Connectivity glossary

The vocabulary of the Connectivity layer.

Plain definitions of the terms that appear in fibre agreements, interconnection price lists, wavelength orders and latency service levels, written for the people who finance and operate the facilities those routes serve.

Read the Thesis Contact Outpace

The terms are in four groups, from the glass in the ground to the economics of a facility that is connected well or badly. Each definition is written for owners, operators and investors who need to read a route map, a carrier quotation or a service level agreement with a critical eye. Where a term bears directly on the economics of an AI facility, the definition says how. The network fabric inside the hall belongs to the Compute layer and is not covered here.

Groups build on each other. Routes and fibre describe the physical path between a facility and the world, interconnection describes where and on what terms traffic is exchanged, transmission describes how much can be carried over a route and how fast, and economics describes what all of this costs and how the arrival of machine-driven traffic changes what a facility must contract for.

Part of the Connectivity layer. All glossaries

83 terms

Group 01

Routes and fibre

A route is a civil engineering asset with its own permits, counterparties and lead time, and it is routinely the last thing to arrive at a site chosen for power. The terms in a fibre agreement decide whether an owner holds an asset for twenty years or rents a service that can be repriced, and whether two routes that look diverse on a map share the same trench.

One route from a facility to another continent. Twenty kilometres of metro fibre to a carrier hotel, a thousand kilometres of long-haul with about a dozen amplifier sites and one regeneration site, backhaul to the landing station, and a subsea cable with a repeater every 60 to 100 kilometres.

Dark fibre

Optical fibre that is installed but carries no light until the customer attaches its own transmission equipment at each end. The buyer controls the capacity, which grows with each generation of optics at no change to the route cost, but also carries the equipment, monitoring and repair coordination a lit service bundles in. It suits an operator with sustained, growing demand on a fixed route.

Show in diagram

Dark fibre is bought at this level: the glass, with the buyer supplying the light and everything above it in the stack.

Lit service

Capacity sold by a carrier over fibre it has already equipped, delivered as a wavelength, an Ethernet circuit or an internet port. The carrier owns the optics and the repair obligation; the customer pays a recurring charge for fixed capacity. Lit services are faster to order and cheaper at low volumes, but each upgrade is a new negotiation and the price carries the carrier's margin on a route the customer does not control.

Show in diagram

The smallest unit: a circuit of stated size on a carrier's wavelength, with the carrier owning everything beneath it.

Fibre pair

Two strands of glass, one per direction of transmission, the usual unit of sale for dark fibre and subsea capacity. A cable holds from a dozen strands on a short metro link to several hundred on a modern terrestrial trunk; subsea cables carry far fewer pairs because their repeaters need power. One pair with current optics carries tens of terabits per second, so pair count says more about future headroom than present need.

Show in diagram

One pair among many in the cable. Current optics put tens of terabits per second on a single pair.

Conduit and duct

The protective pipe in the ground through which fibre cable is pulled or blown. Conduit is the expensive part of a build, carrying the civil works, permits and restoration costs, while the cable inside is cheap to add or replace. Spare duct is an option on future capacity, and rights to a duct, sub-duct or space within one are traded separately from the fibre. Two cables in one duct share every fault it suffers.

Show in diagram

The duct is the costly layer. Everything inside it can be added or replaced for a small fraction of the civil works.

Indefeasible right of useIRU

A long-term, non-cancellable right to use specified fibre strands, duct or capacity, paid largely upfront for a fixed term, commonly fifteen to twenty-five years. The grantor keeps title but cannot revoke the right while the annual operations and maintenance charge is paid. Economically it is ownership for the term, putting a route on the balance sheet without building it; term, maintenance charge and rights on a cut are the clauses that matter.

Show in diagram

The instrument column. An IRU usually sits at the fibre pair level, for a term of fifteen to twenty-five years, with title staying with the grantor.

Metro fibre

Fibre routes within a city or its region, from a few kilometres to a few tens of kilometres long, connecting facilities to carrier hotels, exchanges, cloud on-ramps and each other. Metro routes are dense and competitively priced in established data centre precincts and scarce around a site chosen for its power connection. Over these distances equipment, rather than fibre, dominates latency, so a cluster of sites can operate as one campus.

Show in diagram

The metro segment: a few kilometres to a few tens of kilometres, short enough that equipment rather than distance sets its latency.

Long-haul fibre

Fibre routes between cities or regions, from a few hundred to several thousand kilometres, built along highways, railways, pipelines and power line easements. They need regular amplifier sites, and ownership is usually concentrated in a few carriers, so price depends more on the number of independent owners in a corridor than on distance. For a remote facility, the route to the nearest exchange is often the largest connectivity cost and longest lead time.

Show in diagram

The long-haul segment. Twelve amplifier sites across a thousand kilometres, each one powered, fenced and maintained.

