Paul Baran, Donald Davies, and Leonard Kleinrock — and the messy question of who invented packet switching
Between 1959 and 1966, on two continents, in three institutions that had almost nothing to do with one another, the same idea surfaced three times. A military think tank in Santa Monica, a government laboratory outside London, and a graduate student’s office at MIT each arrived at the notion that data should travel across a network not as a continuous stream on a dedicated line, but chopped into small, independently addressed chunks that find their own way to the destination.
That idea is packet switching. It is the operating principle of essentially every network you will use today. And its origin story is a favourite case study for historians of technology, partly because it is a textbook example of simultaneous invention, and partly because the three men involved did not agree — publicly, and for decades — about who had actually done what.
What the idea actually was
To judge the competing claims, it helps to be precise about what was being invented.
The telephone network of the early 1960s operated on a circuit-switching basis. Placing a call meant reserving a physical path from one end to the other for the duration of the conversation. This is excellent for speech, which flows steadily for minutes at a time. It is catastrophically wasteful for computers, which are silent for long stretches and then transmit in violent bursts.
Packet switching replaces the reserved circuit with three intertwined ideas:
- Segmentation — a message is broken into small pieces of roughly uniform size, each carrying its own destination address.
- Store-and-forward switching — each node in the network receives a piece in full, examines it, and forwards it onward, with no end-to-end path reserved in advance.
- Adaptive, distributed routing — no central authority plans the route. Each node decides where to send each piece based on its own local, continuously updated view of the network.
Store-and-forward on its own was not new; telegraph systems had used a version of it for decades, with messages sometimes literally torn off a tape and carried across the room. What was new was the combination: small pieces, no reserved path, and routing decisions distributed across the whole network rather than concentrated in a switching centre.
Keep those three ingredients in mind. Much of the later argument turns on which of them each man actually proposed.
Paul Baran: designing for the day after
Paul Baran joined the RAND Corporation in 1959. RAND was the US Air Force’s in-house thinking apparatus, and the problem placed before Baran was as bleak as the era that produced it: after a nuclear first strike, could the United States still communicate well enough to make a considered decision about retaliation? The existing long-distance network, built around a small number of large switching centres, would not survive. Its fragility made a hair-trigger response more likely — if you might lose the ability to give any order at all, the pressure to give the order immediately becomes enormous.
Baran’s answer, developed over roughly five years, was a network with no centre. He drew the now-famous distinction between centralised, decentralised, and distributed topologies — a mesh in which every node connects to several neighbours and no single node matters much. His analysis produced a striking practical result: a redundancy level of only three or four links per node was enough to make the network survive attacks that destroyed a large fraction of it.
To move data across such a mesh, he proposed dividing messages into standard units he called message blocks, sized at about 1,024 bits. Each block would carry addressing and a handover count, and each node would run what Baran nicknamed “hot-potato” routing — get rid of the block quickly, toward whichever neighbour currently looked best. Nodes would learn the network’s shape from experience, updating their own routing tables as blocks arrived, with no master map anywhere. Baran also insisted the whole thing be digital, at a time when that was still a contested position.
The work appeared in 1964 as On Distributed Communications, an eleven-volume RAND series, most of it unclassified and openly available.
Then it went nowhere. AT&T, which would have had to build it, was dismissive; Baran later recounted engineers telling him, essentially, that he did not understand how telephony worked. The Air Force approved the project, but responsibility for implementation was routed to the Defense Communications Agency, an organisation Baran judged incapable of building it correctly. Rather than see a botched version discredit the concept, he withdrew the proposal. The most complete design for a packet network in existence sat in a report series for several years, admired by a small number of people and built by nobody.
Donald Davies: designing for the impatient user
Donald Davies came to the same architecture from the opposite direction, and for reasons that had nothing to do with survivability.
Davies had been at Britain’s National Physical Laboratory since the 1940s, where as a young man he worked with Alan Turing on the ACE computer project. By the mid-1960s he had become interested in time-sharing — the then-novel idea that many users at terminals could share one computer interactively. In 1965 he began thinking about what happens when those terminals are not down the hall but across the country, connected by telephone lines.
The economics were absurd. An interactive user types a line, thinks, reads a response, thinks some more. A dedicated telephone circuit sits idle for the overwhelming majority of that session while being billed for every second. Davies concluded that the traffic pattern of interactive computing demanded a fundamentally different kind of network — one that carried short bursts from many users over shared high-speed links, allocating capacity only in the instants it was actually needed.
He wrote his ideas up in an internal note in late 1965, presented them publicly in 1966, and set them out in a proposal for a national digital communication network in June of that year. He also gave the concept its name. Looking for a short word for the small unit of data, one that would survive translation into other languages, he settled on packet. It is Davies’s term, not Baran’s, that we all use.
Davies did not learn of Baran’s work until 1966, after his own design was substantially complete — one of the cleanest documented cases of independent invention in computing. His reaction, characteristically, was not to dispute priority but to point out how thoroughly Baran had already covered the ground.
Britain never funded the national network. What Davies did build was a local packet network at NPL, running from around 1969 and serving the laboratory’s own users — a real, working packet-switched system, small in scale but ahead of almost everything else in operation.
