ARPANET and the Cold War Birth of Packet Switching

A Pentagon problem: a command structure that cannot be beheaded

By the early 1960s, American nuclear war planning had settled on a grim design problem. Command and control ran through a small number of hardened centers connected by long-distance telephone circuits, the same hierarchical switching architecture the Bell System used for ordinary calls between any two American cities. A Soviet strike aimed at a handful of key switching centers could sever the chain of command before any retaliation order ever left Washington, a single point of failure sitting at the center of the country's entire strategic response. The RAND Corporation, an Air Force funded research organization in Santa Monica already known for its work on nuclear strategy, took the problem as a formal research assignment.

Paul Baran, an engineer at RAND, spent the early 1960s working out an answer under Department of Defense funding. He called it distributed adaptive message block switching. Instead of routing a call through a small set of central offices, Baran's design broke a message into small blocks, addressed each block, and let it find its own way across a mesh of many redundant links, node to node, with no office whose loss would silence the whole network. A message could survive the destruction of most of the network's paths, because it never depended on any single one.

Baran's own writing is explicit about the goal: a network that keeps working after an attack destroys much of it. Historians have since complicated the tidier version of the story that grew up around that goal. Baran's RAND work was built for exactly that purpose in the early 1960s, but ARPANET, the network the Pentagon actually built starting in 1966, was chartered for a narrower and more mundane reason: letting university researchers share scarce and expensive computer time over long distance lines. The survivability rationale and the resource-sharing rationale belong to two different projects three years apart, run by different people, and the popular account that flattens them into one origin story is a real point of dispute among historians of the network, addressed further below.

Two engineers, two coasts, one idea

Baran was not alone in reaching for this design. Beginning in 1965, Donald Davies, a computer scientist at the UK's National Physical Laboratory, worked out the same core idea independently, without knowledge of Baran's RAND reports. Davies gave the idea its name. He called the small, addressed blocks of data packets, and the method packet switching, the term that stuck and the one still used for the architecture underneath the modern internet.

The two men were solving different problems with the same tool. Baran wanted a network that would keep working through a nuclear strike. Davies wanted a network efficient enough to let many users share a small number of expensive computers over ordinary telephone lines, a British government funding priority in an era when a single mainframe cost more than most university departments could raise. Neither man had the other's motive, and each arrived at the same architecture. That kind of convergence, two national research programs solving unrelated problems and landing on one design, is itself a marker that the underlying idea was, by the mid 1960s, close to inevitable given the state of computing and communications engineering at the time.

When Larry Roberts took over as ARPANET's program manager in 1967, he built the American network on Davies' packet-switching concept rather than reinventing it, having encountered Davies' published work directly. Credit for the architecture that carries the world's data now belongs, honestly, to a defense researcher on the American West Coast and a government scientist in England, working three years and an ocean apart, neither one able to see the network the other's idea would eventually become.

Building it: from initiation to first message

Bob Taylor, running the Advanced Research Projects Agency's computing office, started the ARPANET project in 1966 to solve a cost problem his office faced directly. Taylor had three separate terminals in his Pentagon office, one for each of three different time-sharing computers his agency funded, because none of the systems could talk to another. Connecting them into one network, building on ideas Taylor had absorbed from the psychologist and computer scientist J. C. R. Licklider, was the plan Taylor proposed, and Larry Roberts was hired away from MIT's Lincoln Laboratory to run it.

Roberts folded Davies' packet-switching design into the plan and, in 1969, the agency awarded the contract to build the network's switching hardware to Bolt Beranek and Newman, a Cambridge research firm. Bob Kahn led the design of BBN's Interface Message Processors, the dedicated minicomputers that would sit at each site and do the actual work of breaking messages into packets and routing them, and wrote the network's first operating protocol.

At 10:30 in the evening on October 29, 1969, a UCLA student named Charley Kline attempted to send the word login from UCLA's computer toward Bill Duvall at the Stanford Research Institute, roughly 350 miles north. The system crashed after the second letter. The first message ever carried across what would become the internet was not a finished word. It was two letters, L and O, sent about an hour before the connection was restored and the rest of the word arrived.

From four nodes to a defense backbone

By December 5, 1969, four sites were connected: UCLA, the Stanford Research Institute, the University of California at Santa Barbara, and the University of Utah. Growth from there was steady rather than explosive. By 1971, secondary accounts of the program's own records put the network at roughly 15 nodes and 23 host computers, a network still small enough that its engineers knew nearly every machine on it by name.