Subsea cable

A fibre optic cable laid on the sea floor between landing points on different coasts. It comprises the wet plant, the cable and its repeaters, and the dry plant at each landing station; capacity, owned by a consortium or single owner, is sold as fibre pairs, spectrum or wavelengths. Cables take years to permit and build and are repaired by ship, so a facility dependent on international traffic needs several systems landing in several places.

Show in diagram

The wet plant: cable and repeaters on the sea floor, owned and sold separately from the land routes at either end.

Cable landing stationCLS

The building near a coast where a subsea cable comes ashore, is powered and terminates on transmission equipment before its traffic continues over land. Its power feed equipment energises the repeaters along the whole cable, and the backhaul from it to the nearest city is a route in its own right with its own owner and price. Proximity to a landing station is valuable only if that backhaul is diverse and competitively held.

Show in diagram

The landing station. It powers every repeater on the cable and is the start of a backhaul route with its own owner.

Backhaul

The route carrying traffic from a point of entry, such as a cable landing station or a remote facility, to the nearest point where many networks meet. The word describes a role rather than a technology: backhaul may be dark fibre, wavelengths or a leased circuit. It is where connectivity economics are most often weakest, because a landing or a facility may be served by a single owner for its first stretch, who prices accordingly.

Show in diagram

The backhaul segment between the landing station and the nearest hub. This stretch most often has a single owner.

Route diversity

Two or more routes between the same points that share no physical path, so one event cannot sever both. Diversity covers the whole route, including building entries, street crossings, bridges and leased third-party ducts. The word is used loosely in quotations: a route sold as diverse may share a river crossing, rail corridor or entry vault, so a claim is only as good as the maps and survey behind it.

Show in diagram

Two routes sold as diverse. For the length of the river crossing they share one bridge, and one event there severs both.

Physical separation

The minimum distance maintained between two routes meant to be diverse, specified in metres and checked against as-built drawings rather than schematic maps. A common requirement is that the routes never come within a stated distance of each other except at the endpoints, and enter the facility through separate conduits on different sides. Separation makes a diversity claim verifiable, and a buyer should insist on seeing it drawn.

Show in diagram

The separation distance the contract states, drawn where the routes run parallel. It is checked on as-built drawings rather than on a map like this one.

Protected circuit

A circuit the carrier keeps in service by switching traffic to a second path when the primary fails, within tens of milliseconds; an unprotected circuit runs over one path and goes dark on a cut. Protection costs more because the carrier holds capacity on two paths, and is worth nothing unless the paths are physically separate. Many operators instead buy unprotected circuits on two diverse routes and switch themselves, keeping diversity under their control.

Show in diagram

The protection path. Traffic switches to it within tens of milliseconds of a cut on route A, provided the two paths are physically separate.

Route kilometres

The distance light actually travels along a fibre path between two points, as distinct from the straight-line distance. Fibre follows roads, railways and easements and detours through amplifier sites and exchanges, so route kilometres run a third to a half longer than the map suggests, and a path named for the cities it joins may loop much further. Latency is set by route kilometres, so a latency commitment needs a route drawing behind it.

Show in diagram

Route kilometres against straight-line kilometres for the same pair of points. Latency follows the longer bar.

Amplifier site

A small powered hut or cabinet on a long-haul route housing optical amplifiers, which boost the light signal without converting it to electricity. Terrestrial amplifiers sit roughly every 80 to 100 kilometres, so a thousand-kilometre route has about a dozen sites, each needing power, access, security and maintenance. Many route failures originate here, and the sites fix a corridor's economics: adding capacity on existing fibre only needs new equipment at the ends and the amplifiers.

Show in diagram

One amplifier hut of the twelve. Spacing of roughly 80 to 100 kilometres is set by how far the light can travel before it needs boosting.

Regeneration site

A location on a long route where the optical signal is converted to electrical form, cleaned up and retransmitted, once accumulated noise and distortion exceed what the receiver can correct. Coherent optics have pushed the distance a signal travels without regeneration to thousands of kilometres on good fibre, so modern routes need few of these sites. Each adds per-wavelength equipment and a little latency, and is a point where wavelengths can be handed between carriers.

Show in diagram

The single regeneration site on the route. Coherent optics let the signal run thousands of kilometres before it needs rebuilding, so modern routes have few.

Submarine repeater

A sealed, pressure-resistant optical amplifier spliced into a subsea cable roughly every 50 to 100 kilometres, powered by high-voltage direct current fed along the cable from the landing stations. Repeaters are built to last the cable's design life, typically twenty-five years, because replacing one requires a cable ship. Their power budget limits the fibre pairs a subsea cable can carry, so a transoceanic cable has far fewer than a terrestrial trunk of the same age.

Show in diagram

One repeater of many, powered over the cable from the shore. Its power budget caps the number of fibre pairs the cable can carry.

