The bet almost everyone lost

Every layered network diagram a student is shown begins with seven layers, and almost every network in the world runs four. That gap is the residue of a twenty-year argument in which the losing side had, on paper, every advantage.

By the mid-1980s Open Systems Interconnection (OSI) had the support of the International Organization for Standardization and the international telecommunications bodies; of the postal, telegraph and telephone monopolies that then owned most of the world's wires; of IBM, Digital Equipment Corporation, Honeywell and their competitors; of the governments of France, West Germany, the United Kingdom and the European Economic Community; of the United States Department of Commerce; and of the United States Department of Defense, which planned to migrate off (Transmission Control Protocol / Internet Protocol) and onto it. Thousands of engineers worked on it. Its worldwide adoption looked inevitable.

The other side was a protocol suite written largely by graduate students, specified in documents titled requests for comments, funded by a research agency, and given away with the source code.

What came before both

The argument did not start as OSI versus the internet. Through the 1970s several answers existed at once. IBM shipped Systems Network Architecture in 1974 and Digital shipped , each excellent and each able to connect only that manufacturer's equipment - which was, from the manufacturer's point of view, a feature. The telephone companies, collaborating internationally, produced in 1976 and built public data networks on it with genuinely global coverage, on the assumption that data networking would be a service like telephony, sold by the minute and by the packet.

OSI was the reaction to all of it: an attempt to produce one complete, -neutral, internationally agreed architecture so that any manufacturer's equipment could talk to any other's. The motive was right, the model was careful, and its seven-layer reference architecture is still the vocabulary the industry teaches - which is the strange afterlife the OSI model article describes: the map outlived the territory it was drawn for.

The mandates

Governments did not merely prefer OSI. They ordered it.

The United Kingdom's Central Computer and Telecommunications Agency published a Government OSI Profile in 1988 and began work with France and West Germany on a common European procurement handbook. The United States published its own as a Federal Information Processing Standard, 146, in August 1988, reissued as 146-1 in April 1991: federal agencies buying networking equipment were required to buy OSI. In 1988 the Department of Defense adopted the OSI mail and file transfer protocols as co-standards alongside its own, designating them experimental "because of the limited operational experience currently available with the OSI protocols".

That last phrase is the whole story in a subordinate clause. The mandate existed; the operational experience did not.

What happened next is documented in the internet's own archives and is worth quoting for its dryness. 1169 was published to explain the role of , and the explanation, as one participant summarised it on a public mailing list, was this: GOSIP required government purchasers to buy OSI, but did not require them to use it as their primary or only means of communication. Agencies bought conforming equipment to satisfy the rule and ran TCP/IP over it to get their work done. Exceptions were granted freely, on the grounds that OSI implementations from different manufacturers could not reliably interoperate with each other - which was the one thing OSI existed to guarantee.

By 1994 the American mandate had been softened to permit the internet protocols outright, and the profile was withdrawn.

Why the better specification lost

Four reasons, and none of them is that OSI was badly designed.

It was specified before it was implemented. OSI documents were produced by committees seeking agreement among national delegations, then handed to vendors to build. TCP/IP was built first and written down afterwards; a specification that describes working code cannot contain a requirement that turns out to be impossible. This is the IETF's rough consensus and running code, stated as a competitive advantage rather than a slogan.

The documents cost money and the code did not. ISO standards were sold. RFCs were free to read and free to implement, and by the early 1980s a working TCP/IP stack shipped in the Berkeley distribution that universities already ran. A generation of engineers learned networking on the free thing because it was the thing in front of them.

Its own openness was procedural rather than practical. The historian Andrew Russell's argument, which has become the standard reading, is that ISO's attempt to use formal democratic mechanisms of international standardisation may have been its downfall: consensus among national committees is slow, and the delegates arriving at those committees represented incumbents - carriers and mainframe manufacturers whose interests were served by a design in which the network was intelligent, centrally administered and billable. The internet's design put the intelligence at the edges and assumed the network was dumb, which was a worse fit for the people funding OSI and a better fit for the people building applications.

It was answering the previous question. OSI was architected for a world of terminals, mainframes and carrier-provided links. While the committees worked, arrived from Xerox and made local networking cheap, workstations replaced terminals, and the interesting problem stopped being how to connect a terminal to a mainframe and became how to connect one network to another. TCP/IP was designed for exactly that second problem, because it had been built to join networks that already existed and were not alike.

The epitaph belongs to Einar Stefferud, an internet advocate of the period, who said as the project stalled: OSI is a beautiful dream, and TCP/IP is living it.

The stacks that lost, and where they went

DECnet ran the world's technical computing for a decade and reached, in Phase V, a design that could carry both its own protocols and OSI's; it lost its market when its manufacturer did. did something no other stack managed, which was to work with no configuration at all - devices found each other and named themselves - and it was dropped in favour of TCP/IP once every Mac had an internet connection, though its zero-configuration ideas came back as Bonjour and are in every printer and speaker on a home network today. X.25 persisted for decades in banking and point-of-sale networks and in places where an operator's contract mattered more than throughput. never really died; it was tunnelled.

None of them was defeated technically. Each lost an installed base, and the four things above are the reason.

What a practitioner should take from it

The useful part is not the history; it is the shape of the decision, because the same shape recurs constantly. A more complete specification, endorsed by more institutions, with more vendors committed to it, lost to something cheaper, running, and freely implementable - and the mandate that was supposed to settle the matter was satisfied on paper by buyers who did the opposite in practice.

Anyone who has watched an organisation buy a strategic platform and quietly keep running the thing that works has seen the same event at a smaller scale. When the question is which of two technologies to build on, the number of institutions endorsing one of them is weak evidence. What people are already running, what a new engineer can learn without permission, and whether the specification describes something that exists are strong evidence. That is what the protocol wars decided, and it is why the network this site is about is the one that won.

Sources