Cisco published an engineering account of how it built direct-liquid-cooled N9000 and 8000 systems around the Silicon One G300. The post describes 102.4 terabits per second of throughput, roughly a 70 percent energy reduction against the prior generation, and a water-glycol direct-to-chip loop running through the switch itself. The detail that matters most is a dedicated 2x8 optics cooling design for the 800G OSFP transceivers. Cisco built it by treating silicon, optics, and cooling as one system rather than three.
The framing is a milestone timeline. A prototype debuted at the Optical Fiber Conference in March 2023. The product was announced in February 2026. Authored by Cisco's Denise Lee with the data center architecture team, the piece sits inside a market projection of global data center electricity demand roughly doubling to 945 TWh by 2030.
In a Blackwell or Rubin class rack, the GPUs went to cold plates first. The CPUs followed. The top-of-rack and spine switching stayed on air the longest because switch ASIC power, while high, trailed accelerator power. Cisco putting the switch on a liquid loop closes that holdout. When the network gear needs direct-to-chip cooling, the rack no longer has an air-cooled island that the facility plant can treat as a rounding error. Every thermal interface in the rack is now a liquid interface.
The optics piece is the underrated part. 800G OSFP modules run hot, and they fail in ways that are sensitive to case temperature. A dedicated cooling design for the transceivers signals that optics thermal management is becoming its own discipline inside the switch, not an afterthought handled by chassis airflow. That is consistent with how direct-to-chip cooling actually works at the component level once power density crosses a threshold.
For operators and CDU vendors, this changes a planning assumption. Coolant distribution units and rack manifolds have largely been sized around compute heat load. If the switching tier is now on the same loop, the flow budget, the manifold port count, and the redundancy plan all have to account for network gear that used to be thermally invisible. This is the practical edge of Nvidia's wattage roadmap spreading outward from the accelerator to everything that sits next to it.
The post points at European waste heat regulation, including Germany's energy efficiency law, as a design driver. A switch on a water-glycol loop produces capturable heat the same way a cold-plate GPU does. As heat reuse mandates move from aspiration to compliance, the network tier becomes part of the recoverable thermal budget rather than exhaust thrown away in a hot aisle. Vendors that can document network-tier heat capture will have an argument that air-cooled switching cannot make at all.