Fibre cut

A physical break in a cable that interrupts every circuit and wavelength on it at once. Most terrestrial cuts come from excavation and construction, then vehicle strikes on aerial routes, rodents, flooding and fire; subsea cuts come mainly from anchors and fishing gear in shallow water. Without diversity a cut takes a facility offline until the splice is made; cut frequency, knowable from a route's history, should inform diversity design and the service levels sold.

Show in diagram

A cut on route A. Every circuit and wavelength on the cable fails together, and repair time depends on how far the splice crew must travel.

Mean time to repairMTTR

The average time from fault detection to restored service, dominated for fibre by travel, locating the break and splicing. A cut on an accessible road is typically restored within hours, one in remote terrain or bad weather can take days, and a subsea cut waits for a repair ship and permits and can take weeks. MTTR sets how much of a facility's availability budget a single unprotected route consumes, and carriers hold the records.

Permits and easements

The legal rights to dig, hang or lay cable across land the builder does not own: road and rail crossing permits, environmental approvals, native title and heritage consents, and easements over private land. Permitting is usually the longest and least predictable part of a route's timetable, and one refused crossing can force a detour of many kilometres. Utility corridors with existing rights are valuable, and a connectivity schedule should rest on granted permits.

Trenching

Cutting an open trench, laying duct at a set depth and reinstating the surface. Trenching is the most robust and, per kilometre, the most expensive way to build a route; cost depends on surface type, depth and required restoration. Microtrenching cuts a slot a few centimetres wide in a road for a shallow duct, faster and cheaper but exposed to resurfacing and utility strikes. Burial depth is a direct input to how often a route is cut.

Directional drilling

A trenchless method that bores a guided path underground and pulls duct back through it, used to cross roads, railways, rivers and other obstacles without opening the surface. It costs more per metre than trenching but avoids an open cut's permits and traffic management, so it is how a route crosses anything it cannot dig through. Crossings are also where competing routes converge, because drillable points under a river or railway are few.

Show in diagram

The drilled river crossing. The bore points under a river are few, which is why competing routes tend to converge here.

Aerial fibre

Cable strung on power or telephone poles, or wrapped around or built into a transmission line's earth wire. Aerial construction is quick and cheap where poles exist and avoids most excavation permits, so power line corridors are a common way to reach a facility at a substation. The cost is exposure: vehicles, storms, fire and pole replacement cut aerial routes far more often than buried ones, and repair depends on the pole owner granting access.

Group 02

Interconnection

A facility earns from traffic only when that traffic can reach the networks, clouds and customers that want it, and the price of reaching them is set by where the facility sits in the interconnection market. Cross-connects and exchange ports are a recurring cost a customer pays for as long as it stays, and the same density that makes a site expensive to enter makes it costly to leave.

One facility as an interconnection market. Three carriers, one exchange switch, one cloud on-ramp and two customers, joined by cross-connects and one private interconnect through the meet-me room. A second meet-me room on a separate entry provides in-building diversity.

Internet exchangeIX

A shared switching fabric, usually run by a neutral operator or a member association, to which many networks connect so that any can exchange traffic with any other across a single port. An exchange turns many bilateral cross-connects into one port and a set of peering sessions. Exchanges concentrate in a few buildings per city, and an exchange's presence in a facility is one of the strongest reasons for networks to locate there.

Show in diagram

The exchange switch. One port on it replaces a cross-connect to each of the networks present.

Peering

An arrangement in which two networks exchange traffic between their own customers directly, without a third network carrying it. Peering shortens paths, lowers latency and removes the per-unit cost of transit for the traffic it covers. It can be done privately over a cross-connect or publicly across an exchange, settlement-free or paid, and a facility's value to a network rises with the number of worthwhile peers already in the building.

Settlement-free peering

Peering in which neither network pays the other, each bearing only its own cost of reaching the exchange point. Networks agree to it when the traffic is of roughly balanced value to both, and large networks publish peering policies setting conditions on traffic volume, ratio and geographic presence. For an operator, reaching a network settlement-free means being present where it peers, a siting decision as much as a commercial one.

Paid peering

Peering in which one network pays the other for the direct exchange of traffic, usually because the paying side sends far more than it receives or the receiving network's customers justify a charge. It sits between settlement-free peering and transit: traffic still flows directly, but at a price. Large content and cloud networks are often party to them, and an AI facility serving many end users may find the networks it needs most charge.

IP transit

A paid service in which a carrier delivers a customer's traffic to, and accepts traffic from, the whole internet, charged by committed capacity and measured usage. Transit is the default way a facility reaches everything it does not peer with, and its unit price has fallen for decades on competitive routes. Prices are usually quoted per megabit per second at the ninety-fifth percentile of measured usage, which rewards steady traffic and penalises bursts.

Autonomous systemAS

A network under one administrative control that presents a single routing policy to the rest of the internet, identified by a number assigned through the regional internet registries. Every carrier, cloud, exchange participant and large enterprise is an autonomous system, and an operator that wants to peer, multi-home across carriers or hold its own addresses needs one. Holding an AS number and address space lets a facility change carriers without renumbering its customers.

