Wherever the GPUs Go, the Network Has to Follow

Inside the facility first, outside next

Today, most AI traffic stays inside the building. A text prompt and its response may amount to a few megabytes or less. The heavy communication happens between GPUs and storage systems within the AI facility.

Meta's AI Research SuperCluster illustrates the scale. Its GPU systems communicate over a 1,600 Gb/s InfiniBand fabric with no oversubscription, supported by a storage service built to serve terabytes per second of bandwidth.¹

AI training has already made the data center network essential. Widespread inference will make the external network strategically essential, with carrier, metro and subsea demand becoming meaningful between roughly 2027 and 2030. The distinction matters, because each network layer reaches that point on a different timeline.

Three shifts that move AI traffic outside the building

The inflection point for external networks arrives as three developments scale together.

Metro networks become the AI on-ramp

Inference places AI closer to the applications and people it serves. Traffic needs efficient paths between AI infrastructure, enterprises, cloud platforms, network exchanges and population centers.

Metro fiber becomes the on-ramp and off-ramp for AI. For carriers, that creates demand for more than bandwidth. It calls for route diversity, lower latency and scalable connections into AI-ready facilities.

Where compute is located matters. So does the network architecture surrounding it.

Carrier networks have to scale with compute

A high-density AI rack concentrates tremendous computing power in a small physical footprint. The effects of that concentration reach well beyond the rack.

As deployments grow from individual installations to multi-megawatt environments, carriers must be ready to carry the traffic those environments produce. That requires higher-capacity optical systems, additional fiber, diverse routes and stronger interconnection ecosystems.

The relationship between the data center and the carrier changes as a result. Connectivity stops being a service delivered into the building.

The carrier ecosystem becomes part of the facility's ability to support AI at scale.

AI is global, and subsea infrastructure carries it

AI runs inside a physical building, but the data it depends on does not stop at borders. Enterprises operate internationally. Cloud platforms span continents. A model can be trained in one market, refined in another and accessed by users around the world. Sovereign AI requirements add another layer, as models and data are replicated within specific jurisdictions.

Subsea cable systems carry that traffic between regions, and their share will grow as AI increases the volume of data moving across borders. Facilities positioned close to subsea infrastructure reduce unnecessary network hops and create more direct paths between domestic and international networks.

As AI deployments scale, the connection between AI infrastructure and subsea infrastructure becomes a design decision rather than an afterthought.

The commercial window opens sooner

The external network becomes strategically essential over the next two to four years. The commercial opportunity begins earlier, between 2026 and 2028, as AI tenants start to require:

• Multiple diverse 100G, 400G and 800G routes
• Direct cloud and GPU cloud connectivity
• Low-latency access to New York and Ashburn
• Dataset and checkpoint transport
• Private connectivity between AI campuses
• Subsea access for sovereign and international AI workloads

None of these requirements are solved by adding GPUs. Each one is a network decision.

NVIDIA describes the AI factory as infrastructure that manages the full AI lifecycle, from data ingestion through training, fine-tuning and high-volume inference.² Every stage of that lifecycle moves data, and increasingly it moves data between sites.

AI is turning compute facilities into networked factories. NJFX provides the carrier-neutral middle mile connecting those factories to clouds, enterprises, subsea systems and other AI campuses.

At our carrier-neutral cable landing station in Wall Township, New Jersey, high-density infrastructure, terrestrial fiber, carrier networks and subsea systems already operate as parts of the same architecture. Through SecureWAVES, customers and network operators can plan beyond adding capacity and build the bandwidth, diversity and resiliency AI requires as it scales.