Cisco and Nvidia are building optical networking for data centers where hundreds of thousands of GPUs work as one

Cisco and Nvidia are preparing a new generation of optical networking technology for data centers where AI workloads increasingly hit a limit in the link between chips rather than in the chips themselves. Cisco President and Chief Product Officer Jeetu Patel told Bloomberg Tech (host Ed Ludlow) that the project aims to make hundreds of thousands of GPUs work as a single system while cutting both the network's power consumption and the cost of AI.

Few technical details have been made public so far: the material gives no timelines, prices or port speeds. The underlying logic is clear, though, and it is telling.

The bottleneck has moved to the network

Once a cluster grows to tens or hundreds of thousands of accelerators, efficiency depends on how fast the GPUs can exchange gradients, weights and intermediate results. If the network can't keep up, expensive accelerators sit idle waiting for data. Money burns on empty waiting.

Hence the interest in optics. At high speeds, copper links work over ever shorter distances, while traditional optical modules draw a lot of power. In a cluster of hundreds of thousands of GPUs, tens of watts per port add up to megawatts. For operators already fighting over available electrical power, that is a direct constraint.

Agents as a new type of traffic

Patel made a separate point about the workloads themselves. In his words, increasingly powerful AI agents can consume far more bandwidth than humans doing the same tasks.

The logic is easy to follow. A person asks a question and reads the answer. Meanwhile, an agent makes a series of calls to models, fetches data, invokes tools, checks its own steps and often launches parallel branches of reasoning. A single user request turns into an avalanche of internal traffic between services, storage and inference clusters. The more autonomous the agent, the stronger the effect.

If that reasoning holds, networks will have to be planned not by the number of users but by the number of agents and how "chatty" they are.

Why Cisco and Nvidia

Nvidia controls a large part of the AI cluster stack, from accelerators to its own networking products. In 2025 the company showed switches with photonics, where the optical components sit next to the switching chip rather than in removable modules. This shortens the electrical signal path and reduces energy loss. Cisco brings networking expertise, equipment for carrier and enterprise networks, and its own silicon platform. For Nvidia, the partnership is another route into enterprise and cloud data centers; for Cisco, it is a chance to stay at the center of AI infrastructure.

The new generation Patel describes continues in this direction. What exactly will go into the product will only become clear from future announcements.

What we know so far

• Partners: Cisco and Nvidia, jointly developing next-generation optical networking technology.
• Scale: clusters of hundreds of thousands of GPUs.
• Goals: faster links, less energy spent on data transfer, lower cost of AI.
• Demand driver: agentic workloads that generate significantly more traffic than a human.
• Not disclosed: release timeline, prices, port bandwidth, compatibility with existing hardware.

What teams building AI infrastructure should consider

First: count the network as part of a cluster's total cost of ownership, not as an add-on to the GPU purchase. Accelerators idling because of a weak network cost more than the network itself.

Second: build agent traffic into your workload models. The request-response profile no longer works for services with autonomous agents, and internal calls can exceed external ones many times over.

Third: measure the energy spent on data transfer separately. If the limit on your data center is megawatts, lower-power optics can free up capacity for additional accelerators.

And fourth: don't rush into contracts before there are numbers. Until Cisco and Nvidia name timelines and specifications, it makes sense to follow the announcements and compare them with solutions that are already available. For developers and businesses that buy AI as a service, the effect will be indirect: if networks really do get cheaper and more efficient, it may show up in token prices, but not immediately.