Who Owns the Internet? One Request, Followed Hop by HopGo back

Who Owns the Internet? One Request, Followed Hop by Hop

No company or government owns the internet. It is about 78,000 separately run networks that have agreed to carry each other's traffic, and each one owns its own routers, its own contracts and sometimes its own fibre. We run one of those networks, AS55285. To show where the ownership sits, we traced a single request from our head office in Ottawa to a server we run in Amsterdam and named the owner of every piece it crossed, along with who pays whom.

Loading...

How the internet works: networks joined by contracts

The internet is a set of separately run networks that have agreed to carry traffic for each other. Each one is an autonomous system (AS): a network with its own number and its own routing policy, which tells its neighbours over the Border Gateway Protocol (BGP) which addresses it can deliver to. Geoff Huston's annual review of the routing table counted about 77,900 of them in IPv4 at the end of 2025.

An AS can be a national phone company, a university, a bank, Google or us. None has authority over another.

What holds the system together is contracts between pairs of networks. A customer pays a provider to reach everyone else, which is transit. Two networks swap their own customers' traffic without charging each other, which is peering. And companies rent each other the physical parts a network needs, from rack space in a building to a strand of glass under a street. A short list of registries hands out the numbers and names, and the owners of everything else are the ones who appear, hop by hop, in a single connection.

One request from Ottawa to Amsterdam

On 19 October 2026, we ran a traceroute from our head office in Ottawa to the server that runs our looking glass at Digital Realty AMS3 in Amsterdam, lg.drt-ams3.serverside.com. Each line is a router that answered along the way. The name and the AS number say who owns it, and the time is the round trip in milliseconds from the office to that router and back.

Hop  Host                                   Address          ASN        ms
  1  -                                      192.0.2.1        -          <1
  2  -                                      192.0.2.1        -          <1
  3  -                                      192.0.2.1        AS577       2
  4  *                                      *                            *
  5  *                                      *                            *
  6  -                                      192.0.2.1        -          13
  7  *                                      *                            *
  8  -                                      142.124.127.122  AS577      15
  9  *                                      *                            *
 10  -                                      64.230.79.89     AS577      12
 11  ae4-43-xcr2.nyk.cw.net                 195.2.17.49      AS1273     12
 12  ae3-xcr1.slo.cw.net                    195.2.28.170     AS1273     88
 13  be42-scr1.amq.cw.net                   195.2.3.117      AS1273     88
 14  be2-scr1.adr.cw.net                    195.2.23.214     AS1273     87
 15  -                                      217.161.90.225   AS1273    100
 16  1.edge.core.drt-ams3.serverside.com    198.73.56.1      AS55285    85
 17  1.spine.core.drt-ams3.serverside.com   198.73.56.47     AS55285    85
 18  access.r01b03.drt-ams3.serverside.com  198.73.56.241    AS55285    84
 19  hosted-by.serverside.com               198.73.56.243    AS55285    85

Our office router, Bell's modem, Bell's first router on the line and one private address inside Bell's network are replaced with a documentation address. Every other line is as captured. Hop 15 answers more slowly than the hops after it because routers reply to traceroute probes at low priority; the traffic they forward is not held up.

HopsNetworkWho owns itWho pays whom
1Office routerServerside.comBought once
2 to 10Bell Canada, AS577Bell CanadaWe pay Bell every month for the office line
11 to 15Vodafone, AS1273Vodafone Group PLCBell and Vodafone: transit or peering (terms private)
16 to 18AS55285Serverside.comOur transit providers and peers carry traffic to us; we pay the providers
19Our looking-glass server at AMS3Serverside.comNobody pays us for it: we run it for the public

Every change of AS number in that capture is a contract between two companies. A traceroute cannot tell transit from peering, because the terms are private, but it does show that no hop belongs to an organisation in charge of the others.

BGP data from the same day names a network the trace missed. In RIPE's route collectors, Bell's route to us read 577 1273 24785 55285: Vodafone handed the traffic to Joint Transit (AS24785), one of our upstreams, and no router in the trace answered from a Joint Transit address.

The same request as a chain of contracts. Bell's terms with Vodafone are private, so that link carries no arrow.

The reply does not have to come back the same way. Each network picks the exit for its outgoing traffic according to its own contracts, so the route from Amsterdam back to Ottawa is chosen by us and the networks after us, not by the ISP that chose the route out. Our looking glass runs traceroutes from AMS3 and our other metros, so you can follow the return path to your own network.

