High-capacity waves and IP transport for AI hubs and data centers: there is no one-size-fits-all

What is driving this

Training and serving AI models concentrates compute in a few places: a colocation hall, a cloud region, or increasingly a facility the customer builds and owns. Data has to move between those places and the places it is created and used, at capacities that make ordinary business internet irrelevant. The result is a wave of requests for 100 gigabit and 400 gigabit transport, often from organizations buying at that scale for the first time, into a market with its own vocabulary and its own constraints on where capacity can physically be delivered.

Three ways to buy capacity

Wavelengths (waves)

A wavelength is a dedicated optical channel on a carrier’s fiber between two points, sold at fixed capacities such as 10, 100 or 400 gigabits per second. You get a private, point-to-point path with predictable latency, without running the fiber or the optical equipment yourself. It is the common choice for data center interconnection and, as the example below shows, for reaching a better market from a worse one.

Dark fiber

Dark fiber is the strand itself, leased unlit. You buy and operate the optical equipment and light it to whatever capacity the equipment supports. It offers the most control and the best economics at very high capacity, with the most responsibility: equipment, monitoring and repair coordination become yours.

IP transport and transit

IP transport delivers routed capacity between your sites over a carrier’s network, or to the internet (transit) through a Tier 1 provider. It is simpler to operate and easier to scale, and its price depends heavily on where you buy it: a 400G-enabled point of presence in a major interconnection facility is a very different market from a remote building served by one carrier.

A real design: 800G of diverse internet to a remote data center

A customer building a new data center in Kentucky needed 800 gigabits of diverse internet capacity into the facility. The location was remote, with access to a single regional carrier. The customer had asked that carrier how it would deliver the capacity and could not get an answer.

We took over the project and designed it differently. Instead of trying to buy 800G of internet at a location where the market could not supply it, we custom-designed two 400G wave connections out of the facility: one to a data center in Ashburn, Virginia, and one to a data center in Columbus, Ohio. Both of those facilities have some of the highest-quality Tier 1 internet providers co-located in the same building, with 400G-enabled points of presence, at a far lower cost per gigabit than anything available at the remote site. The customer’s internet capacity is bought where the market is deepest, and carried to the building over private waves.

The other half of the work was proving the diversity. Two waves are only diverse if they take different physical routes, and three carriers were involved across the two paths. We obtained the route files (KMZs) from each carrier, validated that the paths did not share conduit, fiber or facilities, and documented the design alongside the supporting service level agreements, so the customer could sign knowing exactly what “diverse” meant in the contract.

None of this is exotic. It requires technical proficiency with optical and IP transport, knowledge of which carriers can actually deliver at which facilities, and the project management to push a design involving several parties through to turn-up. The customer’s original carrier was not wrong to struggle; it simply did not have the market, and the design had to go and find it.

Diversity: the specification that matters

A single high-capacity link to a compute facility is a single point of failure for an expensive asset. Diversity means two paths that do not share a conduit, a fiber route, a building entrance or a carrier facility. Carriers describe diversity in different ways, and “diverse” sometimes means two wavelengths on the same fiber. Ask for route maps. Ask which building entrances are used. Ask whether the two paths share any point of presence. Validate the KMZs yourself, or have someone do it for you, and put the answer in the contract.

Latency, and the honest way to specify it

For training clusters split across sites, and for serving workloads with response-time targets, latency between locations is a design input. Specify it as a measured round-trip time on the actual route, with a commitment in the service level agreement, rather than as a straight-line distance. A route that detours through a distant hub can double the latency of a shorter one, and the Kentucky-to-Ashburn path and the Kentucky-to-Columbus path will not be the same.

What decides the schedule

• Lead times. High-capacity services can take months, longer if construction is needed to reach the building. Order early, and ask what is on-net at each address before choosing the site.
• Power. AI hardware is dense; the facility’s power and cooling terms often constrain the project before the network does.
• Cross-connects. The link between your cage and the carrier’s equipment in the same building carries its own fee and its own lead time.
• Equipment. Optics and routers for 400 gigabit interfaces have their own supply timelines. Align the circuit date with the hardware date.
• Coordination. Two waves, three carriers and two remote facilities means one project plan with every party on it, and one advocate chasing every date.

Contract terms worth negotiating

• Term length against the life of the AI project; shorter terms with upgrade rights beat long terms at today’s capacity.
• Upgrade paths: the right to move from 100 to 400 gigabits, or to add a second 400G, without a new term.
• Service levels for availability and latency with automatic credits.
• Diversity language that names the routes and references the validated KMZs.
• Portability if the hub moves, and the cost of early termination.

How to run the purchase

Start with the flows: what moves between which sites, how much, how often, and what breaks if it stops. Turn that into a short specification: capacity, diversity, latency, term and dates. Then look at where the market actually is for that capacity, which may not be where the building is, and take the specification to the carriers that are on-net at the facilities that matter. Compare the responses on the same specification. The differences in price and in diversity are usually larger than buyers expect, and only visible when the asks are identical.