AI Clusters and the New Economics of Data Center Optics

Optics is no longer a supporting accessory in the data center network. As AI infrastructure advances from 400G and 800G toward 1.6-terabit connectivity, optical components are consuming a larger share of network cost, power and operational risk.

In this episode of the Data Center Frontier Show, DCF Editor in Chief Matt Vincent speaks with Bill Gartner, Senior Vice President and General Manager of Cisco’s Optical Systems and Optics business, about how AI is changing the strategic role of optics.

Gartner explains that optics represented roughly 10% of a network port’s bill of materials at 10G. At 400G and above, the optics can cost more than the switch port itself. Reliability has also become critical: A single unstable link can force GPUs operating in parallel to stop, return to a checkpoint and restart. According to data Cisco has seen from hyperscale customers, link flaps can reduce GPU infrastructure efficiency by as much as 40%.

The conversation maps the AI network across three distinct tiers:

• Scale-up: Connections within the rack, carrying approximately 500 times the bandwidth of a traditional WAN environment.
• Scale-out: Connections between racks, commonly using 400G and 800G pluggable optics.
• Scale-across: Coherent optical connections between data centers as AI clusters expand beyond the power limits of a single facility.

Gartner also discusses Cisco’s 1.6T roadmap, routed optical networking, coherent pluggable optics and the emerging debate around co-packaged and near-packaged optics. These architectures promise lower power consumption and greater density, but introduce new questions involving interoperability, replacement and operational resilience.

Looking ahead, Gartner emphasizes that optics is not constraining AI network growth. It is enabling clusters to scale across racks, campuses and geographically distributed data centers, while the coming inference wave shifts the industry’s focus toward cost and power efficiency.