Border Gateway ProtocolBGP

The protocol by which autonomous systems tell each other which address ranges they can reach and through which neighbours. Networks exchange announcements with peers and transit providers and choose among paths by policy and path length rather than measured speed. A withdrawn route propagates across the internet in seconds to minutes, slow beside optical protection, and a facility announcing its own routes under its own AS number can shift traffic between carriers without customers noticing.

Point of presencePoP

A location where a network has equipment and offers connection to customers or peers. A carrier's list of points of presence is the list of buildings where its services can be delivered without a new build; cloud on-ramps and content caches are points of presence by another name. Whether the networks a facility needs already have one within metro fibre reach decides much of its connectivity cost and timetable.

Carrier hotel

A building, usually in a city centre, where a large number of carriers, exchanges and cloud networks have equipment and interconnect. Carrier hotels grew up around the termination points of the earliest long-haul and subsea routes and remain where the most networks can be reached across the shortest cross-connect. A facility outside one reaches those networks over metro fibre, so the route to the carrier hotel is often its most important single link.

Carrier-neutral facility

A data centre whose owner does not sell network services in competition with the carriers present and allows any carrier to enter on equal terms. Neutrality attracts networks, since none will concentrate customers where the landlord is also a rival. A facility owned or operated by a carrier, or whose fibre entries one provider controls, is priced by that provider's interests. Neutrality is structural and should be tested in the entry, meet-me room and cross-connect terms.

Show in diagram

Three carriers on equal terms, each with its own entry. Neutrality is tested at the entry, the meet-me room and the cross-connect schedule.

Meet-me roomMMR

The space in a facility where carriers terminate their fibre and customers' circuits are joined to them. As a building system it belongs to the Data Centre layer; as a market it is where interconnection is priced. Its operator sets, for a customer's whole tenancy, who may enter, cross-connect provisioning time and monthly price. Two meet-me rooms in separate parts of the building, fed by separate entries, give in-building diversity.

Show in diagram

The meet-me room. Its operator decides who may enter, how fast a cross-connect is delivered and what it costs each month.

Cross-connect

A physical cable, usually fibre, run within a facility between two customers, or a customer and a carrier, most often through the meet-me room. It is the cheapest, lowest-latency connection available in one building, and carries a one-time installation charge plus a recurring monthly charge. Recurring charges are small individually and large in aggregate, and because each ties a customer to the building, cross-connect counts measure how hard a facility is to leave.

Show in diagram

Nine cross-connects. Each is a short fibre, a one-time installation charge and a recurring charge that persists for as long as the customer stays.

Cloud on-ramp

A location where a cloud provider's network can be reached over a private connection, usually a router in a carrier-neutral facility or an exchange port. Customers in the same metro reach the cloud with predictable latency and no transit charge. Training data, model weights and the applications calling a model often live in a cloud region, so for an AI facility the distance to the nearest on-ramp is the distance to that data.

Show in diagram

The on-ramp router. Customer B reaches a cloud region over one cross-connect rather than over the public internet.

Dedicated cloud interconnect

A private circuit of fixed capacity between a customer's equipment and a cloud provider's network, ordered at a stated port speed and delivered as a cross-connect at an on-ramp or a circuit from a carrier partner. The provider typically charges for the port by the hour and for outbound data by volume, so recurring cost depends on traffic direction. They move large datasets to and from a cloud region at known cost and latency.

Private network interconnectPNI

A direct, dedicated connection between two networks, usually one or more cross-connects into each other's routers, used when their traffic justifies bypassing the shared fabric of an exchange. Private interconnects carry most of the traffic between large content, cloud and access networks. For an operator, it is the point at which a relationship with a network moves from an exchange port to a bilateral agreement with its own capacity planning and terms.

Show in diagram

The bundle between customers A and B: a private interconnect that bypasses the exchange because their mutual traffic justifies it.

Remote peering

Connecting to an exchange from a facility where it has no switch, over a carrier's circuit or wavelength, to peer in a city without placing equipment there. It widens the reach of a facility far from the exchanges, but adds the circuit's latency to every peering session and puts the circuit's availability inside the peering relationship. It is a useful bridge for a new site and a poor permanent answer for traffic sensitive to either.

Show in diagram

The exchange switch reached from another building over a carrier's circuit. The circuit's latency and availability sit inside every peering session.

Network as a serviceNaaS

Connectivity sold through a software portal or interface: a customer orders a virtual circuit between two points on the provider's fabric, chooses capacity and term, and can change or cancel it in minutes rather than months. Such fabrics commonly connect data centres, exchanges and cloud on-ramps within and between metros. The service converts fixed interconnection into a variable cost, with the trade that the customer rides the provider's routes and diversity rather than its own.

