Summary
- Microsoft reports production use of Cisco Silicon One P200-powered systems.
- Cisco and NVIDIA intend to develop an open 400G-per-lane pluggable optics design.
- Projected density and efficiency improvements are design claims, not demonstrated results for every data centre.
Cisco has detailed new switching, routing and optical interconnect developments at the Open Compute Project summit, including evidence that its Silicon One P200 platform is now running in production at major cloud operators. Microsoft is named as a production user of P200-powered systems. Cisco and NVIDIA also plan to work on an open 400G-per-lane pluggable optics design known as USFP, which remains a standards initiative rather than a universally available product.
The network layer matters increasingly as AI workloads span more accelerators and, in some architectures, multiple facilities. A large compute system may have sufficient processing power but poor utilisation if data cannot move efficiently between resources. Fibre links, switching capacity, buffering and congestion management are therefore physical infrastructure concerns. The challenge is different when connecting devices within a rack, racks inside one data hall and data centres separated by metropolitan or longer-distance links.
Cisco says its 51.2Tbps P200-powered systems are in production and that G300-based systems have begun shipping. These are distinct commercial states. An installed production deployment provides stronger evidence than a chip announcement, while shipping a system does not establish which sites are already running it. The company also outlines additional 1.6Tbps-related silicon and network interfaces that may support later designs.
The proposed USFP format aims to increase the amount of optical bandwidth available in constrained switch faceplate space. Cisco says it could provide four times the bandwidth density and approximately 1.3 times the power efficiency of certain existing 200G-per-lane approaches. Those are architectural comparisons and depend on how the full modules and equipment are configured. Fibre management, heat rejection and component replacement remain relevant even if the interface occupies less panel space.
Cisco's account includes a distributed AI networking concept that links capacity across sites. Compared with traditional data centre interconnect, a coordinated computing job may be more sensitive to delay, packet loss and uneven path performance. Network failure can leave expensive accelerators waiting for data. Deep buffering and alternative paths can improve tolerance of bursts and network impairment, but the value depends on workload and traffic patterns as well as hardware specifications.
An open optical interface can also reduce dependence on a single source, provided multiple manufacturers implement the standard consistently. The planned Cisco-NVIDIA specification will need adoption, testing and agreed interoperability requirements before buyers can treat modules as interchangeable. The partners have not said that every future USFP product will offer identical operating temperatures, reach or power consumption.
The distinction between scale-out and scale-across is important for infrastructure planning. Scale-out generally expands a computing fabric within a site or closely coordinated cluster, whereas scale-across seeks to link resource pools across greater distances. The latter imposes additional transport and failure-domain considerations, because a fibre cut or regional disturbance can affect a job spanning locations. High-capacity coherent optics can help provide the physical link, but software must decide how tasks are placed and how failures are contained.
The standards route proposed for USFP will need coordination across optical component suppliers, switch manufacturers and operators. Connector form factors influence physical density and maintenance, but performance also depends on power budgets and link reach. A specification that works for short intra-building spans may need different optics for metropolitan distances. Evaluation should therefore be based on the exact link class and traffic requirements rather than treating the 400G-per-lane figure as an all-purpose answer.
Production deployments can provide a useful reality check on vendor announcements. Cisco has named Microsoft as using the P200, offering a concrete reference for a shipping technology. The proposed optical format does not have the same level of evidence. European data centre procurement programmes may therefore be able to assess the deployed switching product sooner than the emerging optical specification, which still depends on qualification and vendor support across a longer equipment lifecycle.
Data centre operators assessing new networks should separate currently deployed P200 systems from announced products and proposed standards. Equipment tests will need to cover energy consumption under load, replacement time, congestion behaviour and failure recovery. Cisco's announcement supplies a production reference for one part of its portfolio and an intended path for higher-density optics, without making the two equivalent milestones.
A rack-scale deployment also places considerable demands on fibre documentation and fault localisation. Engineers need to know which links are inside an individual rack, which cross rows, and which pass through separate facility routes. The announced technologies address data movement, but operators will still need to validate loss budgets, connector counts, repair procedures and supported distances for the particular combination of optics and switches selected.
