A Brief History of the Internet
ARPANET → NSFNET → commercial Internet. The forty-year arc and the design decisions that ossified into today's stack.
Summary#
The Internet did not appear. It accreted. Across roughly forty years — from a four-node research network spun up in 1969 to a planetary substrate carrying tens of zettabytes a year by the late 2020s — a sequence of half-deliberate, half-accidental choices ossified into the stack we use today. The major beats: ARPANET (1969, packet switching as a research bet), TCP/IP standardisation (1974 paper, 1983 cutover), NSFNET as the academic backbone (1986–1995), the commercial Internet when NSF withdrew (1995), the web era (1991–2001), the mobile and cloud era (2007 onward), and the CDN and consolidation era that defines today.
Knowing the history is not nostalgia. Every quirk of the modern stack — IPv4 exhaustion, BGP’s trust model, why HTTP is unencrypted by default, why DNS is a tree — is a fossilised decision made under specific constraints at a specific year. The arc explains the artefacts.
Why it matters#
The Internet’s shape is the shape of the choices that built it. Three examples:
- IPv4’s 32-bit address space was set in 1981, when 4.3 billion addresses looked extravagant for a research network. IPv6, NAT, CGNAT, and a global address market all exist because that choice ossified before the consumer Internet arrived.
- BGP’s trust-by-default model was set in 1989, when the Internet had a few dozen autonomous systems run by people who knew each other. Route hijacks (Pakistan-YouTube 2008, Rostelecom 2017, Facebook 2021) are downstream of that choice.
- HTTP being cleartext was set in 1991, when Tim Berners-Lee was prototyping at CERN. The decade-long pivot to HTTPS-everywhere (Let’s Encrypt 2015, HSTS, Chrome’s address-bar warnings) is paying off that debt.
Engineers who know the history understand why “just change it” is rarely an option. Replacing IPv4 took 25+ years. Replacing BGP is a research problem. The protocols we live with were never designed for the scale or threat model they now face — but they got there first, and the cost of replacement is now astronomical.
How it works#
The arc has six recognisable phases.
1. The ARPANET era (1969–1983)#
ARPA (later DARPA), the US defence research agency, funded a packet-switched network in the late 1960s as an experiment. BBN (Bolt, Beranek and Newman) won the contract to build the Interface Message Processors — the routers of their day. On 29 October 1969, the first message was sent from UCLA to Stanford Research Institute. The student typed LOGIN; the system crashed after LO. Four nodes were up by year-end: UCLA, SRI, UC Santa Barbara, and the University of Utah.
The protocols of the era — NCP (Network Control Program) — were tied to ARPANET’s specific hardware. By the early 1970s, multiple incompatible packet networks existed (ARPANET, NPL in the UK, CYCLADES in France, packet radio, packet satellite). Vint Cerf and Bob Kahn published A Protocol for Packet Network Intercommunication in 1974, sketching what became TCP/IP — a protocol stack designed to interconnect heterogeneous networks. The “Internet” name and the idea both come from this paper: inter-net, between networks.
2. The Flag Day cutover (1983)#
On 1 January 1983 — “Flag Day” — ARPANET cut over from NCP to TCP/IP. Hosts that had not migrated were offline. The decision to use a hard cutover rather than a gradual transition was deliberate: NCP and TCP/IP could not interoperate cleanly, and dragging out the migration would have stranded both communities. This is the last successful hard cutover in Internet history. Every subsequent transition (IPv4 to IPv6, HTTP/1 to HTTP/2) has been gradual because the network is now too large for a coordinated flip.
The same year, the Domain Name System was specified (RFC 882, 883), replacing the centrally-managed HOSTS.TXT file SRI distributed by FTP.
3. The NSFNET backbone (1986–1995)#
The National Science Foundation built NSFNET to connect university supercomputing centres. By 1988 it was the de-facto Internet backbone, replacing ARPANET (which was decommissioned in 1990). NSFNET’s Acceptable Use Policy forbade commercial traffic, which forced commercial ISPs to interconnect through their own backbones and peering points. By 1995, NSF withdrew funding and the commercial Internet took over — Sprint, MCI, and UUNET became the early Tier-1s.