Content delivery networkCDN

Servers placed close to end users holding copies of frequently requested content so that most requests are answered nearby rather than from the origin. Content networks are why most consumer internet traffic travels a short distance, and why facilities near large populations can serve it with little long-haul capacity. The arrangement depends on the same content being requested many times, which inference traffic, where each response is computed fresh, does not share.

Edge cache

A server, or a rack of them, inside an access network or a facility holding popular content for the users behind it, often supplied free by the content provider. It trades a little space and power for a large cut in the capacity the host must buy upstream. Caches suit the heavy downstream, light upstream pattern the internet was built for, and a facility whose traffic is machine-to-machine should expect little from them.

Show in diagram

The share of human traffic a cache serves locally, which is the capacity the host does not have to buy upstream.

Interconnection density

The number of distinct networks, exchanges, cloud on-ramps and carriers reachable by cross-connect within a facility or a short metro link of it. Density lets a facility command higher recurring charges for the same space and power, because each additional network makes the building more valuable to every other. It accumulates slowly and is hard to replicate, so new sites chosen for power usually buy their way to density over metro and long-haul routes.

Show in diagram

Every line is a reason for its two endpoints to stay. Density compounds as each new network makes the building more valuable to the rest.

Group 03

Transmission and capacity

The same fibre carries a hundred times more traffic with each generation of optics, so the cost of a route is in the glass and the civil works while the capacity on it is nearly free to add. Latency is set by distance and physics, which no later spending can fix, and the figures a carrier quotes for capacity, latency and availability each hide something that a buyer should ask about.

Distance against latency at about 5 microseconds per kilometre one way. Two hundred kilometres is a millisecond each way; a round trip of a millisecond covers roughly 100 kilometres of route. The bands show where training, inference and agent chains can sit.

Wavelength service

A fixed amount of capacity, such as 100 or 400 gigabits per second, delivered as one colour of light on a carrier's dense wavelength division multiplexing system. Dedicated to the customer and carrying any protocol, it behaves like a private fibre of fixed size. It is the normal unit for long-haul capacity when dark fibre is unavailable or excessive, and its price on a route is a direct read of that route's competitiveness.

Show in diagram

One wavelength bought as a service, at a fixed rate such as 100 or 400 gigabits per second.

Dense wavelength division multiplexingDWDM

Sending many independent signals down one fibre at once, each on a slightly different wavelength of light, and separating them again at the far end. A single fibre carries dozens of channels across the standard transmission bands, each carrying hundreds of gigabits per second with modern optics. DWDM is why a fibre pair laid decades ago still has headroom, and why a route's economic life exceeds that of any equipment attached to it.

Show in diagram

Each tick is one wavelength on the same pair. The channel count depends on spacing and band; dozens per fibre is routine.

Coherent optics

Transmission technology that encodes data in the phase and amplitude of light as well as its intensity, with a receiver measuring the full optical field and signal processing correcting the fibre's distortions. Coherent optics raised single-wavelength capacity from ten gigabits per second to several hundred and let signals travel thousands of kilometres without regeneration. They now arrive as pluggable modules for routers and switches, removing a layer of transport equipment on shorter routes.

Line rate

The nominal speed of a port or wavelength, quoted in gigabits per second and standardised by the IEEE for Ethernet at steps including 100, 400 and 800 gigabits per second. A higher line rate lowers the cost per gigabit of equipment and ports, so operators move to the highest rate they can fill. Line rate is a ceiling; useful capacity below it depends on protocol overheads, error correction and how traffic spreads across parallel paths.

Spectrum sharing

Selling a portion of the optical spectrum on a fibre, measured in gigahertz, rather than a whole pair or a fixed wavelength. The buyer attaches its own transponders and uses it however current optics allow, so capacity grows with technology as on dark fibre, while the seller keeps the line system and sells the rest of the band. Spectrum is now common on subsea systems and long-haul routes where a pair exceeds one customer's needs.

Show in diagram

A slice of the band on one pair, sold in gigahertz. The buyer brings its own transponders and the seller keeps the rest of the band.

Open line systemOLS

An optical line system whose amplifiers, multiplexers and monitoring accept transponders from any vendor, so the route owner and the capacity owner can choose equipment independently. Closed systems tied transponders to the line vendor and let that vendor price upgrades. Open systems make spectrum sharing and tenant-owned optics practical, and underpin a route owner selling the same fibre to many customers at different layers.

Show in diagram

On an open line system any vendor's transponders can light a slice of this band, which is what makes selling the stack at several levels practical.

Optical transport networkOTN

The ITU-standardised framing that wraps client signals, such as Ethernet, into fixed containers for carriage over wavelengths, with its own error correction and performance monitoring. Optical transport runs beneath the IP layer, moving bits between two points without inspecting them, while routers decide where each packet goes. The optical layer costs less per bit and adds less latency than routing, so operators push as much traffic as possible down to wavelengths.

