search
close-icon
Data Centers
PlatformDIGITAL®
Partners
Expertise & Resources
About
Language
Login
banner
Blog

How is AI redefining interconnection and what must enterprises do to adapt?

16 November 2026
author-photo
Chris Sharp

Chief Technology Officer

Everyone's talking about AI compute. Fewer are talking about what connects it.

This quarter, a customer developing frontier artificial intelligence (AI) models committed to two very different deployments on PlatformDIGITAL®: a large-scale campus for training and a distributed deployment spanning multiple sites, each requiring significant capacity.

The customer also ordered more than 1,500 cross connects across the distributed deployment, illustrating the significant interconnection requirements needed to support AI at scale.

It's a good illustration of where AI infrastructure is heading. The campus trains the model. The distributed network puts it to work. Both are part of the same deployment, and both depend on connectivity that can operate at very different scales.

We're used to thinking about interconnection as the way infrastructure connects to other infrastructure. That's still true, however, AI is expanding the job. Models need to reach data, clouds, applications, and people: often across multiple locations and with little tolerance for delay.

The network is no longer plumbing. It’s part of the application.

Different AI workloads, different network demands

AI is often talked about as a single type of workload. It isn't.

Training a large model involves enormous volumes of data moving between processors, concentrated computing capacity, and high-bandwidth connections. We've seen that reflected in campus-scale commitments, including a 200 MW deployment earlier this year.

Inference introduces the opposite challenge. It brings computing closer to where it's needed: near the data and users it serves.

Consider agentic AI, where a single task might trigger 20 to 50 model calls. Those interactions can involve several models, databases, and services, and the time spent moving between them adds up. Network latency becomes something users experience directly, rather than a metric buried in an infrastructure report.

For enterprises, the priorities are often more specific. For example, a medical device company might deploy AI hubs across four metros to run inference close to regulated data. A financial services company may establish infrastructure across three trading hubs, with low latency access a key consideration. Sovereign AI also brings geographic constraints, as we've seen with deployments in Kuala Lumpur, Chennai, and Sydney where in-country placement was decisive.

There isn't one blueprint for all of this. What matters is whether the network can support the workload, wherever it needs to run.

Large and small, on one platform

What's particularly interesting about the referenced customer deployment is that it brings two ends of the infrastructure spectrum together.

Alongside its training campus and smaller sites, the customer standardized a multi-cross-connect product across 13 European metros. It doesn't choose between centralized computing and distributed delivery. It requires both.

And this isn't an isolated pattern. Across the AI opportunities we're seeing, demand spans both major campus deployments and smaller, distributed environments.

Activity in sub-1 MW deployments is also growing, including outside the largest customer commitments. This points to a broader shift in how AI infrastructure is taking shape: alongside large training campuses, enterprises increasingly need smaller deployments positioned closer to their data, applications, and users.

The largest campuses may attract the headlines, but AI infrastructure is becoming increasingly distributed. That makes the connectivity between these environments just as important as the compute within them.

The network must be part of the initial architecture, not an afterthought.

Depth in the metro, reach across the globe

A useful way to think about the network behind these deployments is in terms of depth and reach.

Inside a data center, cross connects provide direct physical links between systems. Across a campus or metro, fiber connects facilities that may be physically separate but operate as part of the same environment. Digital Realty has approximately 237,500 cross connects globally, with Pathway and Metro Connect supporting connectivity across campuses and metropolitan areas.

Then there's the wider ecosystem. A business might need to connect an AI deployment to cloud services, network providers, or models hosted in other markets.

ServiceFabric® extends access across 850+ data centers in 160+ metros, with 350+ cloud on-ramps, 100+ network service provider partners, and 180 service directory partners. Our partnerships with Megaport, Console Connect, and Zenlayer extend that reach into additional cloud and AI environments.

Customers are increasingly needing both: direct, dependable physical connections and the ability to reach services across markets without building every connection from scratch.

Pathway fiber for depth across the campus. ServiceFabric for reach across the globe.

Agents don’t file tickets

There's a further implication of agentic AI that I think deserves more attention: The network increasingly needs to respond to software, not just people.

Today, provisioning a connection can involve a request, an approval, and a series of manual steps. That may be manageable when infrastructure requirements change occasionally. It's less practical when applications need to access services dynamically as part of an automated workflow.

Take 50 milliseconds off each of 20 to 50 sequential model calls and a task could finish 1 to 2.5 seconds sooner. At that point, latency isn't simply a networking concern. It can influence how useful the application is.

Programmability is therefore becoming an important part of interconnection.

Since June 2026, ServiceFabric has supported an open standard that allows agents to discover, provision, and manage private connections in minutes, with sovereignty policies enforced through the platform. Consumption-based pricing is now live, too, giving customers more flexibility in how they use those connections.

We're seeing this move beyond theory. In Barcelona, Blackfuel is running a native AI inference platform in production on Dell and AMD infrastructure, using liquid cooling and ServiceFabric connectivity.

Fiber and cross connects remain the physical foundation. What's new is the ability for software to interact directly with the network built on top of them.

Building for what AI will become

AI isn't heading towards one deployment model. The mix of centralized training, distributed inference, private environments, and autonomous applications will keep evolving.

The opportunity is to make those environments work together without connectivity becoming a constraint.

That calls for physical connections, metro networks, global reach, and programmability to be considered together. That way, the architecture can follow the workload rather than the other way around.

AI will keep changing shape and demand more of digital infrastructure. Interconnection must be ready for whatever comes next.

Learn more about ServiceFabric® and our connectivity.


Tags