The access network you pay for

A monthly internet bill buys one thing: a connection to the ISP's network. In our capture that is Bell Canada (AS577), which owns the equipment in hops 2 to 10 and, usually, the cable or fibre into the street. Past its own edge, the ISP either pays another network to carry the traffic or trades traffic with it. Our office's Bell line has no contract with anyone further along the path, our own network included: the request reached AS55285 only because Vodafone and Joint Transit had agreed to carry it.

Transit and the Tier 1 networks

Transit is the service of reaching the whole internet through someone else's network. The customer announces its addresses to the provider over BGP, and the provider announces back routes to every other network it can reach, which is all of them. We pay our Tier 1 providers on the 95th percentile: the month's traffic is measured in short samples, usually five minutes each, the busiest 5% of samples are discarded, and the highest one left sets the bill. We sell transit ourselves, through our network services, and we buy it.

AS55285 buys transit directly from Tier 1 networks such as Arelion (AS1299) and Cogent (AS174), and through NL-ix's transit service, which runs on two networks registered to the Dutch company Broadband Hosting: Joint Transit (AS24785) for full transit and Open Peering (AS20562) for peering routes. Joint Transit is a transit customer in turn. NL-ix's BGP communities page names its upstreams as Orange (AS5511), Liberty Global (AS6830), Vodafone (AS1273), Arelion and Telxius (AS12956), so our traffic reaches those networks through a provider that pays them, with no contract of our own. Buying from more than one provider means a fault or a dispute at one of them leaves another path open.

At the top of the chain are networks that buy transit from nobody. A Tier 1 network reaches every other network through its own customers and through settlement-free peering with the other Tier 1s. Lumen (AS3356), NTT (AS2914), Arelion and Cogent are four of the networks usually listed. No regulator certifies the title: it is a claim about contracts the networks keep private, which is why published lists of Tier 1s disagree at the edges.

When two large networks fail to agree, the result is public. Hurricane Electric (AS6939) and Cogent have not exchanged IPv6 routes directly, and operators have discussed the gap on the NANOG mailing list since at least October 2009. In 2012, Andrew Ayer described connections to dual-homed hosts on Cogent's network that took so long to establish he blocked the affected IPv6 routes on his own machine. The gap was still there on 19 October 2026: Hurricane Electric's looking glass had no route to Cogent's IPv6 block, 2001:550::/32, from its routers in Fremont, New York or Amsterdam. For anyone connected to only one of the two, part of the IPv6 internet is not there. Nobody can order them to connect, which is the practical meaning of "nobody owns it".

Peering and internet exchanges

Peering is the other contract. Two networks exchange traffic between their own customers and charge each other nothing for it, because each one saves the transit it would otherwise pay for. A peer delivers only to its own users; it does not resell its transit to the network it peers with.

Two networks can peer over a private cable between their routers, but much of it happens at an internet exchange point (IXP), a switching platform inside a data centre. Each member connects a port, and over that one port it can peer with any other member. Most exchanges also run route servers, so a member can peer with everyone who uses them through a single BGP session instead of hundreds. Any fee is for the port, not for the traffic that crosses it.

We connect to six exchanges, all in Amsterdam: NL-ix, Frys-IX, 1-NL, LSIX, Speed-IX and INTERIX. Our policy on PeeringDB is open, meaning we peer with any network that asks, and bgp.tools listed 944 networks connected to AS55285 on 19 October 2026: 932 peers, plus our upstreams and customers.

Our Arista 7280CR3 switch at Nikhef in Amsterdam, above a fibre patch panel. Cross-connects from the building land on the back of the panel, and yellow single-mode patch cables carry them up to the switch.

Buildings, cross-connects and dark fibre

We own our servers, switches and routers. We do not own the buildings they sit in. AMS3 belongs to Digital Realty, FR5 in Frankfurt to Equinix and DFW1 in Dallas to DataBank; our London servers are in Telehouse West, and our Chicago servers are in Netrality's building at 717 South Wells. We rent cabinets from those operators and pay them for the electricity the cabinets draw.

Inside a building, two networks meet through a cross-connect: a fibre the building operator runs from one customer's cabinet to another's, usually through a shared meet-me room, and bills for every month it stays in place. Every exchange port and every private peering session starts with one.

The top of our cabinet at Digital Realty AMS3, where the Ottawa traceroute ended: a fibre panel with single-mode patches above the first of our Arista switches.

Between buildings, the fibre in the ground belongs to someone as well. In Chicago, our servers at 717 South Wells reach Digital Realty's 350 East Cermak, the city's main carrier hotel, over dark fibre from Zayo: strands we lease unlit and light with our own optics. Most of our Chicago connectivity is picked up in that building's meet-me room.