Bandwidth

The maximum rate at which a link can carry data, quoted in bits per second; in commercial use, the capacity a customer has bought. Bandwidth is a ceiling set by the port, wavelength or committed rate of a service, and a quotation giving it alone says nothing about what the link delivers under load. The figure to pair with it is throughput, what traffic actually achieves after loss, latency and protocol behaviour take their share.

Throughput

The rate at which useful data actually crosses a link, measured rather than quoted, and always below the bandwidth. Throughput falls with packet loss, with latency on protocols that wait for acknowledgement, and with contention where flows share a port. A long, high-bandwidth circuit can deliver low throughput to a single flow for reasons unrelated to the carrier, so bulk transfer between distant sites uses many parallel flows and tuned protocols.

Latency

The time a signal takes to travel from one point to another, measured one way in milliseconds or microseconds. On a fibre route it is propagation time, which depends only on distance, plus small delays from each piece of equipment. Propagation dominates on long routes, so latency is fixed by route kilometres and cannot be bought down. A quote should state whether it is one-way or round-trip, typical or guaranteed, and over which route.

Show in diagram

The fixed block at the left is equipment delay. On a metro route it dominates; past a few hundred kilometres distance takes over.

Jitter

The variation in latency from one packet to the next across the same path, caused by queuing in routers and switches, competing traffic and path changes. A link can have low average latency and high jitter, and jitter is what disrupts synchronised work between machines, because each step waits for the slowest message. It is controlled by keeping links lightly loaded, carrying sensitive traffic at the optical layer and avoiding paths that reroute often.

Show in diagram

Bursts queuing behind each other at a shared port. The variation in arrival time is jitter, and synchronised work waits for the slowest message.

Round-trip timeRTT

The time for a signal to reach a destination and the reply to return, twice the one-way latency on a symmetric path. Round-trip time is what most protocols and applications experience, because a sender waits for an acknowledgement before continuing, and it limits the throughput of a single flow on a long link. Carrier quotations sometimes give one-way latency where a buyer is thinking in round trips, which halves the apparent distance of a route.

Show in diagram

The lower scale: twice the one-way figure on a symmetric path, and the figure most protocols actually experience.

Speed of light in fibre

Light travels through optical fibre at about two thirds of its vacuum speed, close to 200,000 kilometres per second, slowed by silica's refractive index. The practical rule is roughly 5 microseconds per kilometre one way: 200 kilometres of route adds about a millisecond each way, and a millisecond round trip covers roughly 100 kilometres. These figures bound every latency commitment; a quote below what distance allows means a shorter route than advertised or a wrong figure.

Show in diagram

The propagation scale: about 200,000 kilometres per second in glass, roughly 5 microseconds per kilometre. No later spending moves this line.

Latency budget

The total delay a workload can tolerate between two points, allocated across fibre distance, equipment and software. Synchronous training across sites needs round trips short enough that exchanging gradients does not idle the accelerators, keeping cooperating sites within a metro or a short regional distance. Inference for a human tolerates tens of milliseconds, but an agent making many sequential calls spends the budget many times over, so distance matters more as calls are chained.

Show in diagram

An agent chain spends the inference budget once per call. Five sequential calls over 1,000 km cost 50 milliseconds in round trips before any computation.

Data centre interconnectDCI

High-capacity optical links between two or more facilities, usually in the same metro, run as if the facilities were one. Interconnect is built on dark fibre or many wavelengths, sized to the traffic between the halls rather than the traffic leaving the campus. For AI operators it lets a cluster span buildings when one facility cannot supply the power, and its latency and capacity then count as cluster performance rather than an external service.

Show in diagram

Interconnect between two halls 20 kilometres apart adds about 0.1 milliseconds each way, close enough to run the halls as one cluster.

Oversubscription

Selling more capacity in total than a link or port can carry at once, on the assumption that customers will not all use it at the same time. Oversubscription is normal and sensible for bursty, uncorrelated traffic, and is what makes shared services cheaper than dedicated ones. It fails when demand becomes continuous and correlated, the shape of machine-to-machine traffic, and a buyer of any shared service should know the ratio applied to it.

Show in diagram

The sold capacity line. Human curves leave headroom beneath it to share; machine traffic sits against it, so the sharing assumption fails.

Packet loss

The share of packets sent across a path that never arrive, caused by congested queues, faulty optics or dirty connectors, and errors the link's correction cannot recover. Even a small loss rate cuts the throughput of protocols that slow when they detect it, and long links amplify the effect because recovery takes a full round trip. Loss is usually bounded in the service level agreement, and is the first check when a link underdelivers.

Encryption in flight

Encrypting data while it travels over a route so that a tap on the fibre or an intermediate carrier cannot read it. It can be applied at the optical, Ethernet or IP layer, with lower layers adding less latency and overhead at line rate. Model weights and training data are among the most valuable assets an operator moves, so line-rate encryption is now expected on inter-site links and any route leaving the operator's control.