This is also the era of the foundational protocols: DNS matured, BGP (1989) replaced the earlier EGP for inter-domain routing, SMTP was standardised (1982), FTP was already in use, and HTTP plus HTML (Tim Berners-Lee at CERN, 1989–1991) appeared just in time for the commercial transition.
4. The web era (1991–2001)#
The World Wide Web was released to the public in 1991. Mosaic (1993) made it visual; Netscape (1994) made it mainstream; the dot-com boom (1995–2000) funded its commercialisation. IPO valuations untethered from revenue, fibre was laid coast-to-coast in volumes that would not be lit for a decade, and the bubble burst in 2000–2001.
But the dark-fibre overhang and the trained engineers turned out to matter more than the lost capital. Google (founded 1998), Amazon (1994), eBay (1995) were the survivors. The first wave of Internet infrastructure — Akamai’s CDN (1998), Cisco’s routers, Sun’s servers — was built in this era.
5. The mobile and cloud era (2007 onward)#
Two events reshaped the Internet in the late 2000s:
- The iPhone (June 2007). The mobile Internet was a real product. Always-on, location-aware, push-notified, app-mediated. Traffic shifted from desktop browsers to mobile apps and from HTTP to HTTPS-by-default within a decade.
- AWS (2006) and the cloud. EC2 made it normal to rent compute by the hour. By 2015, “deploy to the cloud” was the default; by 2020, three providers (AWS, Azure, GCP) carried most of the public Internet’s serving infrastructure.
This is also when CDN consolidation took off (Cloudflare 2010, Fastly 2011 alongside Akamai), HTTPS became the default (Let’s Encrypt 2015), and HTTP/2 (2015) and QUIC / HTTP/3 (2021) replaced HTTP/1.1’s connection model.
6. The consolidation era (today)#
The Internet today is centralised in ways its early architects did not anticipate. A handful of CDNs front a majority of web traffic. Three clouds host most serving infrastructure. Two app stores gate most mobile distribution. The 2021 Fastly outage took down GitHub, Reddit, the New York Times, and the UK government in a single propagation window — a control-plane bug at one CDN.
The protocols still work the same way. The control points moved.
Variants and trade-offs#
Three lessons from the arc that recur in system design:
- The thing that ships first sets the contract. TCP/IP was not the best-designed protocol of the 1980s — OSI’s TP4/CLNP was arguably cleaner. TCP/IP won because it shipped on real networks first and the deployments compounded.
- Hard cutovers stop working at scale. The 1983 Flag Day worked because there were a few hundred hosts. IPv6 has been “five years away” since 2005 because there are now billions of hosts and dual-stacking is the only viable path.
- Trust models ossify with the protocol. BGP, SMTP, and DNS were designed in a small-world era and have spent the last twenty years bolting on authentication (RPKI, DKIM/SPF/DMARC, DNSSEC) with mixed success.
When this is asked in interviews#
Rarely as a pure history question — but the historical framing comes up constantly as context for “why is X the way it is?”
Common framings and what to know:
- “Why is IPv4 32 bits?” — chosen in 1981 when the network was a research artefact. Address exhaustion was projected by the late 1980s; IPv6 work began in 1994 and the spec landed in 1998 (RFC 2460).
- “Why does BGP trust route advertisements by default?” — designed in 1989 between a small number of mutually-known ASes. RPKI and BGPsec are the post-hoc fixes.
- “What was the first website?” —
info.cern.ch, Tim Berners-Lee, 1991. Still served (a historical snapshot) at the same URL today. - “Walk me through how the Internet became commercial.” — NSFNET’s AUP forbade commercial traffic; commercial ISPs peered at MAEs (Metropolitan Area Exchanges) and NAPs; NSF withdrew in 1995; the Tier-1s emerged.
- “Why is the Internet centralised on CDNs?” — performance economics. Serving from one origin to a global audience has 100–300ms RTTs at the edge of the network; serving from a CDN POP has 10–30ms. The web architecture (origin + static assets) is amenable to caching, and CDNs captured that wedge.
Asked most in SRE / networking / infrastructure loops as colour. In product loops, history shows up as “why can’t we just X?” — and the honest answer is usually some flavour of “because the install base from 1995 still has to keep working.”
Related concepts#