AI's Hidden Bottleneck: The Network Between Your Data Centers
In this blog
Every conversation about AI eventually arrives at the same realization: the data center you have is not the data center you need. What is less obvious and far more consequential is that the network between your data centers is becoming just as strategic as the GPUs inside them.
This is the story of data center interconnect, or DCI. It is the part of your infrastructure that has not historically commanded boardroom attention, and the part that will quietly decide whether your AI investments deliver in the years ahead. The architectural choice that makes modern DCI possible is Segment Routing. And the decision window for planning, for budget, and for partner selection is now.
A new kind of data center problem
For two decades, "the data center" has meant one building, or one campus. Cloud rewrote that definition. Now AI is doing it again. Training a frontier model, serving inference at global scale, or simply keeping pace with the data growth that AI workloads create now routinely requires multiple data centers to behave as a single, coordinated platform.
The market is signaling the shift clearly. Data center capital expenditure surged 57% in 2025, driven almost entirely by AI build-outs. The optical transport market, the layer that physically connects those data centers, grew 20% year-over-year in the first quarter of 2026, prompting Dell'Oro Group to raise its full-year forecast to 16%. Direct DCI purchases alone grew an estimated 40%.
Those numbers tell you something simple: the companies that have already figured out their AI strategy are spending heavily on what connects their data centers. The traffic patterns AI creates (unpredictable bursts, massive east-to-west flows, latency-sensitive coordination between GPUs hundreds of miles apart) are not what enterprise WANs were ever designed to carry.
Proven at the most demanding scale in the world
Segment Routing has been running in production across the largest, most demanding networks on the planet for years: the same networks that now carry the training runs, inference workloads and east-west data flows that define what modern AI infrastructure looks like at scale. That is not a frontier bet. It is an architecture that has been stress-tested, refined and operationalized at a level that dwarfs any enterprise deployment, supported broadly across the major core networking platforms with a well-understood operational playbook. Any organization running AI across more than one location (primary plus disaster recovery, on-premises plus colocation, or private data center plus public cloud) stands to benefit directly from that body of validation. The architectural choices made now will dictate what your network can do for the rest of the decade.
The architectural shift behind a modern, AI-ready WAN
The most useful way to think about Segment Routing is not as a protocol, but as a shift in how the network behaves.
A traditional network is a committee. Every router consults with its neighbors, negotiates the path, and hopes everyone agrees. It works, but it is fragile, slow to change, and difficult to engineer for the kind of performance AI demands.
A network running Segment Routing is more like a guided tour. The point of entry decides the route, writes it into the packet, and every router along the way simply follows the directions. That single shift unlocks a set of business outcomes that matter:
- Predictable performance. Critical workloads can be pinned to specific paths with specific service levels. Tail latency (the silent killer of AI training efficiency) becomes something you can manage rather than hope for.
- Faster recovery from failure. When a fiber cut or hardware failure happens, the network reroutes in milliseconds without operator intervention or human delay.
- A single operating model. Segment Routing consolidates multiple legacy WAN protocols into a single, consistent architecture, replacing complexity with clarity across your edge and core. Lower operational overhead, fewer specialized tools, and a network that's easier to troubleshoot when something goes wrong.
- Programmability. New services (network slices for different AI workloads, secure paths for sensitive data, premium routes for inference) can be added in software rather than re-engineered in hardware.
- Standards-based, not proprietary. Segment Routing (whether SR-MPLS for existing infrastructure or SRv6 for modern greenfield builds) is an open, standards-based architecture supported across all major platforms, with no vendor lock-in and no proprietary hardware requirements.
Those are the outcomes that show up on the executive scorecard, and they are the outcomes your AI program will live or die on.
The window is open now
The question isn't whether to modernize. It's how fast, and from what starting point. Segment Routing has an answer for both. If your network has a healthy installed base of MPLS hardware, Segment Routing layers on top of it cleanly. You can capture the operational and performance benefits without rebuilding what is already paid for, and you can plan the modernization to align with your normal refresh cycle.
If you are building new AI infrastructure (a new region, a new fabric, a new colocation footprint), the right answer is to skip the legacy step and adopt the modern, IPv6-native version of Segment Routing (SRv6) from day one. That is the path the hyperscalers are taking, and it is the path that supports the multi-cloud, multi-data-center realities of an AI roadmap.
Most enterprises will need a mix of both, and that is perfectly fine. The earlier you engage, the more runway you have to get it right. The decisions made this year and next will shape what your network can do for the rest of the decade. Getting it right is a conversation worth starting now.
Where WWT comes in
The complexity of getting from the network you have today to the one your AI strategy needs is real. Every business arrives at this conversation from a different starting point: existing investments, operational maturity, the AI workloads on the roadmap, the partners already in the environment. There is no universal answer, and there is no shortcut to discovering yours. WWT brings the Segment Routing and DCI expertise, the lab environment, and the partner relationships to help you translate your AI requirements into an architecture that fits your business, your timeline, and your budget.
Your team stays at the center of this decision. A typical first engagement starts with a briefing or workshop to align on architecture direction and identify the right path for your business. From there, WWT can validate the proposed design in the Advanced Technology Center, where the architecture can be built and tested before any production budget is committed.
AI workloads no longer respect data center walls and the network that connects your data centers is now the most consequential infrastructure decision of the decade. The shape of the AI-era network is becoming clearer every quarter, and the decisions made now will define what your network can do for the rest of the decade. If any of this lands, if you are already having this conversation, or sense that you should be, we would welcome the chance to compare notes. Reach out through your WWT account team or send a note to CoreNetworkingHelp@wwt.com.