Software-defined wide area networkSD-WAN

A method of running a wide area network in which a central controller steers traffic across several underlying links, such as a private circuit and two internet services, by policy and measured performance. It improves resilience and cost for branch and enterprise traffic. For a facility's own backbone it matters less than the quality of the underlying routes, since software can choose among paths but cannot create diversity that is not in the ground.

Segment routing

Steering packets through a network by stamping each at its entry with the list of segments it should traverse, so intermediate routers forward it without per-flow state. It lets an operator pin traffic to a chosen route, for example to keep it within a jurisdiction or on the lowest-latency path, and pre-compute a backup path that takes over within milliseconds of failure. Segment routing is how a contracted route becomes an enforced one.

Group 04

Economics and the agent era

Connectivity has been priced for human traffic: bursty, asymmetric, cacheable and forgiving of a dropped route for a few seconds. Machine-driven traffic is continuous, symmetric and intolerant of failure, and distributed training makes an inter-site link part of the compute. Routes, latency and diversity become attributes of the capacity a facility sells, which means they must be contracted, measured and underwritten rather than assumed.

A day of link load for human traffic and for machine traffic. The human curve is diurnal, heavier downstream and largely served from cache. The machine lines are flat, symmetric and uncached, with millisecond bursts to the line rate.

Monthly recurring chargeMRC

The fixed amount paid each month for a circuit, wavelength, port or cross-connect for the life of its term. Connectivity budgets are built on it, and across a facility the sum of small recurring charges for cross-connects and ports often exceeds the cost of the long-haul routes. Recurring charges on lit services are reviewed at renewal, so a route whose market price has fallen can be renegotiated while a route with one owner cannot.

Non-recurring chargeNRC

A one-time fee for installing a service, covering the construction, splicing, equipment and provisioning done before the recurring charge begins. Carriers often trade non-recurring against recurring charges and term, waiving installation for a longer commitment or charging it in full for a short one. Where a route needs new construction to reach a facility, it can be the cost of that build, and ownership of what is built should be settled in the same document.

IRU term and accounting

The treatment of an IRU in a facility's accounts and financing. The upfront payment is usually capitalised and amortised over the term, the annual operations and maintenance charge is expensed, and under current lease standards an IRU over identified fibre is commonly a right-of-use asset with a matching liability. A fifteen to twenty-five year term outlasts several generations of optics and most customer contracts, so the route is rarely what limits a facility's life.

Price erosion

The steady fall in the price of a unit of capacity on a route, driven by each generation of optics multiplying what a fibre carries and by new entrants. Erosion is strongest between major hubs with many owners and weakest where one or two owners control the glass. For a buyer, erosion argues for short terms on lit services over competitive routes and for ownership or long rights where it will not come.

Constrained route

A corridor where fibre is owned by one or a few parties, or is nearly full, so capacity is priced by scarcity rather than cost. They are common on the last stretch to a remote facility, landing station backhaul and international paths where cables are few. The cure is to build, take an IRU on a competing route or anchor a new build; a facility sited for cheap power should expect the route to be the constraint.

Cross-connect pricing

The one-time and recurring charges a facility levies for each physical connection through its meet-me room. Cross-connects are cheap to install and almost never removed, so their recurring charges are among the highest-margin revenue a facility earns, and for a tenant with many carriers and partners they are a material share of the colocation bill. A customer comparing facilities should compare the cross-connect schedule as closely as the rate for space and power.

Show in diagram

The recurring charge per connection, levied at the meet-me room. Small on each line and material in aggregate.

Carrier neutrality

Allowing any carrier into a facility on equal terms and charging each the same for a cross-connect. As a pricing lever it works for the customer inside the building, where carriers compete across the meet-me room, and for the facility at the gate, where a neutral site attracts more carriers and charges more for interconnection. A facility that grants one carrier an exclusive entry has sold that lever for a single payment.

Contracted route diversity

Route diversity written into a customer's capacity agreement, with the separation, entry points and carriers named and a remedy if the facility fails to maintain them. Diversity merely described in a sales document is lost the first time a carrier regrooms a route or a duct is shared. An operator selling diversity must hold matching back-to-back rights with its route providers, which turns a connectivity attribute into a contracted asset of the facility.

Show in diagram

Route A as named in the customer's agreement, with its separation, entry and carrier stated. Diversity that is only described in a brochure is lost the first time the carrier regrooms it.

Latency service level

A contractual commitment that latency between two stated points will not exceed a figure, usually a monthly average or a percentile, with service credits if it does. The figure is only meaningful with the route it is measured over, because a carrier may reroute protected traffic onto a longer path and still meet an average. What a facility can commit to customers running synchronised or chained workloads is bounded by the commitments it holds.

