Dark Fiber, Lit Fiber, and the Terrestrial Backbone

Before data crosses an ocean it crosses a continent, and it does so on terrestrial fiber, the strands of glass buried along rail lines, highways, and pipeline rights-of-way. Understanding this layer requires a distinction that outsiders often miss: the difference between the fiber and the light in it.

A fiber-optic cable is a bundle of glass strands. Each strand can carry data, but only once it is lit, meaning that transmission equipment has been attached at both ends to send and receive light through it. A strand with no equipment attached is dark fiber. It is physically present, capable of carrying data, and carrying none. The distinction matters economically because these are two different products. Buying lit capacity means renting a service: a certain bandwidth between two points, with the operator running the electronics. Buying or leasing dark fiber means acquiring the physical strand and supplying your own electronics, which grants control over how much capacity to extract and how to upgrade it over time.

For a large compute operator, this distinction is strategic. A single pair of fibers can carry vastly different amounts of data depending on the transmission equipment attached, because a technique called wavelength-division multiplexing sends many separate channels of light down the same strand at once, each on a slightly different color. As the electronics improve, the same dark fiber can be made to carry more without digging up any ground. Whoever controls the fiber captures those upgrades. Whoever merely rents lit capacity pays for them indirectly. This is why operators building at scale increasingly want to own or hold long leases on dark fiber along the routes that matter to them.

The scarcity here is subtle. There is a great deal of fiber in the ground in the developed world, much of it laid during earlier construction booms and never fully used. But fiber is only valuable where it goes where you need it, and where enough strands are available on that specific path. A route between two ordinary cities may have abundant dark fiber. A route into a newly-chosen data-center site in a rural area chosen for cheap power may have almost none, and building new terrestrial fiber means securing rights-of-way, trenching or hanging cable, and negotiating with every landowner and jurisdiction along the path. The physical build is slow and the permitting is slower.

This creates a recurring tension in the siting of AI compute. The best places to put a large cluster are often chosen for reasons that have nothing to do with connectivity: available power, cheap land, favorable climate for cooling. But those same remote virtues often coincide with thin fiber. An operator can find a site with abundant power and discover that connecting it to the fiber backbone will take longer than building the data center itself. The network, once again, is the constraint that arrives late and unbudgeted.

The terrestrial backbone also determines how clusters in different locations can cooperate. As single sites strain against the limits of available power, operators have begun to consider distributing a training run across multiple data centers. That only works if the sites are connected by enough bandwidth at low enough latency, which throws the spotlight directly onto the long-haul fiber between them. The distance imposes a latency floor set by the speed of light in glass, and the available fiber imposes a bandwidth ceiling. Whether two sites can act as one machine is a question answered by the terrestrial network between them.

The durable lesson is that terrestrial fiber is a slow-moving, physical, right-of-way-bound asset, and that control over it is control over the ability to deliver compute where the power is. Silicon can be procured in months. Fiber routes are measured in years. When those timelines collide, the fiber wins, because you cannot light a strand that has not been laid.