Our Chicago cabinet at 717 South Wells. Arista 7280 switches sit under an AFL fibre panel, and the passive DWDM multiplexer at the bottom puts several wavelengths onto the dark fibre to 350 East Cermak.

Undersea cables

The path from Ottawa to Amsterdam crosses the Atlantic, so one stretch of it runs through a submarine cable. TeleGeography tracks more than 600 active and planned submarine cables in 2026, with over 1.5 million kilometres in service as of early 2026. Carriers built most of them in consortia, splitting the cost and the capacity, and content providers such as Google, Meta, Microsoft and Amazon are now major investors in new ones. TeleGeography also reports that private network operators have recently deployed more capacity than internet backbone operators.

A traceroute does not name the cable. It shows a jump in round-trip time between a router on one side of the ocean and a router on the other: in our capture, from 12 ms at Vodafone's ae4-43-xcr2.nyk.cw.net to 88 ms at ae3-xcr1.slo.cw.net, the next router along. Vodafone (AS1273) carried the packet across, on fibre it owns on the cable or capacity it leases. We do not pay Vodafone. We pay Joint Transit, and Joint Transit buys transit from Vodafone.

Numbers and names: coordinated, not owned

Two things every network needs come from registries rather than from a supplier: its IP addresses and AS number, and the domain names its users type.

Addresses start at the Internet Assigned Numbers Authority (IANA), which ICANN runs through an affiliate, Public Technical Identifiers. IANA hands large blocks to five regional internet registries: AFRINIC, APNIC, ARIN, LACNIC and the RIPE NCC. IANA worked under a contract with the US Department of Commerce until 1 October 2016, when the contract expired and oversight passed to ICANN's community.

AS55285 is registered with ARIN, the registry for Canada, the United States and parts of the Caribbean. Registration is a membership with an annual fee set by how much address space, or how many AS numbers, an organisation holds: ARIN's 2026 fee schedule runs from $275 a year for an IPv4 /24 or less to $282,240 for the largest holders. In Europe, the RIPE NCC charges €1,800 per LIR account in 2026, plus €50 for each AS number (RIPE-848). The registry record is what other networks check before they accept our route announcements.

Domain names have the same structure. The DNS root zone, the list of every top-level domain, is published by IANA and served by 13 root server identities, lettered A to M and run by 12 independent organisations (root-servers.org). Each identity answers from many sites that share one address. Verisign runs two of them, A and J, and operates the .com registry under its contract with ICANN, so a registrar selling a .com is selling a record in Verisign's database.

Who can switch it off

Nobody can switch off the whole internet, because no single network carries everyone else's traffic.

A government can switch off the access networks inside its borders by ordering the ISPs there to stop, and many do: Access Now and the #KeepItOn coalition documented at least 313 shutdowns in 52 countries in 2025. Traffic between other countries keeps flowing around them.

A network can also remove itself. On 4 October 2021, a command meant to assess Facebook's backbone capacity "unintentionally took down all the connections in our backbone network", in Meta's own account. Its DNS servers were built to withdraw their BGP announcements if they lost contact with Facebook's data centres, so they did, and every other network lost the routes to Facebook's name servers. Those other networks kept running.

What an independent network pays for

Seen from our side, those contracts come down to a list of suppliers and bills. None of them goes to anyone in charge of the internet.

What we pay forWho we payHow it is billed
TransitArelion, Cogent and Joint Transit95th percentile of measured traffic, for our Tier 1 providers
Exchange portsThe six Amsterdam exchangesA port fee, where the exchange charges one
Cross-connectsEach building's operatorMonthly, per cable
Cabinets and powerEach building's operatorMonthly
Dark fibre in ChicagoZayoLease of unlit strands
AS number and addressesARINAnnual membership fee
Routers, switches and opticsHardware vendors, Arista among themBought outright

The registries come closest to being a centre, and they are funded by fees from members like us. Every other network on the internet runs on the same arrangement: it owns its equipment, rents what it cannot build, and signs its own contracts with its neighbours.

Looking up AS55285 yourself

PeeringDB lists our exchanges, facilities and peering policy, bgp.tools lists our upstreams and peers, and the looking glass runs traceroutes and BGP route lookups from our own routers in each metro, so you can trace a path to your network from ours. The network page lists every metro and facility. If you run a network and want to peer, our policy is open; if you want transit from us, network services has the details.

Clay Berndt

Written by

Clay Berndt

CEO, Serverside.com & Host Havoc

Clay is the CEO of Serverside.com and Host Havoc, with more than a decade of experience running globally distributed hosting infrastructure and a game-server platform that has served over 200,000 customers.

Keep reading

View all articles