Availability service level

A contractual commitment that a service operates for a stated share of time, such as 99.99 percent in a month. The figure hides outage distribution: a route cut once a year for a day fails a 99.9 percent target, while a protected circuit switching in fifty milliseconds many times a month meets 99.999 percent yet may disrupt synchronised work on every switch. Credits for shortfalls rarely match the customer's loss, so the figure is a design input rather than a remedy.

Machine-to-machine traffic

Traffic between software with no person at either end: models calling models, agents calling tools, sites exchanging state, accelerators exchanging gradients. Human traffic follows a working day, flows mostly downstream and is read many times from caches; machine traffic runs around the clock until the work is done, flows both ways and is produced fresh on every request. Connectivity was engineered and priced for the first pattern; the second removes the assumptions that made it cheap.

Show in diagram

Machine traffic across a day: continuous, with no overnight trough to recover in.

Agent traffic profile

Network demand from software agents that plan, call tools and models, and act on the results. Each step is short and waits on the last, so an agent is latency-bound at the millisecond scale, bursty within a task and continuous across a fleet. Every call is a fresh inference that caches do not absorb, so routes serving agents see sustained load with sharp peaks, and a failed route stops work rather than slowing it.

Show in diagram

One second of agent traffic. Short bursts to the line rate, each waiting on the last, and a fleet of them filling the day.

Symmetric traffic

Traffic whose volume is similar in both directions, as distinct from the heavily downstream pattern of people reading and watching. Machines exchanging state, models exchanging context and sites replicating data send as much as they receive. Access networks, peering ratios and many pricing models assume asymmetry, so a facility whose outbound matches its inbound finds settlement-free peering policies, cloud data transfer charges and transit billing treat it less favourably than a consumer-facing network.

Show in diagram

Upstream equal to downstream. Peering policies, cloud transfer charges and transit billing assume otherwise.

Route failure tolerance

How far a workload continues through the loss of a route: high for human traffic, close to nil for synchronised machine work. A person reloads a page; a training step that misses a gradient exchange stalls every accelerator, and an agent whose tool call fails abandons or repeats its task. Tolerance is engineered in the network: diverse routes, protection or fast reroute within tens of milliseconds, and capacity on the second path so failover does not congest.

Cache-defeating traffic

Traffic whose every response is new and never requested again, so storing copies near the user achieves nothing. Inference is the clearest case: each response is computed from the request, and even identical prompts are rarely identical once context is included. Caching has held down long-haul capacity needs for two decades; uncacheable traffic is carried end to end every time, raising the capacity a facility needs on its routes rather than at its edge.

Show in diagram

The hatched region is what caches absorb for human traffic. The machine chart has no such region; every response is computed fresh and carried end to end.

Inter-site training link

A dedicated optical link between facilities carrying a single training job's gradients, activations or parameters, so two sites behave as one cluster. Its capacity must match the accelerators' exchange rate, and its latency and jitter set how long each synchronisation step idles them. The fibre is part of the compute: a slower or less reliable route lowers output on both sides, so a multi-site campus turns on owning or long rights over the glass between its buildings.

Show in diagram

The synchronous training band. Sites that exchange gradients every step stay within tens of kilometres of each other, inside a metro.

Sovereign routing

Keeping a data flow's physical path within one jurisdiction, so one country's law applies throughout and no foreign cable, landing station or exchange is transited. Internet routing follows policy and cost, so a flow between domestic points can leave the country unintentionally. It needs domestic routes and exchange points, plus segment routing or dedicated wavelengths to pin traffic to them; for capacity sold as sovereign, the route is part of the product.

Submarine cable geopolitics

The influence of states on where subsea cables land, who may own or supply them and which waters they may cross. Governments review ownership and equipment on security grounds, deny or delay landing permits and contest routes through disputed or hostile waters, while cables in shallow, narrow seas face deliberate damage as well as accidents. International routes therefore have political as well as physical diversity, and the newest or shortest cable is not always the most reliable.

Cable permitting

The approvals a subsea cable needs to be surveyed, laid, landed and repaired: landing licences, environmental and seabed permits, consents from each coastal state it crosses, and permits for repair ships when it breaks. Permitting takes years for a new system and can take weeks for a single repair, so a cut cable stays dark longer. Landing a cable at a new site is a regulatory programme before it is a construction one.

Time to fibre

The interval between committing to a site and having diverse, usable routes in service there, measured in months and often years for a greenfield location. It is driven by permitting, construction season, the lead time of the nearest competitive route and carriers' willingness to build to a site without contracted demand. A facility energised before it is connected earns nothing, so the fibre schedule belongs on the critical path with power and the building.

Where this leads

What Outpace does with this vocabulary.

The terms of each layer describe the gap between what an asset could do and what it does. Outpace underwrites that gap. The thesis sets out why contracted, productive output matters more than installed capacity.

Read the Thesis

Start with the project you have.

Outpace works with project owners, infrastructure partners, capital partners and enterprise customers. The first conversation takes thirty minutes and ends with a straight answer on fit.

Discuss a Project Contact Outpace