Transcription
Bill Busby stood in the basement of the Mesa Lab at the National Center for Atmospheric Research on July 11th, 1977, watching a 5 and 1/2 ton C-shaped machine being lowered through a service hatch into a refrigerated computing chamber outside Boulder, Colorado. Encar had just taken delivery of serial number three, the first Cray 1 ever sold to a paying customer. The price was $8.86 million, of which $7.9 million covered the machine itself, and almost a million more bought the disc drives. Busby, who ran the scientific computing division, had spent 2 years arguing for this purchase. The CDC 7600 it would replace, was failing four or five times every day. Cray's new machine would run for days without interruption. Around the base of that cabinet, padded benches hid the power supplies and the Freon refrigeration plumbing. Engineers sat on them to read printouts while the machine they were debugging hummed behind their shoulders.
This is the story of how a Wisconsin engineer named Seymour Cray built the fastest computer on Earth out of integrated circuits, copper plumbing, and an idea about geometry and how the industry that worshipped him let his machines die. You remember what it cost to use one? Every minute on a Cray was billed, every job submitted through a batch queue, every line of Fortran scrutinized by a compiler that would either vectorize your inner loop or refuse to and tell you why. If you wanted speed, you learned to write code that the machine could see as parallel. Eight elements at a time through a single pipeline, chained add and multiply instructions firing on every clock tick. The clock cycled every 12.5 nanoseconds, yielding a frequency of 80 megahertz. Modern phones run at 30 times that rate, but in 1977, nothing else on Earth ran physics at that speed.
The smell stayed with you. Cray 1 machines used Freon refrigeration. The Cray 2 used fluorinert immersion, an inert fluorocarbon liquid that the cooling system pumped through and around the circuit boards directly. You watched bubbles rise through clear plastic tubes if a board was running hot. Consoles for the supercomputer were never built into the supercomputer itself. A Data General Nova at first, then a Sun workstation acting as the front-end terminal. The Cray was a back-end engine. Programmers fed it through whatever machine sat at its front. The operating system was COS through the early years and UNICOS, Cray's flavor of Unix, from 1986 onward. Some sites at the Department of Energy ran their own variant called the Cray Time Sharing System, written in a Fortran dialect ported up from CDC machines. Operators learned to read kernel dumps no other vendor on Earth produced. If you needed help, you called Chippewa Falls. Operators wore lab coats and worked 2-hour shifts because the machine was that valuable. They knew which jobs to kill and which to let finish. If you had a deadline for a paper or a weapons review, you learned their names by heart.
Printouts piled up beside every terminal. Vectorization reports ran to 600 pages, tracing where the compiler had succeeded or given up. A loop that the Cray translator could not vectorize was a loop that would run no faster than on a CDC machine you had walked away from. The Cray was not a faster general-purpose computer. It was a faster computer for code that had been written with the machine's pipeline in mind, which meant you and your colleagues had to rewrite physics codes from the 1960s, line by line, to extract the performance you had paid $8 million for. That trade-off defined an industry for 20 years.
By 1991, Cray Research held an 80% share of the world's supercomputer market. The figure came from International Data Corporation, which tracked the industry year-by-year. There were not many machines to track. Total industry sales that year were close to a billion dollars spread across fewer than 200 deliveries from all vendors combined. A single high-end Cray YMP or C90 sold for $15 to $30 million. The customer list was small and specific. The Department of Energy Weapons Labs at Los Alamos, Livermore, and Sandia. The National Security Agency at Fort Meade, Encar in Boulder for atmospheric simulation, the UK Met Office in Bracknell, Oakridge National Laboratory in Tennessee, Boeing in Seattle for aerodynamic modeling, Exxon, Shell and Mobil for seismic processing on oil exploration data, General Motors and Ford for finite element crash analysis. That customer base concentrated revenue in ways that mattered. Approximately 31% of Cray's 1989 sales went to governments directly. Roughly 75% of total revenue between 1987 and 1989 came from new computer sales rather than maintenance or upgrades, which meant the company depended on a steady stream of brand new machine shipments to keep its books healthy. 61% of those shipments stayed in North America. The rest went to a handful of national laboratories in Europe and Japan and to oil and aerospace customers who could justify a Cray on the basis of one specific class of problem.
Performance metrics tell the same story. The Cray 1 peaked at 160 megaflops in 1977. Steve Chen's XMP, introduced in 1982, pushed peak performance to 800 megaflops across four processors. Three years later, the Cray 2 hit 1.9 gigaflops. By 1988, the YMP became the first commercial machine to sustain more than one gigaflop on real applications. C90 in 1991 doubled that. T90 in 1995 doubled it again. Each generation got faster. The slope of improvement, though, was being matched and then overtaken by something coming from a direction Cray Research had not chosen to defend. Despite that, the XMP and YMP together became the most commercially successful supercomputers Cray Research ever shipped, generating over a billion dollars in cumulative revenue across their lifetimes. They were the machines that paid for everything else the company was built around.
