switches behind them move from upgrade cycle to capacity requirement. (lightreading.com, newsroom.cisco.com)
