Everyone’s talking about AI fabric. No one’s talking about what it connects to.

The AI industry is increasingly being built on a layer it cannot see. Models, fabrics, agents, gateways, and exchanges—these are all evolving quickly. The assumption is that compute already exists where it is needed, at the scale it needs, with the power, cooling, and proximity that AI requires. That assumption may no longer hold in many cases.

We have built an industry that talks about AI infrastructure as if it were a connectivity problem:

“Connect this model to that GPU”
“Route this inference request through that gateway”
“Federate this policy across those clouds”

All of it is real, and all of it matters, but there is a more fundamental question that should be asked out loud: Where does the workload actually run?

The answer, increasingly, is in places most traditional networks were never designed to reach, with characteristics that have no fixed duration, no static throughput requirements, and variable performance needs. This environment should be highly elastic and dynamic to serve ever-changing data requirements.

This is not theoretical. AI infrastructure is constrained by physical reality: power, cooling, land, and in-country availability. These are not software problems.

Scaled enterprise AI needs more than a fabric

Enterprise AI is converging on multi-megawatt footprints.

Inference at scale needs power density
Training needs liquid cooling
Agentic systems need a persistent, low-latency state across regions
Sovereign AI needs in-country capacity, not in-country routing rules

None of this is expected to be fully solved by a fabric.

There is a narrative that the next category of AI infrastructure is a private exchange, a hub, or a fabric layer that allows enterprises to plug into AI providers without crossing the public internet. That is half of the story.

The other half of the story is about the layer underscoring the fabric. The foundation layer should be purpose-built, multi-megawatt, with AI-ready capacity, and distributed across the data oceans where AI compute, data, and connectivity converge. Without that layer, the fabric is connecting clouds that do not exist at the scale AI demands.

Private AI Exchange (AIPx), our underlying architecture, is patented for agentic control first and was put into production before the market began describing it as a new category.

Fabric-first approaches assume capacity already exists. In practice, it often does not meet the density, proximity, or sovereignty enterprise AI requires.

Our strategy is straightforward: build the base, rent the spike, and make the foundation programmable. This is not just an interconnect business. This is an infrastructure business.

I have lived this pattern before. I helped pioneer the first private cloud interconnect, built a global cloud exchange, and helped stand up the cloud business that came with it. The lesson from that work is one the market is relearning: connectivity matters because of what it connects to. Workload creates the need for fabric, not the other way around.

What enterprises need are three things in one place:

The capacity to run AI workloads at the density they require—not racks, but megawatts
The private interconnection to make those workloads usable across regions, clouds, and providers, on an open platform rather than a single vendor's fabric
The geographic distribution to satisfy sovereignty with real in-country capacity, not routing announcements

The only providers who can deliver all three are those who built the underlying capacity first and added connectivity on top—not the ones who started with connectivity and are now hunting for capacity.

Digital Realty has been building this for two decades. We are also our own ‘customer zero’, running AI workloads on our own platform. Insights from internal deployments directly inform our roadmap and accelerate the development of the most critical features to enterprise Private AI environments.

From our experience, we know enterprise AI needs a strategic foundation. This is why our roadmap includes:

Capacity: 3.5 gigawatts of capacity under contract
Pipeline: 6 gigawatts of future development capacity
Efficiency: An industry-leading power usage effectiveness (PUE) of 1.38 and water usage effectiveness (WUE) of 0.59—the first corporate disclosure of both at this level of rigor
Global footprint: 17.8 million square feet of certified data center space across every region, anchoring enterprise AI, North America, EMEA, and Asia-Pacific
ServiceFabric^®: Our open, partner-extended interconnection platform, that runs on top of it all, accessible through the industry’s largest ecosystem of network and cloud partners. ServiceFabric extends this foundation, enabling private, open interconnection because the underlying capacity exists first.
Partner solutions: We’ve recently launched the AI Solutions Program with partners including ePlus, Lenovo, and Dell, built on infrastructure from silicon leaders such as NVIDIA and AMD

The third exchange, the AI exchange that comes after the internet exchange and the cloud exchange, is not going to be a fabric. It is going to be a place. Many places. Wherever AI workloads need to live at scale. Connected, sovereign, open, and built for the density the AI era actually requires.

That is the foundation. We are building it.

The foundation for AI: Why the third exchange needs more than fabric