Cray's economic structure was unusual even for the computer industry. A Cray 1 in 1977 cost between $5 and $8 million depending on memory configuration. By 1991, the C90 sold for $15 to $30 million. Maintenance contracts ran several million a year. The supercomputer demanded its own electrical substation. Encar paid roughly $35,000 a month just to power and cool serial number three. Customers covered those costs because the machine paid for itself in a specific way. Boeing engineers running a finite element analysis of a wing structure on a Cray YMP could complete in 8 hours a simulation that would have taken a week on the next fastest machine. An oil company processing seismic data could turn around a survey before a competitor finished its first pass. Weapons designers at Livermore could iterate on a hydrodynamic model in a week instead of a quarter. Customers did not buy Cray because they were cheap. They bought them because the alternative was waiting, and waiting cost more than the machine.
Many customers leased rather than purchased outright. A monthly lease on a top-of-the-line Cray covered the hardware, the operating system, the compiler, and on-site engineering support from Chippewa Falls. The economics worked for an oil company that ran the machine flat out for 12 months of seismic processing and could justify the operating expense from a single field discovery. They worked less well for a research group whose grant funded only a quarter of that time. That economic logic rested on a particular assumption: that the only path to a sustained gigaflop was custom bipolar silicon, hand-routed wires under 4 feet long, immersion cooling, and a vector pipeline architected by Seymour Cray himself in Chippewa Falls. The assumption held from 1976 until roughly 1989. Cray Research built no shareholder defenses against the possibility that it might stop holding because nobody at Chippewa Falls or at the corporate headquarters in Eagan, Minnesota believed it could stop holding. The company's competitive moat was the physical impossibility of competing. No one else could build a machine like this, so no one else did.
By 1989, the assumption was already wrong. The people inside Cray Research who recognized that it was wrong were leaving the company. The standard explanation for what happened to Cray is the talk Eugene Brooks gave at the Supercomputing '89 conference in Reno, Nevada in November of that year. Brooks, a researcher at Lawrence Livermore, titled his presentation, "Attack of the Killer Micros." His argument was that commodity microprocessors from Intel and from the RISC vendors were closing the performance gap on vector machines so fast that the gap would vanish within a decade. Brooks was right about the trajectory. Within 8 years, Sandia and Intel together built a machine called ASCI Red that crossed the Teraflop threshold in December 1996 using 9,072 Pentium Processors. ASCI Red cost the Department of Energy a little over $50 million. A Cray T90 capable of one sustained teraflop would have cost the better part of a billion, and Cray did not have one to sell. The killer micros explanation is true. It is also incomplete. Microprocessor performance closed the flops per dollar gap, but that gap had been closing for a decade before 1989. And Cray Research had survived it. Something else was happening at the same time. And the something else is what actually killed the company.
The value layer of high-performance computing migrated from hardware to software during the early 1990s, and Cray did not own the software. Through the 1980s, a customer who wanted to run a large physics code at high speed had two choices: buy a Cray and write to its vector compiler, or wait. The compiler was the moat. Cray's Fortran translator was tuned by a team in Chippewa Falls that understood every cycle of the pipeline, and code optimized for it would run nowhere else as fast. By 1994, that moat was gone. Tom Sterling and Don Becker built the first Beowulf cluster at NASA Goddard in the summer of that year. 16 DX4 processors connected by channel-bonded Ethernet running Linux, executing scientific code through a new standard called the Message Passing Interface. MPI had been ratified the same year by a working group that included researchers from Argonne, Oakridge, and the major university supercomputing centers. The standard was free. Linux was free. The Beowulf approach meant a research group with $100,000 and a room could assemble a machine that ran a meaningful fraction of a Cray workload. What changed was not the speed of the hardware. Hardware speed had been improving on the same curve since the early 1980s. Portable scientific code, on the other hand, was new. The MPI standard, the Linux kernel, and the open-source compilers and numerical libraries that started to ship with them meant a Fortran or C program written for a Beowulf cluster in 1996 would run on a different Beowulf cluster in 1997 with no modifications. Running that same code on a Cray would have required rewriting it for the vector compiler. Customers had not stopped wanting speed. They had stopped wanting to be locked into one vendor's compiler to get it.
By 1995, university researchers and national laboratory groups were running production scientific code on Beowulf clusters of their own assembly. The migration was not driven by a vendor selling something better than a Cray. It was driven by users discovering they could buy nothing at all and assemble what they needed from parts. That is the structural shift. No proprietary supercomputer company could survive. Hardware competitors can be beaten on price or performance. Open standards cannot be beaten at all.
