You tap a link on your phone in a Tokyo subway. Milliseconds later, bytes from a server in Frankfurt are painting pixels on your screen. Nobody planned that route in advance. No central switchboard looked up "Tokyo phone → Frankfurt server" and connected the call. There is no such switchboard — there is no central anything. And yet it works, billions of times a second, across roughly seventy thousand independently run networks that mostly do not know or trust one another.
That is the puzzle this page unpicks. The word "Internet" is a contraction of inter-network: it is not one big network but a network of networks, glued together by a shared addressing scheme and a shared willingness to forward each other's traffic. Understanding its shape — who plugs into whom, who pays whom, and who (if anyone) is in charge — is the difference between treating the Internet as magic and treating it as an engineered, economic, political system you can reason about.
Start at the outside. Everything you actually use the Internet to reach — laptops, phones, game consoles, web servers, sensors, thermostats — sits at the edge and is called an end system or host. Hosts run the applications; they are the source and the destination of every byte. Everything between hosts — routers, switches, links — is the network core, whose only job is to shuttle packets from one edge to another.
A host does not touch the core directly. It first joins an access network — the "last mile" that connects a home, a phone, or an office into the Internet:
The access network hands you off to your Internet Service Provider (ISP). And here is the first key idea: your ISP is itself just another network that has to reach all the other networks. So who does it plug into?
If every one of the ~70,000 networks had to run a physical cable to every other one, you would need billions of cables — hopelessly unscalable. Instead the networks arrange themselves into a loose hierarchy, so that a small number of well-connected networks can relay traffic for everyone else. It is traditionally sketched as three tiers.
Treat "three tiers" as a teaching sketch, not gospel. The real graph is messier — content giants and big clouds (Google, Meta, Amazon, Cloudflare, Netflix) build their own global backbones and peer widely, sitting outside the neat pyramid. But the core intuition holds: reach is bought going up, and shared going sideways.
Two networks can relate in exactly two commercial ways, and the distinction runs the whole economy of the Internet.
Why peer? Suppose two big regional ISPs each send a lot of traffic to the other's users. Routing it up through a tier-1 costs both of them transit fees and adds latency. If they instead run a direct cable between themselves, the traffic between their customers flows for free and takes a shorter path. Peering is thus a pure win when the traffic is roughly balanced — which is exactly why the map of who-peers-with-whom is a constant negotiation over money and fairness, not just engineering.
Your phone has never exchanged a single byte with that Frankfurt server, and neither has your access ISP. So how does the packet get there? No single network knows the whole route. Instead, each network only needs to know one thing: "for a destination I can't reach myself, who do I hand it to that's closer?" Your access ISP hands the packet up to its regional ISP; the regional ISP hands it to a tier-1 (or peers with a network that can reach Frankfurt); the tier-1 carries it across the globe and down the other side.
The glue that makes this work is a routing protocol called
We keep saying "a network," but what counts as one? The technical unit is the Autonomous System (AS): a collection of IP address ranges and routers under a single administrative control, with a single, consistent routing policy. Your ISP is an AS. Google is an AS (several, in fact). A large university is often an AS.
Each AS is identified by a globally unique Autonomous System Number (ASN) — for example, AS15169 is Google, AS32934 is Meta. ASNs are 32-bit numbers, handed out by the regional registries (below). When two networks peer or buy transit, what they are really doing is telling BGP: "AS X will accept and forward routes to and from AS Y." The Internet's routing table is, at heart, a graph whose nodes are ASes and whose edges are these business relationships.
| Concept | What it is | Example |
|---|---|---|
| Host / end system | A device that runs applications at the edge | Your phone, a web server |
| Access network | The "last mile" linking a host to its ISP | Fibre, cable, 5G |
| Autonomous System (AS) | A network under one administrative control | An ISP, Google, a university |
| ASN | The globally unique number naming an AS | AS15169 (Google) |
| IXP | A shared facility where many ASes peer at once | DE-CIX (Frankfurt), LINX (London) |
Peering one-to-one is fine for two networks, but a big city might have a hundred networks that all want to peer with each other. Running a hundred-choose-two = ~5,000 separate cables would be absurd. The solution is an Internet Exchange Point (IXP): a neutral, shared facility — often literally a room, or a set of rooms across a city, full of switches — where any member can, over a single connection to the IXP's switching fabric, peer with any other member.
IXPs are where an enormous fraction of the world's traffic changes hands. DE-CIX in Frankfurt, AMS-IX in Amsterdam, and LINX in London each move terabits per second at peak. They are among the most important pieces of infrastructure you have never heard of — and they are run as neutral cooperatives or nonprofits precisely so that no single member controls the meeting place. An IXP is the physical embodiment of "network of networks": a place built for otherwise-independent networks to shake hands.
All of this rides on real, buried, submerged, blinking hardware. It helps to picture the layers of glass and copper the abstraction sits on:
A single fibre carries dozens of wavelengths of light at once, each an independent channel — which is how one cable moves terabits. And because it is physical, it is vulnerable: a dropped anchor, a fishing trawler, or an earthquake can sever a cable and reroute a continent's traffic in an instant. The Internet's resilience comes from having many paths, not from any one link being safe.
If the Internet is 70,000 independent networks with no centre, who makes the rules? The answer is a set of coordinating bodies that manage the shared resources everyone must agree on — but crucially, none of them runs the Internet or can switch it off.
Notice the pattern: these bodies coordinate the few things that genuinely must be globally unique — names, numbers, and protocols — and nothing else. Peering deals, pricing, what runs where, what content is served: all of that is decided independently by each network. The Internet is governed the way a language is governed: there is a dictionary and a grammar everyone shares, but no one owns the conversations.
It is tempting to imagine a master control room, a "main server," or a company that "owns the Internet" and could turn it off. There is no such thing. The Internet is a voluntary interconnection of tens of thousands of separately owned networks that have simply agreed to speak the same protocols and forward each other's packets. No single organisation operates it, and no single switch turns it off.
The coordinating bodies — IETF, ICANN/IANA, the RIRs — manage shared naming and numbering so that identifiers don't collide; they do not run the pipes or route the traffic. That is why a government can censor or throttle the Internet within its own borders (by leaning on the networks physically inside the country) but cannot "shut down the Internet" globally — there is no global switch to reach. The lack of a centre is not an oversight; it is the core design principle that makes the Internet robust, and it is exactly why it survived growing from a handful of research computers to the whole planet without ever being re-architected.
Most routers, when they don't know where to send a packet, fall back on a default route — "if in doubt, send it upstream to my provider." But the core routers inside tier-1 networks have no upstream to defer to. They must know a route to every address block on the entire Internet explicitly, because there is no "up" left. This full-knowledge core is called the default-free zone (DFZ), and the routing table it must hold — the list of every reachable prefix on Earth — now exceeds a million entries and grows every year. It is the closest thing the Internet has to a complete map, and even it is stitched together from advertisements rather than authored by anyone.