Summary

  • ARPANET was formally decommissioned in 1990, yet the Internet continued because TCP/IP had turned it from the network into one network among many.
  • Continuity was engineered before retirement. The 1983 MILNET split separated operational military traffic, successor backbones grew, and a 1989 routing change let NSFNET reach the Defense Data Network without traversing ARPANET.
  • The episode supports a demanding rule for institutional change: protect the function, migrate every dependency, and make the incumbent replaceable before withdrawing it. Protocol freedom does not remove the need for funded paths, identifiers, operations or accountability.

A shutdown that did not become an outage

The most revealing fact about ARPANET’s end is what did not happen. In 1990 the original packet-switching network was formally decommissioned. The Internet did not close. Email did not return to a four-node experiment. A single government decision did not erase the growing collection of university, federal, regional and commercial networks that had learned to speak IP.

That sounds obvious only in retrospect. Institutions that begin a system often become indistinguishable from the function they perform. Their histories become claims of necessity: without this operator, this database, this backbone or this committee, continuity supposedly ends. ARPANET offers a harder standard. Its achievement was large enough to survive the network itself.

The exact shutdown date is less secure than the familiar story suggests. Popular accounts often name 28 February 1990, while historical discussions note that different equipment and connections ceased at different times. The authoritative record is firm at the year and the outcome. DARPA says the original ARPANET was formally decommissioned in 1990; RFC 1302 says it was laid to rest that year; RFC 1174, published in August, speaks of its demise as a completed condition. That is enough. The mechanism matters more than a ceremonial minute.

First the network, then a network

ARPANET began as a specific communications facility. In 1968 ARPA contracted BBN to build Interface Message Processors, and the first four sites came online in 1969. Hosts reached one another through a packet-switching subnet of leased lines and IMPs. At that stage, access to the network and participation in the experiment were closely aligned.

Internetworking changed the noun. DARPA research connected packet radio, satellite and terrestrial networks that could not be reduced to one transmission system. TCP/IP provided a common layer across those differences. After the 1 January 1983 protocol conversion, an Internet host’s identity and reachability no longer had to mean attachment to one kind of packet switch. ARPANET remained important, but it was becoming a component of a larger system.

RFC 1009 captured the changed geometry in 1987. Its “global-interconnect system” included ARPANET, the NSFNET backbone, NSF regional and consortium networks, MILNET, SATNET, WBNET and other networks. The document did not describe a homogeneous machine. It described independently administered networks joined by gateways and common protocol behaviour.

This is the architectural precondition of institutional replaceability. If the Internet had remained merely a larger ARPANET, shutting ARPANET would have meant shutting the Internet. Once the service was communication across networks, any one network could in principle be replaced—provided that its remaining dependencies were first removed.

Separation before retirement

The retirement sequence began years before 1990. RFC 881 is a domain-name deployment plan, not a narrative history of the split. Its schedule nevertheless dates the ARPANET/MILNET logical separation to 4 October 1983 and gateway access controls to February 1984. RFC 942 subsequently describes ARPANET and MILNET as the DDN's two major networks.

The separation distinguished research from military operations while preserving controlled interconnection, reducing the functions for which the original network had to remain the common home. RFC 878, sometimes confused with this evidence, specifies the 1822L host-access protocol; it is not the split record.

Civilian research connectivity was also moving. NSF funded CSNET and then NSFNET, which launched in 1986 to connect supercomputer centres and academic networks. NSF says NSFNET connected roughly 2,000 computers in 1986 and more than two million by 1993. Those numbers should not be read as a simple transfer of a crown from DARPA to NSF. They show that capacity and users were accumulating outside the original substrate.

Other federal and regional networks mattered too. ESnet, NASA’s science networks and campus networks formed part of the expanding interconnect. Funding, equipment and operating responsibility were plural. TCP/IP made them compatible, but compatibility alone did not move packets. Gateways, circuits, route announcements and people answering failures supplied the physical and institutional substance.

The route that made removal real

RFC 1133 provides unusually concrete evidence of how a network becomes dispensable. Until mid-1989, NSFNET and the Defense Data Network exchanged traffic through intermediate routers attached to ARPANET. Three gateway arrangements provided redundancy. ARPANET was no longer the whole Internet, yet a path between two large parts of it could still depend on traversing that aging network.

During the first half of 1989, engineers installed Ethernet interfaces in two DDN Mailbridges, one at NASA Ames on the west coast and one at Mitre in Virginia. Those systems could exchange routing information directly with NSFNET nodes. The RFC states the consequence plainly: reaching MILNET sites from NSFNET no longer required traversal of ARPANET.

The details prevent the transition from becoming a slogan. The two interconnections used different autonomous-system views. Network announcements were divided across them, metrics expressed preferred paths, and either link could take the full load if the other failed. NSFNET and DDN administrators retained control over what they announced and accepted. Continuity was produced by two operational paths, explicit routing policy and fallback—not by declaring that the old network was obsolete.

This is the moment the retirement thesis becomes testable. A dependency had been identified, successor paths were installed, traffic could be rerouted, and separate operators could exercise bounded control. ARPANET could then disappear without silently severing DDN reachability from the civilian research backbone.

What survived the hardware

The physical subnet was only one layer. Network numbers, names, host information, documentation and support also had to persist. RFC 1302 explains that SRI’s original Network Information Center had grown from the task of maintaining the official host table and RFC repository. After the ARPANET/MILNET split and ARPANET’s retirement, the DDN-NIC continued military information services while other networks operated their own NICs.

RFC 1174 shows the administrative model changing at the same historical boundary. “With the demise of the ARPANET and the growth of a global Internet,” it argued, identifier administration had outgrown a client base of military, government and government-sponsored research organisations. The old binary label of globally “connected” status no longer captured a world in which each network chose its interconnections and transit policies.

The recommendation was not to abandon uniqueness. It was to separate registration from a single US sponsorship test, retain policy information that networks could use, and consider distributing registry functions across multiple centres once registration was divorced from policy enforcement. The original network could die because the invariant—globally useful identifiers—was distinguishable from the institution and eligibility rule that had once administered it.

No miracle of decentralisation

ARPANET’s retirement is sometimes compressed into a triumphal story: decentralised protocols made the network impossible to kill. That is too easy. The Internet survived this particular retirement because governments had paid for research, successor backbones and interconnection; institutions cooperated on routing; operators moved dependencies; and common identifiers and protocols remained available.

Nor did 1990 mark instant privatisation. NSFNET remained a federally supported backbone until 1995. DDN functions continued. Registration still required administration. The Internet’s institutional centre of gravity changed over several years, and public funding remained central to the transition.

The accurate claim is narrower and more useful. No single physical network remained constitutionally necessary. Shared protocols permitted different institutions to carry the function, but only after real alternative capacity existed. Decentralisation was an option created by architecture and completed by operations.

Sources and evidence limits

DARPA’s ARPANET history supplies the institutional chronology. RFC 881 supplies the dated deployment schedule; RFC 942 provides background on the resulting two-network DDN. The multi-network architecture appears in RFC 1009, and the decisive routing transition in RFC 1133. RFC 1174 records the contemporaneous identifier-policy problem, while RFC 1302 describes continuity of network information services. The NSF account and the Internet Society’s participant-authored brief history provide successor-network scale and synthesis.

The evidence supports formal retirement in 1990, not one universally verified switch-off instant. It also supports continuity of the Internet, not continuity of every ARPANET circuit, host or user experience. Retirement should therefore be understood as a staged operational process whose success can be judged by migrated functions and reachable networks.