The Internet's New Factories

For most of the Internet’s history, a data centre was a place the network connected. But that relationship changes with the introduction of AI factories. Accelerators, storage and data pipelines depend on several network layers, and when compute is spread across halls, campuses, regions, etc., the links between them become part of the computing system itself.

In The Internet’s New Builders, I wrote about content and cloud platforms moving from buying capacity to building cables and private backbones. The Internet’s New Edge Builders followed many of the same companies closer to users, where their own caches and delivery infrastructure now operate alongside traditional CDNs.

AI brings the same shift into compute itself. For many operators, AI will show up first as a request for private fibre, dedicated wavelengths and data-centre interconnection, or as crawler and agent load on their origins, long before it shows up as a jump in public transit.

In this article, I want to show how these systems use networks, why much of that activity stays out of public view, and what this means for the operators connecting them.

The network becomes part of the machine

An AI factory is made up of accelerators, storage, data pipelines and several separate networks all run as one system. Inside one campus, those networks carry model state, datasets and checkpoints. When compute is distributed across several campuses, dedicated optical links can help those sites operate as a larger pool.

Availability of power is a strong deciding factor in where this system can be built. The International Energy Agency uses roughly 10–25 MW as the scale of a conventional data centre and 100 MW or more for a hyperscale AI-focused facility, with individual projects now reaching gigawatt scale. Facilities of this size cannot just be added to an established Internet hub. Rather, their location has to bring together available electricity, cooling, land and several independent fibre routes.

Google states the logic plainly: moving electricity is harder than moving data over fibre, so it locates data centres near energy sources and uses its network to pool AI compute across campuses. Microsoft similarly links its Fairwater sites in Wisconsin and Atlanta through a dedicated optical AI wide-area network. And at RIPE 92, Meta described how power and space constraints were already spreading AI infrastructure across wider metropolitan areas.

AI traffic is not one flow

'AI traffic' is a label for several systems that behave differently and are visible from different points. For an operator the practical question is where each one can be seen: inside the factory, between campuses, or on the public Internet.

Inside the factory, accelerators exchange model state within a rack, cluster fabrics carry scale-out operations between racks, and storage feeds datasets and checkpoints to compute. The volumes are enormous: Google says its Virgo fabric can connect up to 134,000 TPU 8t chips with as much as 47 Pb/s of non-blocking bisection bandwidth in a single fabric. Meta has described checkpoint bursts that congested its data-centre network and left GPUs idle. None of this appears on an Internet transit graph, yet this is where the network decides how much of the compute actually works.

Figure 1. Google separates scale-up, Virgo scale-out and Jupiter front-end networks. Storage, general-purpose compute and connections to other campuses use the front-end network. Author’s illustration based on Google Cloud.

Between campuses, private interconnection and dedicated AI wide-area networks move checkpoints, replicas and distributed workloads over dark fibre or wavelengths. Meta says one pair of sites in its AI Backbone is twice the size of the global backbone it built during the preceding decade. For a fibre or DCI provider, this creates demand for private capacity with specific delivery dates.

On the public Internet, two flows reach operators directly. Inference sessions, where a small prompt returns streamed text, images, audio or video: in Cisco’s measurements, median inference flows lasted about twice as long as non-AI web flows. And agents and crawlers, which read pages, call APIs and trigger other services: in one controlled Cisco task, an agent generated up to 450% more total traffic than a manual workflow completing the same task. These high request rates can create operational pressure even when bandwidth use remains modest.

PeeringDB reported degraded service from AI scrapers and introduced rate limits for unauthenticated web access. Fastly found that more than half of the AI-bot requests it classified reached origin, compared with less than 9% of human requests.

AI changes the backbone and the edge

Training calls for large amounts of compute working together, so access to power, cooling and connectivity strongly influences where clusters can be built. Inference - i.e., running a model to answer a request - may benefit from being closer to users when response time matters. But cost, model availability and data-residency requirements also influence where that processing happens.

This is not the same model as CDN caching. A CDN can place reusable software, media and other objects close to users, while an inference request normally requires model execution. The nearest cache, the service ingress and the GPUs running the model can be in three different places. AWS Bedrock cross-region inference, for example, can accept a request in one region and process it in another over the AWS network.

For the same reason, an operator-hosted cache does not automatically become a GPU edge. At RIPE 92, Meta said it did not expect forward-deployed appliances inside ISP networks to host power-hungry GPUs, and estimated that five to eight European locations could serve 99% of its current AI use cases.

Figure 2. Meta’s AI Backbone connects sites through optical multiplexing and protection switching. Its published design supports 64 × 800G, or 51.2 Tb/s, per fibre pair. Author’s illustration based on Engineering at Meta.

Who builds the factory, and who connects it?

An AI factory may carry the name of a cloud or AI company, but it is assembled by several parties. The compute operator defines which sites must work together, how much capacity and path diversity they require, and how quickly the system must expand. Utilities supply power, developers build the facility, and fibre operators provide the routes and optical systems. Power availability, permits and existing fibre still limit what can be built and where.

End users do not normally connect to the factory itself. They reach an API or service edge. The factory’s external links are more likely to connect other compute campuses, storage and data platforms, private clouds, enterprise customers and network providers.

The Microsoft–Lumen agreement shows this division of work. Microsoft selected Lumen to expand connectivity between its data centres through a custom private fabric combining access to existing fibre with construction on existing and new routes. In another model, Zayo says an anchor customer enabled 8,000 route miles of new builds and overbuilds, while Zayo will build, own and operate the routes.

The compute customer steers where the carrier invests without replacing it as builder and operator. What changes is the product: dark fibre, dedicated wavelengths, diverse site entrances, inter-campus paths, restoration commitments and capacity delivered in phases as the factory grows.

An anchor customer can make a new route possible, but it can also leave the carrier dependent on one project’s timetable. If that project is delayed or reduced, can the route serve other customers, and can later construction phases be deferred?

Public routing shows the edge, not the factory

An operator’s public ASN doesn't reveal the full infrastructure of its AI factory. BGP, the RIPE Database, PeeringDB and RIPE Atlas can show its public edge: announced prefixes, registered resources, declared presence and selected paths. They do not expose accelerator fabrics, private data-centre interconnection, storage traffic or workload placement. A traceroute can reach the service front door without revealing which campus or accelerator handled the request.

CoreWeave’s AS33425 represents part of its customer and interconnection surface, while Direct Connect carries private customer links into virtual private clouds without crossing the public Internet. Lambda lists AS11588 and AS398090, but separately documents its private network and 400 Gb/s InfiniBand fabric.

Public traffic measurements tell a different story from the scale of AI infrastructure investment. In one mobile network studied by Ericsson, 21 identified generative-AI applications accounted for 0.06% of bytes during one week in April 2025. AppLogic reported AI traffic growth of 90% on fixed networks and 80% on mobile, while still describing it as a small share of total traffic. Interviews with public IP operators likewise found no massive AI-driven surge.

Both are right. Public AI traffic grows fast from a small base, and the largest transfers never leave the private fabrics.

Traffic to an "AI company ASN" tells operators little about the workload; the same cloud network can carry an AI API, storage, ordinary software services and unrelated customer workloads. Direction, flow duration, request rate and the type of interconnection can provide additional context, but do not necessarily identify traffic as AI-related.

Firebird: one regional example

When I mentioned Firebird AI in my first article, its Hrazdan project was still an early signal. Firebird announced the operational launch of the factory in August 2026.

Firebird says Team Telecom Armenia’s network can provide the project with capacity of up to 1 Tbps and scale further. The operator separately reports four diversified international directions from Armenia. These are published network-capability statements, not measurements of used traffic.

“For an AI factory, connectivity has to be planned alongside power and compute. Capacity needs to grow with each phase, with diverse routes to help maintain service when a link fails. That means designing for large data transfers and disruptions from the outset, as well as for everyday demand.” — Arpine Chalkian, COO, Firebird AI

Firebird operates the compute platform, while a regional operator supplies its external connectivity.

AI Factory France: connecting national supercomputers

In France, AI Factory France gives researchers and companies access to national supercomputers. In October 2026, its coordinator GENCI and the Bristol Centre for Supercomputing announced plans to connect France’s Jean Zay and Adastra with the UK’s Isambard-AI for federated learning, with technical support from NVIDIA. Each site would train on its own data and exchange model updates rather than transfer the underlying datasets. The connection is targeted for the end of 2026 and would let separately operated computing centres contribute to a shared training process.

What operators should prepare for

For access and CDN operators, the useful starting point is service behaviour: which automated requests reach origins, where they create load, and whether enterprise customers need private cloud connectivity. Request rate, errors and response time can expose pressure that a traffic-volume graph misses.

For fibre, DCI and hosting providers, three questions should be settled before capacity is committed:

• What must be transferred, and by when? Establish dataset and checkpoint sizes, frequency and completion windows; averages alone do not define the link requirement.
• What must continue operating after a route failure, and how quickly must the remaining services recover? Agree the required capacity on the surviving path and check for shared ducts and site entrances.
• When is the next capacity step? Align wavelength or fibre delivery with power and compute commissioning, and agree the lead time for expansion.

Conclusion

I set out to show how AI factories use networks, what stays out of public view, and what this means for the operators connecting them. Much of their activity takes place on private networks, so public traffic figures alone won’t tell operators what connectivity these systems need.

Companies investing in compute increasingly shape network development around their sites and workloads. A location far from an established Internet hub can attract substantial investment if power and connectivity are available. For fibre and DCI providers, this means planning routes, capacity and delivery dates alongside the customer’s power and compute deployment.

Notes and sources

The discussion of power and location draws on the IEA’s Energy and AI, Microsoft’s Fairwater architecture and the Microsoft–Lumen agreement. The network examples use Google’s Virgo and Jupiter architecture, Meta’s AI Backbone and storage publications, and its RIPE 92 presentation.

Figures 1 and 2 are the author’s illustrations based on the published Google and Meta designs. The capacities shown are design specifications, not measured traffic. The traffic discussion draws on reports from Cisco, Ericsson, AppLogic and Fastly, together with PeeringDB’s operational reports.

The Firebird example uses the company’s launch announcement and connectivity announcement. The photograph of its Hrazdan facility is credited to Firebird.

Disclosure: I am CEO of SkyTel, in which Team Holding reports a 40% strategic equity stake. Team Telecom Armenia, part of the same group, is Firebird’s connectivity provider. Firebird co-founder Alexander Yesayan is also a co-founder of Team Group.