What Is the Internet?

Summary#

The Internet is not one network. It is a network of networks — tens of thousands of independently operated autonomous systems (ASes), each running its own internal routing, that have agreed to forward each other’s traffic according to a shared contract. That contract is the IP datagram: a self-describing chunk of bytes with a source address, a destination address, and a payload. Any device that can read an IP header and forward a packet toward its destination is a participant. Everything else — TCP, HTTP, DNS, video calls, BGP itself — runs on top.

The Internet’s genius is what it does not require. No central authority routes packets. No global directory lists every host. There is no shared clock. Each AS makes its own decisions about who it peers with and what paths it advertises. The whole thing holds together because every router, no matter who owns it, agrees on one thing: how to read an IP header.

Why it matters#

Knowing the Internet is “ASes + IP + BGP between them” reframes a lot of operational questions. Why does a packet from your laptop to a server take a particular path? Because each AS along the way advertised a route via BGP, and your AS’s policies preferred that next hop. Why did Facebook go dark for six hours in 2021? Because their own AS withdrew its BGP routes — the rest of the Internet stopped seeing them and there was no oracle to fix it externally. Why does CDN performance vary by region? Because the peering relationships between your AS and the CDN’s AS determine the latency, not the geographic distance.

It also matters for system design. When you place a service in us-east-1 versus eu-west-1, you are choosing which ASes will most efficiently reach it. When you buy a transit link, you are buying access to one AS’s routing table. When you choose between unicast Anycast or DNS-based GSLB, you are choosing whether to push routing decisions into BGP or into the application layer.

How it works#

Three pieces compose the Internet:

+---------------+      +-------------+      +---------------+
|   end host    | ---> |   router    | ---> |   end host    |
|  (laptop,     |      |  (forwards  |      |  (web server, |
|   phone,      |      |   IP        |      |   API,        |
|   server)     |      |   datagrams)|      |   peer)       |
+---------------+      +-------------+      +---------------+
        ^                     ^                     ^
        |    +---------------+|+----------------+   |
        +----| autonomous    |||  autonomous    |---+
             | system A      |||  system B      |
             +---------------+|+----------------+
                              |
                       BGP between ASes

End hosts run applications and the full TCP/IP stack. They originate packets and consume them. Phones, laptops, servers, IoT devices — anything with an IP address and a process talking on a socket.

Routers forward packets. They read the destination IP, consult a forwarding table, send the packet out the appropriate interface. Routers do not care what’s inside the packet beyond the IP header (with rare exceptions like deep packet inspection middleboxes that violate this purity).

Autonomous systems are administrative groupings of routers under one routing policy. A university, an ISP, AWS, Cloudflare, Facebook — each has at least one AS number. Within an AS, an interior gateway protocol (OSPF, IS-IS, or static routes) chooses paths. Between ASes, BGP advertises which prefixes each AS can reach and at what AS-path length.

The whole stack runs on the IP datagram — a header plus payload, addressed source-to-destination, with a TTL that decrements at every router hop. IP is connectionless and unreliable: routers may drop, reorder, or duplicate datagrams. Reliability is pushed up the stack to TCP or to the application.

Variants and trade-offs#

Other axes worth knowing:

• IPv4 vs IPv6 — the Internet runs both in parallel. IPv4 is exhausted; IPv6 has unlimited addresses but rollout is uneven. Most servers are dual-stacked; consumer last-mile is increasingly IPv6 with carrier-grade NAT for IPv4.
• Tier-1 vs Tier-2 vs Tier-3 ISPs — Tier-1s peer with every other Tier-1 for free and never buy transit; the global default-free zone is their table. Tier-2s buy transit from at least one Tier-1. Tier-3s buy transit from Tier-2s. Settlement-free peering is the prize.
• Anycast — the same IP advertised by many physical sites; BGP picks the topologically nearest one. Cloudflare, Google DNS (8.8.8.8), AWS Route 53 all use it.

When this is asked in interviews#

Most often as a warm-up: “what is the Internet?” or “how does a packet get from your laptop to a web server?” The strong answer names the layers (end host, router, AS), the contract (IP datagram), and the inter-AS glue (BGP) — not just “TCP/IP”.

Follow-ups to expect:

• “What’s an autonomous system?” — an independently-administered routing domain with its own AS number, advertising prefixes via BGP.
• “What does a router do that a switch doesn’t?” — forwards by IP (L3) using a routing table built from a protocol like BGP/OSPF, instead of by MAC (L2) using a learned forwarding table.
• “Walk me from a browser to a server.” — DNS lookup, ARP for the gateway MAC, IP datagram out the NIC, hop-by-hop forwarding across ASes via BGP-selected paths, server’s NIC reads it, TCP demultiplexes by port, app delivers the response.
• “What guarantees does IP give you?” — almost none. Best-effort delivery, may drop, reorder, duplicate. TCP layered on top supplies reliability.

Asked most in infrastructure / SRE / network-engineering loops. In product or backend loops, the question is usually quicker and rolls into “and how does that explain DNS/CDN/latency?”.

Related concepts#

• The TCP/IP Model
• The OSI Reference Model
• Encapsulation, Headers, and the Envelope Metaphor
• Access Networks
• Interdomain Routing — BGP