Thinking Machines, What Happened To The Connection Machine Company

In June 1993 the first TOP500 list of the world’s fastest computers was published. The top four machines on it were all built by the same company, and so were five of the top ten. Fourteen months later that company filed for bankruptcy protection.

Thinking Machines Corporation built one of the most striking computers anyone has made, a black cube filled with blinking red lights that went into museums and into Jurassic Park. It employed Richard Feynman, on and off, for five years. It was, briefly, the fastest thing in computing, until, as its designer now tells us, a rival boxed it in with export limits and a Pentagon upgrade law just when it needed the money.

What follows is assembled from federal audit reports, bankruptcy filings, the company’s own hardware manuals, the recollections of people who were there and a letter from Danny Hillis on 27 October 2026. Thinking Machines was privately held and never filed public accounts, so the revenue and profit figures below are the company’s own statements as reported at the time and cannot be checked against an audited filing. Figures taken from the federal audit are identified as such.

A machine with sixty-five thousand tiny brains

Thinking Machines was founded in May 1983 and set up in an old mansion outside Boston, the Robert Treat Paine estate in Waltham. The architect was Danny Hillis, whose doctoral work at MIT was the machine itself. The company and the dissertation were built in parallel, which is why the thesis reads like a design document.

One correction before going further, because it appears in nearly every account. Hillis’s doctoral supervisor was Gerald Sussman, not Marvin Minsky, and the front matter of the thesis says Sussman on both the signature page and the abstract.

Hillis has told us why the accounts differ. His committee, he wrote, “consisted of Gerald Sussman, Marvin Minsky, and Claude Shannon, with John Tukey as the outside reader.” Sussman, much younger than the others, was the official supervisor, but Hillis was living in Minsky’s basement when he started work on the machine.

Minsky and Shannon cared more about using the machine for artificial intelligence than about its engineering, and both helped start the company, Minsky as a co-founder and Shannon as a founding investor. Minsky also recruited Howard Resnikoff as a research director, and our earlier version was wrong to say that Minsky was not a co-founder.

The idea was to invert how computers work. A conventional machine has one fast processor that fetches data, operates on it and puts it back, over and over. Hillis proposed giving every piece of data its own processor, so that instead of one clever thing working very fast you had an enormous number of stupid things all working at once.

The CM-1 as built carried 65,536 processors, each one bit wide with 4,096 bits of memory to itself, running at four megahertz inside a cube about 1.3 metres on a side and drawing roughly twelve kilowatts. Sixteen processors shared a router chip, and the 4,096 routers were wired as a twelve-dimensional hypercube, so no router sat more than twelve hops from any other.

That last detail is worth stating carefully, because it is routinely mangled. The processors are not the corners of the hypercube. The router chips are. Descriptions of a twenty-dimensional machine refer to a million-processor design that was never built, and it survives in the literature only because of a story about Richard Feynman.

Richard Feynman and the router

Hillis wrote the definitive account himself, in Physics Today in February 1989, a year after Feynman died. It is the source for most of this section, alongside the letter Hillis sent us in October 2026, and it is worth reading in full.

Hillis mentioned over lunch that he planned to start a company to build a parallel computer with a million processors. Feynman’s response, in Hillis’s telling, was “That is positively the dopiest idea I ever heard.” For Feynman a crazy idea was an opportunity to prove it either right or wrong, and by the end of lunch he had agreed to spend the summer at the company.

Told his job would be advising on the application of parallel processing to scientific problems, he said that sounded like a bunch of baloney and asked for something real to do. He was sent out to buy office supplies. While he was gone the team decided the part of the machine they were most worried about was the router, and when he came back with the pencils they gave him the router to analyse.

By the end of that summer he had an answer, and he presented it as a set of partial differential equations. Treating a set of boolean circuits as a continuous differentiable system is a natural move for a physicist and a very strange one for a computer designer. His variables were things like the average number of one bits in a message address.

Five buffers or seven

The company’s own discrete analysis said the router chip needed seven buffers. Feynman’s equations said five. They decided to play it safe and ignore Feynman. That was September, and by the following spring the chips they had designed were slightly too large to manufacture, and the only way to fix it was to cut back to five buffers.

The machine worked, and the first program it ran, in April 1985, was Conway’s Game of Life. Feynman tested the machine against quantum chromodynamics, modelling a section of space-time as a discrete four-dimensional lattice, and since the only language he really knew was BASIC he invented a parallel version of it and simulated the program by hand to estimate the speed.

Hillis added a postscript when he wrote to us. Feynman’s calculation, he said, “really just showed that a pattern that would stump the router was very unlikely.” Wati Taylor, whom Hillis describes as a mathematician, later found one such pattern, which would have cycled forever without all seven buffers.

The five-buffer machine never got stuck in practice, but the company added some clever software to make sure that it never would. So Feynman’s equations had been right about the odds, which was enough to build the machine but not enough to guarantee it.

Feynman and quantum computing at Thinking Machines

In a 2016 oral history interview with the Computer History Museum, Hillis said that the project Feynman was working on was quantum computing, and that he was quite sure Thinking Machines was the first company ever to work on it. Our earlier version said the claim did not survive checking. Hillis has since told us that Feynman did work on quantum computing at the company, although the claim to have been first remains his and not one we can verify.

We relied on his own 1989 account in Physics Today, which lists what Feynman did at the company. The list runs from the router and Hopfield associative memories to quantum chromodynamics, cellular automata, fluid simulation with Stephen Wolfram and simulated evolution, and quantum computing is not on it.

Hillis says the omission was his own choice. Feynman kept working on quantum computing at Thinking Machines, Hillis told us, “trying to figure out how to compute with SQUIDs.” These superconducting quantum interference devices are close cousins of the circuits inside many of today’s quantum chips.

“Only a few of us understood what he was doing, so I didn’t attempt to describe it in 1989,” Hillis wrote. His first attempt to describe it came in his 1998 book The Pattern on the Stone.

The dates still show that the interest came first. Feynman’s founding paper of the field, Simulating physics with computers, appeared in the International Journal of Theoretical Physics in June 1982, roughly a year before Thinking Machines was incorporated. It came out of the 1981 Physics of Computation conference at MIT.

Hillis was in the room. He first met Feynman at that conference, along with Charles Bennett, Paul Benioff and Rolf Landauer, and watched them share their early ideas about quantum computing.

The man who wrote the founding paper of quantum computing spent his last five years working, on and off, on a very classical machine, analysing its router and writing a parallel BASIC for it. Alongside that work, and understood by only a few of the people around him, he was trying to work out how to compute with quantum devices.

The lights, and what they actually did

The Connection Machine is the only supercomputer most people can picture, and it is because of the lights. The usual assumption is that they were decoration. The usual correction is that they were purely functional. Both are wrong, and the documentation is unusually clear about it.

Status lights are ordinary things to put on a circuit board, showing whether power is on or a chip is seated. Carl Feynman, Richard’s son and an employee, described a fantasy of the machine as a vast cloud of lights flickering as they sent messages, like neurons firing. The designer, Tamiko Thiel, made those board-level lights visible through the skin of the cabinet.

So they were real status indicators, deliberately exposed. And a program could seize them. The CM-2 instruction set contains an operation that determines whether the lights display internal status or are driven by the user program, and another that latches an arbitrary pattern onto them. A machine with nothing to do went dark, which is why finished programs sometimes carried extra operations purely to keep the face blinking for visitors.

The CM-2 that carried them arrived in April 1987, not 1986 as is often written. It kept the 65,536 processors but gave each one sixteen times the memory, 512 megabytes in total, and added an optional floating-point accelerator using Weitek chips. Alongside it came the DataVault, five gigabytes spread across thirty-nine disk drives. Each 64-bit chunk arriving from the machine was split into two 32-bit words, and each of those took seven error-correcting bits to make 39, one bit to each drive, with three spare drives standing by. It is an early commercial disk array, and Thinking Machines called rebuilding a failed drive healing the database.

Who was actually buying

This is where the story turns, and the source is a federal audit rather than a memoir. In May 1993 the General Accounting Office reported to the House Armed Services Committee on the high-performance computing programme run by ARPA, the Pentagon’s research agency.

The report is blunt. ARPA had facilitated the placement of 44 systems made by Intel and 24 made by Thinking Machines, and had not supported procurements from several major rivals with which it had no development contracts. Government officials, researchers and vendors told the GAO that this had distorted the market for massively parallel machines. The excluded competitors are named as Kendall Square Research, nCUBE and MasPar.

Adding up every disclosed cost in the GAO’s placement table gives about $30.2 million across four funding sources, ARPA, the Navy, NASA and the Strategic Defense Initiative Organization. That is a floor rather than a total, because several rows are grouped and some show no figure at all, and it covers procurement rather than research funding. A figure of $55 million circulates widely; it traces to a single magazine article citing a 1991 newspaper investigation, and it should be attributed rather than asserted.

The audit also has a blind spot, which Hillis pointed out to us. It “covered only non-classified U.S. government sales,” he wrote. Many more Connection Machines, he says, went to the oil industry and to customers abroad, mostly in Europe, Japan and Australia, and a number of highly classified purchases never appeared in the report.

So the GAO’s 24 machines are one slice of the unclassified government business, not the company’s customer list. Our earlier version leaned on them too hard, and Hillis called the question of who the customers were the part of the article that was least accurate.

What the inquiry actually found

One founding detail explains a great deal. Among the co-founders was Marvin Denicoff, who until August 1983 had directed the Office of Naval Research’s Information Sciences Program. He is recorded as a co-founder in 1983 and as having directed that programme until August 1983. Nothing in the audit connects him to any particular procurement decision, none of the sources alleges he did anything improper, and we make no such suggestion here.

It is often implied that an investigation caught ARPA and Thinking Machines doing something improper. It did not. The GAO made three headline management recommendations, that the placement programme be broadened to more vendors, that a database of project status be published, and that more emphasis go on system software. There was no finding of illegality, fraud or breach of procurement regulations, and ARPA officials generally agreed improvements could be made.

The GAO also noticed something the company would rather it had not. The CM-2 and CM-200 machines being placed were, in its words, essentially refinements of the original CM-1 design rather than significant departures from it. A programme meant to get novel prototypes into researchers’ hands was buying updated versions of the same computer.

The CM-5, and the switch to standard chips

The CM-5 was announced in October 1991 and it abandoned almost everything that made the Connection Machine distinctive. The custom one-bit processors went, replaced by ordinary SPARC microprocessors of the kind Sun put in workstations. The hypercube went, replaced by a fat tree. The operating system was a version of SunOS.

Thinking Machines did not present this as a retreat, and Hillis now says it was not one. Its own technical summary calls the CM-5 a universal architecture combining the best features of existing designs, able to run nodes in lockstep or independently as the problem required. The marketing answer was that it was neither one thing nor the other but both.

Our earlier version read the switch as the surrender of a company that had bet everything on custom silicon, and Hillis says that is one of the things the article “gets a bit wrong.” Using standard microprocessors “was actually always the plan,” he told us. They “just weren’t small and self-contained enough for the first machines, so we had to build our own, which I think were the first multi-core chips.”

What he did resist for too long, by his own account, was high-precision floating-point arithmetic, which he considered unnecessary for AI. “Eventually, I had to accept that AI was not our near-term market,” he wrote.

The technical summary names the topologies the CM-5 left behind, listing the hypercube and the two-dimensional mesh used in the CM-1 and CM-2 as the designs its fat tree improved upon. The processors were now standard SPARC parts, which Hillis says is what he had planned from the start.

A node could be fitted with four vector units alongside its SPARC chip. By the November 1993 technical summary those gave 160 megaflops of peak 64-bit floating point and 640 megabytes per second of memory bandwidth, up from the 128 and 512 quoted at the October 1991 announcement. The vector units could not fetch their own instructions; the SPARC issued to them. The technical summary calls a single node with vector units a supercomputer in itself, which was true and also the problem, because so was a rack of workstations.

Fourteen months

In June 1993 the first TOP500 list appeared. A CM-5 with 1,024 nodes at Los Alamos was first in the world at 59.7 gigaflops measured. A 544-node machine at the Minnesota Supercomputer Center was second, a 512-node machine at the National Center for Supercomputing Applications third, and FROSTBURG at the National Security Agency fourth. A 256-node CM-5 at the Naval Research Laboratory was seventh.

Thinking Machines filed a voluntary Chapter 11 petition on 17 August 1994, in the United States Bankruptcy Court for the District of Massachusetts. When the resulting lease dispute reached the First Circuit the following year, Judge Bruce Selya noted drily that the environs of the company’s Cambridge building were not sufficiently conducive to fertile thought.

The numbers behind the filing contradict the account usually given. Computergram International, reporting on the day of the filing, had the company losing $20 million in 1993 on sales of $82 million, and last making a profit in 1990, when it earned $1 million on sales of $65 million. Peak revenue was therefore 1993, the year of the record loss, not 1990. The same report gives 425 employees, about 140 of whom went at the filing, 112 machines in the field, and six months of unpaid rent on the building.

The double block

The numbers show the cash running out, and Hillis has now told us why it ran out when it did. The company was selling so many machines abroad, he wrote, that Cray Research persuaded the US government to cap how powerful a supercomputer Thinking Machines could export. The limit, he says, “was set so that Cray could export their most powerful machines, but we could not.”

Around the same time, Hillis says, Congress was lobbied into passing a law making upgrades to Pentagon supercomputers software-compatible with the machines already installed, “which just happened to be Cray’s.” He counts the export limit and the law as one “double-blocking operation,” and notes the irony that the fund for buying more Crays was called the DoD High Performance Computing Modernization Program.

He wrote that the company was blindsided by both the redirection of funding to Cray and the export controls, “although in retrospect, we should have seen them coming.” Together, he says, they made the company miss its 1993 sales projections and start running out of cash just as it was preparing for its first sale of shares to the public.

“So we were successfully out maneuvered by the incumbent trying to keep people using their outdated technology,” Hillis wrote. “They knocked us out of the market, but it was not enough to save them.” Cray Research was itself sold to Silicon Graphics in 1996.

Why it failed, according to the people who watched

Before Hillis wrote to us, most accounts put the failure down to economics rather than to Cray Research, the rival Seymour Cray had founded. Gordon Bell put it most directly, saying the company had extrapolated from a few sales directly related to government placements that there was a market for huge processing machines, and that there never was a real market there. Lew Tucker, one of the company’s research directors, told Gary Taubes of Inc. that its charter had not been to look at a machine and work out the commercial profit, but to build an interesting machine.

Hillis’s letter cuts against Bell’s premise. He says oil companies, overseas customers and classified buyers took machines the audit never counted, and that a rival then shut the company out of top-end exports and Pentagon upgrades.

The GAO had named the deeper problem a year before the filing, reporting that while much progress had been made in hardware, software remained too primitive to make massively parallel systems useful. A 1990 assessment by Robert Schreiber, at the NASA Ames institute that ran the agency’s supercomputing research, was harsher still about the hardware, concluding that the floating point had been strapped on rather than designed in, and that without a full custom implementation future Connection Machines would not keep pace with microprocessor-based machines. Hillis told us that this criticism was valid.

Then there was the bet. In 1990 Gordon Bell wagered Danny Hillis that in the last quarter of 1995 the highest sustained performance would come from a machine with fewer than a hundred processors rather than one with more than a thousand. The judges were John Hennessy and David Patterson. By 1996 Hillis was saying that from a changed market standpoint they had both lost, and that the supercomputing market basically did not exist any more as a definable market.

Where the pieces went

The company reorganised in October 1994 as a software-only business and reportedly ran four profitable quarters in that form. The Connection Machine line was discontinued rather than sold, which is worth stating plainly because the usual account has Sun buying the hardware business in 1996. What Sun negotiated for, as Hillis explains below, was the people who built it.

What actually happened, according to trade reporting at the time, is that Sun hired Thinking Machines hardware engineers from late 1994, and then in late 1996 agreed to buy the company’s GlobalWorks assets for an undisclosed sum. GlobalWorks was software, a parallel computing environment that ran on clusters of Sun servers.

Our earlier version called this a company being picked over, and Hillis says the handover was more orderly than that. Both Cray and Sun wanted the team, he told us, but the long leases on so much office space in Kendall Square were then seen as a liability. So the company went through a bankruptcy negotiated with Sun to shed them.

Sun then took the people with the hardware and software skills it wanted, giving them one Sun stock option for every Thinking Machines option. Hillis says that “worked out very well for both Sun and for the people that joined.” Hillis sees one more irony in the leases, which he says would have become extremely valuable had Sun kept them, as other companies moved into Kendall Square.

The last valuable thing was Darwin, a parallel data-mining toolkit that American Express had been using to sift credit card transactions on a Connection Machine. Oracle bought it in 1999, for a price nobody disclosed. The patents went to a partnership formed under the plan of reorganisation, which then sued IBM. We could not establish how that case ended, so nothing should be read into it either way.

The people turned out to be the asset. Hillis went to the MIT Media Lab, and later to Disney. Others went to Sun and helped turn a workstation company into a web company. Brewster Kahle built the WAIS system at Thinking Machines and went on to found the Internet Archive. Hillis’s own observation is that the best investment anyone could have made would have been to back everybody in that building, whatever they did next.

Asked about the collapse twenty-two years later he said he had been a bad business person who made a lot of stupid mistakes, and that had he known a quarter as much then as he did later, he could have kept the company alive long enough for the web to arrive.

The Jurassic Park question, settled

Connection Machine cabinets stand in the control room in Jurassic Park, in place of the Cray specified in Michael Crichton’s novel. Our earlier version said nobody could tell whether they were working CM-5s or empty front panels wired for their lights, because the evidence on both sides was thin.

Hillis has settled it. “The CM-5s used for the filming of Jurassic Park had actual CM-5 light panels in them, but nothing much else,” he wrote, so the lights were real and the supercomputer behind them was not.

He cleared up a smaller film question too. The Thinking Machines laptop that the hacker asks for in Mission: Impossible “never existed,” he wrote, so the request on screen was for a product the company never made.

Questions and answers

The machines, in order

Specifications below come from the Thinking Machines technical summaries where one survives. The CM-200 is included for completeness, but no manufacturer documentation for it could be found, and the figures usually quoted for it are unsourced.