For its first years, ARPANET stayed what it was chartered to be: a research tool for a small circle of university and defense computer scientists. That changed administratively in 1975, when operational control passed from the Advanced Research Projects Agency to the Defense Communications Agency. The move formalized what the network had already become in practice, a piece of working defense infrastructure rather than an open-ended research experiment, and it kept the network under military operation for another fifteen years even as the wider network of networks built on its ideas grew up around it.

The French government funded CYCLADES project deserves its place beside this history as a parallel state-funded contributor to the design that came after ARPANET. Directed by Louis Pouzin and funded by the French government through IRIA starting in the early 1970s, CYCLADES worked out the principle that the network itself should not be responsible for guaranteeing delivery of a message, that job should sit with the computers at each end of the connection. That end-to-end principle, developed on French government money at the same time the Pentagon was funding its own network an ocean away, became a foundation of the Internet Protocol design that eventually let ARPANET, and every network like it, interconnect into one system.

State money, private fortune

None of this was built to make money. ARPA's funding ran on defense research budgets, evaluated against military and scientific goals, with no expectation of commercial return and no patent wall built around the protocols that resulted. Kahn's Interface Message Processor design, Davies' packet-switching concept, the queueing-theory proofs Kleinrock produced, and the end-to-end principle out of the French program were all published and shared among researchers rather than owned and licensed. This was the ordinary practice of a defense research program, not a deliberate act of generosity, and it turned out to matter enormously for what came after.

A privately funded telephone company or computer maker building the same network in the same years would have had every commercial reason to keep the design proprietary, the way IBM's own networking systems and the international carriers' switched data services of that period were built and licensed. A network built on public defense money, evaluated by researchers rather than a sales division, had no such incentive. The specifications could be published, extended, and copied by anyone with the technical means to do it, and over the following two decades, that is roughly what happened.

The Defense Communications Agency ran ARPANET as military infrastructure through the late 1970s and 1980s while, alongside it, the wider set of ideas it had proved kept spreading through universities, corporations, and eventually consumer markets. By the time ARPANET itself was formally shut down in 1990, partnerships between its research successors and the telecommunications and computer industries had already assured the network's ideas would carry on as a private, commercial internet, no longer under any single owner's control.

The dollar figure on the other side of that handoff is not one number, and no accounting method reduces the entire commercial internet to a single return on a specific defense budget line. What can be said is narrower and still large: a communications architecture developed with public money, released without patent enclosure, became within a generation the substrate for a global economy measured in the trillions, and no government treasury holds equity in any part of it. The same openness that let a university lab, and later a two-person startup, build on the network for free also meant that no public fund ever captured a share of what was eventually built on top of it. Both things are true about the same decision, made for reasons that had nothing to do with either outcome.

A network that could not be beheaded could not be owned either

The design choice that gave ARPANET its shape, no center whose loss would kill the whole system, was made to answer a military threat. It carried a geopolitical consequence its funders were not thinking about at the time. A network with no headquarters has no headquarters to seize. No single government, company, or command post ever sat at the center of the system that grew out of ARPANET's architecture, because there was no center built into the design for anyone to occupy.

That structural fact has shaped real disputes for decades, not abstract ones. Arguments over who governs core internet functions, over which country's laws bind an infrastructure with no national capital, over whether any single authority can or should be able to cut a region off from the rest of the network, all trace back to an architectural choice made for a Cold War threat that never materialized. Decentralization diffused power away from any one actor, which is part of why the network could grow across borders and companies that never agreed on much else. It also meant no single body was ever fully accountable for what moved across it, a real and still unresolved cost sitting beside the benefit the same design produced.

The line from that 1969 design choice to the present is a long one, and it should not be forced. But there is a real thread worth naming plainly: the open, no-single-owner architecture built to survive a Cold War attack is the same substrate that today's search engines and AI answer systems crawl, read, and cite from when they decide which businesses and institutions to name in an answer. The infrastructure has no gatekeeper by design. What decides who gets found and read on top of it is a separate and much more recent contest, but it is being fought on a network whose fundamental openness was set, by engineers who never imagined it, in a Pentagon research contract half a century before the first answer engine existed.