Then came the collapse of Cray's most concentrated customer segment. The Berlin Wall fell in November 1989. By December 1991, the Soviet Union had dissolved. With its disappearance, the Department of Energy Weapons program, the single largest source of Cray revenue for 15 years, entered a procurement freeze that lasted until the Accelerated Strategic Computing Initiative was announced in 1995. Cray Research lost its anchor customer at the moment when its compiler moat was dissolving and its hardware advantage was being arbitraged away by commodity Pentium and Alpha chips. Three failures arrived at once. The company had architected itself against only one.
The years between 1989 and 1996 played out like a slow, controlled demolition. Steve Chen, who had designed the XMP and built the team that became the commercial backbone of Cray Research, left the company in September 1987 after John Rollwagen, the chief executive, killed his next-generation multiprocessor project. Chen took 45 engineers with him and tried to build his own machine at Supercomputer Systems Incorporated in Eau Claire, Wisconsin, with $150 million of investment money from IBM. That company went bankrupt in 1993, leaving more than 300 employees jobless. Cray himself spun off the Cray-3 project to a new entity called Cray Computer Corporation in Colorado Springs in 1989, taking with him another large block of engineering talent. The new company's only launch customer was Lawrence Livermore, which canceled its order in 1991 after the Cray-3 failed to meet a demonstration deadline. Eventually, the machine ran in 1993 as a loaner at Encar. Cray Computer filed for Chapter 11 in March 1995.
Meanwhile, Cray Research itself was running out of room. The company posted a $14.86 million loss in 1992, even as its low-end YMPL line found 70 new customers. Market share measured by IDC dropped from 80% in 1991 to 74% in 1993 to 52% in 1996. Japanese competitors NEC and Fujitsu took the high end. Beowulf clusters and commodity MPP machines from Intel and Thinking Machines took the low end. Cray's own MPP entry, the T3D, shipped in 1993 using DEC Alpha processors. The company had finally accepted that custom silicon was a losing position. But the T3D and its successor, the T3E, ran into a market that no longer believed proprietary supercomputers were worth premium prices. Silicon Graphics announced its acquisition of Cray Research on February 26th, 1996, for $740 million. The deal closed in April. Cray Research's 1995 fiscal year had ended with a $226 million loss on $676 million in revenue. SGI's chief executive Ed McCracken talked about combining the two product lines into a unified scalable architecture. That plan never delivered. SGI sold the Cray business to Tera Computer Corporation in March 2000 for $35 million and 1 million shares, less than 5% of what SGI had paid 4 years earlier.
Seymour Cray was driving his Jeep Cherokee on Interstate 25 north of Colorado Springs on September 22nd, 1996, when another driver clipped a third vehicle while attempting to pass. Cray's Jeep rolled three times. He died of head injuries at Penrose Hospital on October 5th, 1996, one week after his 71st birthday. SRC Computers, the company he had founded 6 weeks earlier with five employees, continued operating without him.
What survives of Cray today exists in three layers. The brand is intact. Hewlett Packard Enterprise acquired Cray Incorporated in 2019 for $1.3 billion, and HPE Cray supercomputers occupy the top three positions of the Top 500 list as of June 2025. The Frontier system at Oak Ridge and HPE Cray EX machine broke the exaflop barrier in 2022. Even the Cray name has outlasted Seymour Cray by nearly 30 years. And the engineering culture in Chippewa Falls, where HPE still manufactures interconnect hardware for those machines, traces directly back to the original Chippewa Falls lab that built the Cray-1. Vector instructions came back. The architecture Seymour Cray pioneered in the early 1970s: single instructions operating on long arrays of data through a pipeline execution unit, was reborn as SIMD inside every modern Intel and AMD processor. SSE, AVX, and AVX-512 on Intel chips. The Neon instructions on ARM processors. The Tensor units inside GPUs from Nvidia and AMD. Every one of them is a vector unit by another name. Chip designers who killed Cray ended up borrowing his idea and embedding it in the silicon that replaced him.
The people went on. After Supercomputer Systems Incorporated collapsed, Steve Chen joined Sequent Computer Systems and later founded Galactic Computing in Shenzhen, China. Burton Smith, who built the Tera MTA architecture that became Cray Incorporated under the 2000 merger, joined Microsoft Research in 2005. Many of the T3E engineers who designed Cray's last great proprietary interconnect went on to design the NumaLink interconnect at SGI, which itself was absorbed into HPE in 2016. Linux now runs on the descendant Cray machines that were once the exclusive province of UNICOS.
What did not survive was the business model. Nobody in the supercomputer industry today builds custom silicon for a customer list of a hundred buyers. The machines are constructed from commodity parts at scales their designers never imagined. The benches around the Cray-1 were not seating in the conventional sense. They were power supplies wrapped in vinyl and foam, designed to hold the cabinet's auxiliary hardware and to give engineers somewhere to rest while a code went through the pipeline. Visit the Computer History Museum in Mountain View, California today, and you can still sit on one. The machine in front of you ran an entire generation of physics, weather, and weapons research. And the company that built it no longer exists in any independent form. If this kind of computing history is useful to you, subscribe.