Leonard Kleinrock: the mathematics of queues
Leonard Kleinrock was a doctoral student at MIT from the late 1950s, working in a building saturated with early computer networking interest. His thesis, completed in 1962, addressed message delay in communication networks with storage — that is, in store-and-forward networks where messages queue at intermediate nodes waiting for an outgoing line.
This was a genuinely hard problem. Queueing theory could handle a single queue cleanly, but a network is a system of interlinked queues where the output of one becomes the input of the next, and the mathematics becomes intractable. Kleinrock’s contribution was a set of analytical techniques — most famously an independence assumption that made the equations solvable while remaining accurate enough to be useful — that let engineers predict delay and throughput in networks of queues. The thesis became the 1964 book Communication Nets.
In 1963 he moved to UCLA. When ARPA began building its network at the end of the decade, Kleinrock’s laboratory was designated the Network Measurement Center, and his was the site chosen for the first node. The first Interface Message Processor was installed there in September 1969, and on 29 October 1969 his team sent the first message across the ARPANET to Stanford Research Institute. The system crashed partway through the word “login,” which is why the first thing ever transmitted over the network’s ancestor was the fragment “lo.”
Convergence: Gatlinburg, 1967
The three streams met at an ACM symposium in Gatlinburg, Tennessee, in October 1967.
Lawrence Roberts, then planning ARPA’s network, presented the ARPANET concept. Roger Scantlebury, from Davies’s group at NPL, presented the British packet-switching design — and told Roberts about Baran’s RAND volumes, which Roberts had not read. The encounter had immediate, concrete effects. ARPA’s plan had assumed links running at 2.4 kilobits per second; on the strength of the NPL analysis, that was revised upward to 50 kilobits per second. Roberts subsequently consulted Baran directly on routing.
The ARPANET that went live in 1969 was therefore not any one man’s design. It was Baran’s distributed routing philosophy, Davies’s terminology and traffic reasoning, Kleinrock’s analytical tools, and a great deal of original engineering from Roberts, the BBN team who built the IMPs, and the graduate students who wrote the host protocols.
The argument
The polite version of this story ends with “three independent inventors.” The actual history is more contested, and it would be dishonest to leave it out.
Kleinrock has long presented his 1961–1962 MIT work as the invention of the mathematical theory of packet switching, and this framing appears in a great many institutional biographies, awards citations, and popular accounts. A number of his contemporaries have publicly disagreed. Their objection is specific rather than personal: the thesis analyses delay in store-and-forward networks carrying whole messages of variable length. It does not propose breaking messages into small fixed-size pieces, it does not propose distributed adaptive routing, and it does not describe a network of the kind Baran and Davies described. On this reading, Kleinrock analysed message switching brilliantly, and the distinct innovation that made packet switching came from elsewhere.
Donald Davies spent part of his last years working on a careful historical assessment of the question, published after he died in 2000. Several ARPANET veterans, including Bob Taylor and Alex McKenzie, made similar arguments in public over the years. Kleinrock has consistently maintained his account, pointing to sections of his work that model the effect of breaking messages into smaller units.
Paul Baran took the most gracious line, and probably the most historically accurate one. He compared the invention of the network to the building of a cathedral: many people contributing over many years, each convinced that their own stone was the important one, and none of them wrong exactly, but none of them the whole story either. He consistently credited Davies for the name and for independent invention, and declined to claim the thing for himself.
Why three times?
Simultaneous invention is not a coincidence; it is a signal that the preconditions have arrived. By the early 1960s, three conditions held everywhere at once. Digital computers had become fast enough that the telephone network was now the bottleneck. Solid-state electronics had made it economically plausible to put a small computer at every switching point rather than a room full of relays. And the analytical tools for reasoning about traffic in networks were mature enough to prove that the resulting design would actually work.
Given those conditions, the idea was available to anyone who looked hard at the mismatch between how computers communicate and how telephone networks were built. Baran looked at it through the lens of surviving a nuclear exchange. Davies looked at it through the lens of a user waiting for a terminal to respond. Kleinrock looked at it through the lens of queueing mathematics. They saw overlapping pieces of the same object.
The lesson is not that credit should be split three ways in equal shares. It is that “who invented it” is often the wrong question to ask about an infrastructure. Nobody invented the network. A dozen people, over a decade, each solved the part of the problem they could see from where they stood — and the thing that emerged belonged to all of them and to none of them.
Further reading
- Paul Baran, On Distributed Communications, RAND Corporation, 1964 — the eleven-volume series, freely available from RAND’s website.
- Donald Davies’s 1966 proposal for a national digital communication network, and the NPL papers presented at Gatlinburg in 1967.
- Leonard Kleinrock, Communication Nets: Stochastic Message Flow and Delay, 1964.
- Janet Abbate, Inventing the Internet (MIT Press, 1999) — the standard scholarly history, and even-handed on the priority question.
- Katie Hafner and Matthew Lyon, Where Wizards Stay Up Late (1996) — the readable popular account of the ARPANET’s construction.
- The Computer History Museum and Charles Babbage Institute oral history collections, which include long interviews with most of the principals in their own words.