Skip to content
Friday, October 23, 2026 · Today’s edition, 6:30 a.m. ETFri, Oct 23 · 6:30 a.m. editionSign inMy Desk 0EditionsSign in

AI's fiber ceiling is set in a shed the industry barely counts

The long-haul routes feeding AI traffic run into their real ceiling at the amplifier huts spaced every 80 to 100 kilometers, where power draws are climbing from hundreds of watts to 12-kilowatt cabinets and no capex line obviously owns the bill.

At a glance

  1. Data Center Dynamics describes these in-line amplifier sites as thousands of tiny data centers that go unnoticed by the world, which raises the question of whether they are ready for the age of AI.
  2. A small network point of presence on a long-haul route, an ILA exists to give space, power and cooling to the inline amplifiers inside DWDM systems.
  3. The physical plant has not moved: a mix of containers, pre-cast concrete, and other prefabricated huts, nondescript by design, largely unchanged since the 1970s and a far cry from a modern data center, in Data Center Dynamics' telling.

Every 80 to 100 kilometers along the world's long-haul fiber routes sits a shed that goes largely unmentioned even within the industry, its contents a stripped-down data hall: racks of networking gear, redundant air-cooled units, uninterruptible power supplies, a diesel generator, security monitoring. Data Center Dynamics describes these in-line amplifier sites as thousands of tiny data centers that go unnoticed by the world, which raises the question of whether they are ready for the age of AI.

A small network point of presence on a long-haul route, an ILA exists to give space, power and cooling to the inline amplifiers inside DWDM systems. Those amplifiers are compact modules holding coils of erbium-infused fiber; lasers irradiate the isotopes so they release energy that re-powers the light, and spaced every 80 to 100 kilometers to offset attenuation, they keep it strong enough to travel hundreds of kilometers without being converted back to electrical form. That is how a fiber route carries more data over longer distances.

They are held to data center standards because, without them, the long-haul inter-city links between data centers fail. "They are crucial," Patrik Gylesjö, head of tech and delivery at the Nordic fiber firm GlobalConnect, tells Data Center Dynamics. "They have everything that a big data center has, but in a smaller format. They have that A and B power. They have uninterruptible power supplies. They have diesel generators."

The physical plant has not moved: a mix of containers, pre-cast concrete, and other prefabricated huts, nondescript by design, largely unchanged since the 1970s and a far cry from a modern data center, in Data Center Dynamics' telling.

The curve from 800 watts to 12 kilowatts

The deciding number is watts per rack: early ILA designs supported roughly 600 to 800 watts, while today's designs typically draw 2 to 4 kilowatts, a three- to five-fold increase, and future designs are expected to reach 5 to 6 kilowatts, with up to 12 kilowatts per cabinet contemplated for high-density hyperscaler deployments. A site generally totals around 24 racks, so a single hut is a modest load, while a route — huts at 80-to-100-kilometer spacing for its whole length — is a different proposition.

That curve repeats the pattern the rest of the digital buildout knows: the power envelope that made grid capacity the binding constraint on new data halls is now landing on a far smaller asset. Grid permission has become the thing that gets digital infrastructure built, with interconnection positions and power rights pricing ahead of electrons; in-line amplifiers make that logic awkward. An AI campus fights for one interconnection, while a long-haul route depends on many at close intervals, and at each the question is whether the hut has the power feed, the cooling headroom and the footprint to move from today's 2-to-4-kilowatt racks toward the 5-to-6-kilowatt designs expected next and the 12-kilowatt cabinets contemplated for the densest hyperscaler work.

The huts matter because they are what let a route carry more data over longer distances in the first place: an amplifier site that cannot take the next increment of power and cooling sets the ceiling for the span it sits in, and the glass may be capable of more, but a route can only pass what its amplifiers can re-power. That makes the upgrade a route-wide capex program, with thousands of small sites each needing more of the three things hardest to add to a prefabricated box: power, cooling and floor space.

An operator that leaves its sheds at 2 to 4 kilowatts is banking the fiber it already owns while capping the product it can sell over it.

The capex that is easiest to defer

Infrastructure of this shape has no natural champion: scattered rather than clustered, small in dollar terms at any one site, essential in a way that becomes visible mainly when one fails, and none of that wins a budget fight against a new route or a new campus. All of it argues for treating the amplifier layer as a program, not a line item that can slip a year.

Amplifier-site upgrades are being deferred precisely because they are easy to defer: the spend is small against a hyperscale campus, the sites are invisible from the road, and there is no ribbon to cut. That is the middle of the network getting starved while the endpoints get funded. An operator that leaves its sheds at 2 to 4 kilowatts is banking the fiber it already owns while capping the product it can sell over it.

The answer an investor can act on is written in watts per rack, and those numbers are being set one prefabricated hut at a time; the first long-haul route announcement that names its amplifier-site program the way campus announcements now name their power will mark the moment the sheds stopped being deferrable.

Reported by the Data centers & fiber desk from Data Center Dynamics. Published . Corrections

Related

Infrastructure

Krolmap plans 225MW data center campus in Spain

Six buildings across 25 hectares in the La Legua industrial park; on-site generation would supply power, with capacity and tenants not yet published.

Infrastructure

Nvidia in talks to invest in Firmus after neocloud scrapped its IPO

The Information reports talks, not an agreement; a several hundred million dollar ticket would be a fraction of the $2bn–$3bn private raise Bloomberg reported Firmus was seeking.

Infrastructure

BLM weighs 35 data center-related applications on western public land

Seventeen filings are for campuses and 13 for AI-supporting infrastructure, per a Center for Western Priorities tracker built on BLM land records.

Infrastructure

Aleria plans 8,640 Blackwell GPUs for 25MW US site

Aleria plans to expand the US build to 16,000 GPUs and calls the UAE Vera Rubin deployment one of the first in the region.

Infrastructure

Foresight signs term sheet for 50.01% of Polish data center developer

Warsaw is planned at 130MW on a secured 32,000 sqm plot with an option on a further 22,000 sqm.

Private Infrastructure Dailyfrom Private Wealth Daily

Energy transition, digital infrastructure and core-plus funds. Weekdays at 7:25 a.m. ET.

Free. Weekday mornings. Unsubscribe from any edition in one click.

From other sections