For most of the early 1980s, if you wanted to run serious Lisp, you bought a machine designed to run nothing else. Symbolics in Cambridge, Lisp Machines Inc. with its CADR-derived workstations, Xerox with the Dandelion and Daybreak, Texas Instruments with the Explorer. Each box was a small architectural argument: stack hardware, tagged memory, a microcoded instruction set tuned to the cost profile of a language with garbage collection and dynamic typing. The machines ran the Genera environment or its cousins, which many people who used them still describe as the most coherent development experience they ever had. A CADR sold for around fifty thousand dollars, an LM-2 closer to seventy, in the money of the time.

The business case held as long as the alternative was a conventional minicomputer struggling to execute the same code through a software interpreter. By 1987 that case was gone. Sun's workstations, particularly the Sun-3 line, were running compiled Common Lisp from Lucid and Franz on commodity Motorola silicon at prices a research lab could put on a regular procurement form. An Apple Mac II with an MicroExplorer board sat in the same office. The premium for going specialist had become a tax, not an investment.

What turned a slow erosion into a market collapse was DARPA. The Strategic Computing Initiative, launched in 1983, had been the quiet backbone of the Lisp machine business. Many Symbolics customers were ultimately spending federal AI grant money. When funding for the program contracted in 1987 as the Reagan-era defence build-up cooled, that procurement channel narrowed in the same fiscal year that commercial buyers were already pulling back. Symbolics's revenue did not decline gracefully; it fell off a cliff.

Underneath the hardware story was a software story everyone in the industry could already feel. The expert systems boom that had justified the optimism, XCON at DEC, the Authorizer's Assistant on American Express phones, MYCIN in research, was running into the qualification problem. Rules did not generalise. Updates required the original knowledge engineer. Maintenance costs in year three or four often exceeded the system's payback. By 1988 corporate buyers had concluded that what they had been sold as artificial intelligence was, mostly, a brittle and expensive form of structured programming.

LMI went bankrupt before its K-machine reached customers. Symbolics restructured repeatedly through the early 1990s and emerged as a software company selling Open Genera into a niche that has never quite closed. Xerox quietly folded the AI workstation work back into the rest of PARC. The hackers dispersed into Common Lisp standardisation, into the early internet companies, into academia. The hardware did not die for technical reasons. It died because the market discovered it could buy ninety percent of the experience for ten percent of the money, and ninety percent was enough.

The lesson was not new even in 1987. It was the same lesson Pierce's panel had delivered about machine translation twenty-one years earlier: a field can be made to look unviable simply by removing the subsidy that was holding it up. The funders left, and the rest followed. Anyone working on AI infrastructure in 2026 should at least know the shape of the room they are standing in.

Sources: