Thinking Machines: Topping the TOP500, Then Bankrupt in 14 Months

Sweeping the top four spots on the global supercomputing rankings could not withstand a market with only a single buyer.

In June 1993, the first global TOP500 supercomputer list was released. The four highest-scoring machines all bore the same trademark: CM-5.

Los Alamos National Laboratory's flagship system, with 1,024 nodes, achieved 131 GFLOPS to take the world number one spot. The second, third, and fourth places also came from this company.

The company was called Thinking Machines. In the summer of 1993, it stood at the pinnacle of the entire computational science world.

Fourteen months later, it filed for bankruptcy protection in a Delaware court.

A Doctoral Thesis About Wiring

The story began at the MIT AI Lab. Founder Danny Hillis was a PhD student of AI pioneer Marvin Minsky. At the time, the computer industry followed Seymour Cray's philosophy: make a single central processor bigger, more precise, and faster, using expensive liquid cooling or fluorocarbon immersion to squeeze out extreme clock frequencies.

But Hillis was troubled by another physical reality: in the von Neumann architecture, only a narrow bus connects processor and memory. No matter how fast the chip computes, massive data queues up waiting for the bus to move it.

His idea sounded like madness at the time: don't make one processor bigger; shatter it into thousands. Use over 65,000 ultra-simple, cheap processors working simultaneously, each glued directly to its own local memory. This dissertation on massively parallel computing architecture won the 1985 ACM Doctoral Dissertation Award.

In 1983, Hillis and Sheryl Handler founded Thinking Machines in Waltham, Massachusetts, moving the next year to Kendall Square in Cambridge, right across from MIT.

The Physicist Who Calculated Buffers

The first-generation Connection Machine CM-1 had an extreme, almost obsessive specification: the entire machine packed 65,536 processors.

To fit tens of thousands of chips into one machine, Hillis reduced each processor's width to just 1 bit. It was a Single Instruction Multiple Data (SIMD) machine: tens of thousands of chips executed the exact same instruction from a single clock at the same moment, each manipulating its own 1 bit of data.

The single-core computation wasn't hard; the fatal bottleneck was the communication network between chips.

Directly wiring every pair of 65,000+ chips was impossible — copper wires would burst the chassis. Hillis's solution was a 12-dimensional hypercube network: each chip communicated only with 12 fixed neighbors, so data between any two chips took at most 12 hops.

But when the network congested, data piled up at nodes, causing deadlock.

Chip engineers calculated that to guarantee lossless routing in the hypercube, each chip needed 7 internal message buffers.

Laying out 7 buffers per chip made the die too large for the semiconductor lithography of the day to yield working parts.

Richard Feynman was at Caltech at the time, spending summers as a consultant at Kendall Square. Instead of computer simulation, he derived a set of partial differential equations to model the statistical probability distribution of address bits containing "1" in a high-dimensional hypercube under random traffic.

A few days later, Feynman walked into the design room and told the young engineers: 5 buffers are enough.

Richard Feynman, photographed in 1984. He consulted on the Connection Machine design and solved the processor-to-processor data buffer depth problem. Photo: Tamiko Thiel, CC BY-SA 3.0, via Wikimedia Commons

The team respun the masks with 5 buffers, and this time the chip taped out successfully; the CM-1 network lit up on first try.

Feynman later ported the logarithmic algorithms he had used in the Manhattan Project to the machine, replacing multiplication and division with shifts and adds. He even ran quantum chromodynamics Feynman diagrams on the CM-1, measuring speeds faster than a dedicated array computer at Caltech worth millions of dollars.

The Machine That Had to Be Seen

The CM-1's industrial design came from sculptor Tamiko Thiel.

It wasn't the typical beige metal cabinet of the era, but an assembly of eight matte-black cubes, each exactly 1.5 meters on a side, wrapped in dark translucent acrylic doors.

Through the glass doors you could see 4,096 red light-emitting diodes (LEDs).

Connection Machine CM-1, 1985, now in the Computer History Museum. LEDs on the cube surface show real-time processor status. Photo: Carlo Nardone, CC BY-SA 2.0, via Wikimedia Commons

Each LED was bound in real time to a small group of processors' running state. When the machine ran at full load, the entire black cube glowed with a rhythmic blood-red pulse in the dark, like an artificial brain breathing.

This visual impact instantly pierced the geek community. When Spielberg filmed Jurassic Park, he wheeled two rows of CM-5s with blinking red LEDs into the dinosaur island's genetic engineering control room. In Hollywood's eyes, that was what a supercomputer should look like.

But the design also produced a cold-humor side effect.

In theory, the optimal SIMD parallel algorithm keeps all nodes working simultaneously, so all red LEDs should stay solidly lit; any blinking means some nodes are idle waiting for data transfer.

Yet many university and lab directors who had paid millions for the machine loved nothing more than a wall of wildly flashing red lights. Engineers often had to insert dummy no-op instructions into the main loop just to make the cabinet's red lights blink .

Prosperity With Only One Buyer

Riding its surreal looks and extreme compute throughput, Thinking Machines enjoyed a brief golden age.

In weather forecasting, fluid dynamics, and three-dimensional nuclear weapon hydrodynamic simulation, the hypercube's parallel matrix math was unrivaled. The company turned profitable in 1989; by 1990 annual revenue hit $65 million , overtaking the veteran supercomputing giant Cray Research.

But beneath the boom lay a fatal fact: there simply was no civilian parallel computing market.

In the late 1980s, ordinary software developers struggled with single-threaded pointers and memory alignment; nobody knew how to write parallel code for 65,000+ cores. Taming the beast required custom *Lisp or CM-Fortran.

A Connection Machine cost $3–5 million . Wall Street investment banks and Detroit automakers one glance at the price tag and the obscure parallel compiler, and walked away.

Thinking Machines' only real cash register was the U.S. Defense Advanced Research Projects Agency (DARPA) and the Department of Energy's national laboratories.

In the late Cold War, the Strategic Defense Initiative and the Strategic Computing Initiative needed the most ferocious compute power to simulate ballistic trajectories and nuclear reactions; the Pentagon had an unlimited defense budget.

Thinking Machines built its entire commercial edifice on that single purchase order.

The Fall of the Berlin Wall and the Commodity Chip Squeeze

In 1991, two external storms hit simultaneously.

The first came from Washington. As the Soviet Union collapsed and the Cold War ended, U.S. federal defense budgets began shrinking sharply. At the same time, competitors Cray, nCUBE, and MasPar jointly protested to Congress, accusing DARPA of unfairly funneling federal research funds to a Massachusetts favorite.

Facing congressional scrutiny, DARPA quickly cut its exclusive procurement from Thinking Machines and tightened export controls on high-performance computers. A company with hundreds of top scientists and engineers was instantly cut off from its financial lifeline.

The second storm came from a deeper shift in computing architecture.

In 1991, Thinking Machines launched the CM-5. This generation made a compromise: it abandoned the pure 1-bit SIMD architecture and switched to commercial SPARC RISC chips made by Sun, while the network topology changed from the complex hypercube to the easier-to-wire fat tree.

Connection Machine CM-5, released 1991. Nodes switched to commercial SPARC processors, network structure changed to fat tree. Photo: brewbooks, CC BY-SA 2.0, via Wikimedia Commons

This effectively acknowledged one thing: the Moore's Law pace of commodity RISC chips had already left custom proprietary silicon far behind.

When a single general-purpose microprocessor doubled in speed every 18 months, businesses preferred buying dozens of cheap Sun workstations to build clusters, or simply buying Cray vector machines, rather than spending millions to maintain a black cube that required rewriting all their software.

In 1992, Thinking Machines plunged into massive losses, and co-founder Sheryl Handler was ousted by the board.

In August 1994, barely a year after sweeping the top four spots on the TOP500, Thinking Machines formally filed for bankruptcy.

A Monument Left in the Eighties

Post-bankruptcy assets were quickly divided.

Its hardware technology and patents were acquired by Sun Microsystems, becoming the underlying DNA of the later Sun Enterprise server family; its data mining team and software business eventually ended up with Oracle.

The brilliant minds who once crowded Kendall Square coffee shops scattered into the modern internet:

Greg Papadopoulos joined Sun, rising to Executive Vice President and CTO;

Brewster Kahle expanded the Wide Area Information Server (WAIS) concept he developed on the machine, later founding the legendary Internet Archive ;

A group of alumni founded data-processing software company Ab Initio, embedding themselves in Wall Street's high-frequency trading systems.

Thinking Machines was not wrong in its technical intuition. Three decades later, the GPU clusters the entire AI industry uses to train large models — from node-to-node interconnects to massive data parallelism — carry the same instincts born on Kendall Square.

But it died in a brutal commercial mismatch: it sold the most advanced geek art object to the government's only budget window, and when that window slammed shut with geopolitics, it discovered there was no commercial foundation beneath its feet.

Feynman could calculate that five buffers would sustain the data flow, but no physics formula can sustain a market with only a single buyer.