AI will run in many places. The enterprise fabric should remain one.
Earlier this year at Cisco Live, our team spent an evening talking with several NeoCloud providers alongside colleagues from Cisco’s AI Infrastructure organization. We expected the conversation to revolve around GPUs, AI factories, and inference capacity.
Instead, the same request kept surfacing. Their enterprise customers wanted an easier way to connect AI workloads running in these environments into their existing Cisco SD-WAN networks.
They weren’t simply asking for another connection. These customers had already built routing, segmentation, security, and application policies into their SD-WAN fabric. They wanted that operating model to extend to AI infrastructure.
AI is distributing compute. It shouldn’t fragment the enterprise.
AI will not have one natural home #
The first wave of cloud computing concentrated a large share of enterprise workloads into a relatively small number of hyperscalers. Enterprise networking evolved with it. Cloud on-ramps made those environments easier to reach, while SD-WAN abstracted the transports connecting branches, campuses, data centers, and clouds.
AI is developing differently – from core to edge.
Hyperscalers remain important, but enterprises are also turning to specialized NeoCloud providers for GPU capacity and AI infrastructure. NeoClouds initially emerged around specialized GPU capacity. As enterprises move AI into production, these environments increasingly become part of enterprise infrastructure, where networking, security, performance, and policy need to follow the workload.
At the same time, more AI processing is moving closer to where people and businesses operate. You can already see this in retail, where stores are beginning to combine local inference with AI models and services delivered from the cloud. Similar patterns are emerging through edge, local, and deskside inference, wherever latency, economics, privacy, or sovereignty favor running AI closer to the business.
Where a workload runs will increasingly depend on performance, GPU availability, economics, security, and sovereignty.
AI will not have one natural home.
Don’t break the abstraction #
The value of SD-WAN has never been another way to build tunnels. It is the abstraction that lets enterprises use different transports and providers while defining networking and security intent on a common overlay fabric.
AI shouldn’t break that abstraction.
An enterprise should not need to learn a different networking model every time a team selects another AI provider. Nor should network and security teams recreate routing, segmentation, and policy because an inference workload happens to run somewhere new.
Cloud on-ramps extended this abstraction into public clouds. AI now expands the problem across a much broader set of destinations.
The more places AI can run, the less practical it becomes to engineer connectivity one destination at a time.
We saw that directly with NeoCloud providers. Some were considering their own transit capabilities to simplify enterprise connectivity. That solves part of the problem inside the NeoCloud, but the enterprise still needs to connect that domain back into its WAN and reconcile policy on both sides.
A better approach is to make the AI environment part of the enterprise fabric.
AI Transit connects specialized AI environments to the SD-WAN fabric #
That customer requirement is what Cisco AI Transit is designed to address.
Existing Cisco SD-WAN customers can use AI Transit to bring workloads running in NeoClouds, private and sovereign AI clouds, and enterprise AI factories onto their existing fabric.
Cisco SD-WAN Cloud OnRamp extends the enterprise fabric into public clouds. AI Transit builds on the same foundation to connect these specialized AI environments.
AI Transit can be deployed as a physical or virtual appliance, depending on the environment. In participating NeoClouds, where deployment is virtual, a Cisco secure virtual router is available as-a-service.
Once connected, customers can use SD-WAN application policy to classify AI and agentic traffic. They can then apply consistent routing, segmentation, security, performance, visibility, and geographic intent across the fabric. Secure Internet or private connectivity can be used while preserving one SD-WAN policy model.
We’re working with AI infrastructure providers including Armada, Cirrascale and Latitude to bring this model to life, giving enterprises greater choice in where they run AI without creating another networking and security domain.
The AI workload can live somewhere new without creating another network to operate.
This also matters for sovereignty. For some organizations, controlling where an AI workload and its data reside is only part of the requirement. They may also need to control where traffic is allowed to travel. Geographic intent therefore becomes another policy consideration alongside security and performance.
The best path may not always be the shortest one. It may be the path that satisfies all three.
The SD-WAN overlay becomes the connective tissue for distributed AI #
In an earlier blog, How AI is changing enterprise SD-WAN requirements, I wrote about how AI is changing network traffic itself. AI applications and agents create interactions that can be more distributed, latency-sensitive, and machine-generated, increasing the importance of workload awareness, assurance, security, and visibility in SD-WAN.
AI Transit addresses another part of that transition: connecting specialized AI environments to the common SD-WAN fabric.
Today, NeoClouds make the shift particularly visible. Tomorrow, AI will increasingly span infrastructure from core to edge – from hyperscalers, NeoClouds, and sovereign AI environments to enterprise AI factories and edge, local, and deskside inference.
Enterprises should be able to choose where AI runs without rebuilding networking and security each time that choice changes. With Cloud OnRamp and AI Transit, Cisco SD-WAN provides one policy-driven fabric for that world.
The challenge isn’t connecting AI to the enterprise. It’s extending the enterprise to wherever AI runs.
As AI distributes compute, the SD-WAN overlay fabric becomes the connective tissue that keeps the enterprise together.
Learn more about how we’re powering AI networking for today and tomorrow with Cisco SD-WAN.
Connecting AI environments with Cisco AI Transit #
What is Cisco AI Transit? #
Cisco AI Transit is designed to connect specialized AI environments to an enterprise’s existing Cisco SD-WAN fabric. Its purpose is to extend the enterprise’s established networking and security policy model to workloads running in those environments.
How does AI Transit relate to Cisco SD-WAN Cloud OnRamp? #
Cisco SD-WAN Cloud OnRamp extends the enterprise fabric into public clouds. AI Transit builds on that foundation to connect specialized AI environments, including NeoClouds, private and sovereign AI clouds, and enterprise AI factories.
Which AI environments can AI Transit connect? #
AI Transit is designed to connect workloads in NeoClouds, private and sovereign AI clouds, and enterprise AI factories to an existing Cisco SD-WAN fabric. Before planning a deployment, confirm support for the selected environment and provider.
How does AI Transit support consistent network and security policy? #
AI Transit connects the AI environment to the existing SD-WAN fabric. Customers can then use the SD-WAN policy model to address routing, segmentation, security, and application requirements. Teams should validate the supported controls and configuration requirements for their deployment.
How is AI Transit deployed? #
AI Transit can be deployed as a physical or virtual appliance, depending on the environment. The deployment plan should confirm the supported appliance, provider integration, and responsibility for managing the connection.
Why do geographic traffic requirements matter for AI workloads? #
Organizations may need to consider both where AI workloads and data reside and where their traffic travels. Network and security teams should document those geographic requirements alongside security and performance needs when evaluating connectivity.
What should network architects confirm before adopting AI Transit? #
Confirm the supported AI environment, deployment model, connectivity options, and operational responsibilities. Also validate the required routing, segmentation, security, traffic classification, visibility, and geographic controls against current product documentation.
Cisco Cybersecurity Viewpoints
Where security insights and innovation meet. Read the e-book, see the video, dive into the infographic and more…