Seymour Cray Bet on Gallium Arsenide and Lost Cray Computer to Chapter 11

A farewell memo and a new company

In 1989 Seymour Cray left Cray Research, the company he had founded in 1972 and whose machines had set the standard for supercomputing. He had already stepped away from day-to-day management; the board had grown tired of his long development cycles and his insistence on building every machine from scratch. Rather than retire, he started Cray Computer Corporation in Colorado Springs with a small engineering team and a single goal: build a supercomputer using gallium arsenide chips instead of silicon.

The departure was less a rupture than a rearrangement of the same instincts. Cray had spent his career arguing, in effect, that a supercomputer is a system designed as a whole — logic, packaging, cooling, and software considered together — and that buying commodity parts and assembling them was a different activity wearing the same name. Cray Research had grown into a company that needed predictable revenue, service contracts, and a product cadence its customers could plan around. Cray wanted to keep chasing the leading edge one machine at a time. The new company was the institutional form of that disagreement.

The decision was not impulsive. Cray had been watching gallium arsenide for years. Electrons move through the material several times faster than through silicon, which promised clock speeds that silicon could not reach. If the physics held, a gallium arsenide machine could be smaller, faster and cheaper to operate than the silicon-based Cray Y-MP. Cray Computer would not sell many units. It needed only a handful of customers willing to pay millions for the fastest machine on earth.

That business model is worth pausing on, because everything that followed depends on it. A company selling a handful of machines per year is not really selling computers; it is selling a research program with a delivery date attached. Revenue arrives in lumps, tied to individual contracts, and each contract is large enough that losing one is a crisis. There is no installed base generating maintenance income, no midrange product smoothing the cash flow, no upgrade path that customers buy on a schedule. The margin for error is measured in months, not quarters.

The Cray-3 and the cooling problem

The Cray-3 was the company’s first product. It packed processors into a compact cabinet and ran them at a clock rate that silicon designs could not match. The machine also generated enormous heat. Gallium arsenide circuits were dense and power-hungry, and the packaging had to keep them cool enough to operate reliably. Cray’s team developed a liquid cooling system that pumped a fluorocarbon fluid through the modules, a design that was elegant on paper and difficult to manufacture at scale.

The cooling problem is the part of the story that gets least attention and explains the most. Speed in a machine of this kind is not primarily a logic-design problem; it is a heat-removal problem. Push a clock rate up and every gate in the machine dissipates more power in the same volume, and the density of the packaging determines whether that heat can leave fast enough to keep junctions within tolerance. Fluorocarbon immersion was the answer Cray’s team settled on because it let them pack modules tightly and still pull heat out of the interior of the cabinet rather than only its surfaces. The difficulty was not the concept but the plumbing: seals, pumps, fluid handling, and the assembly steps that turn a laboratory demonstration into something a customer’s operators can maintain. Each of those steps is a place where a schedule slips.

The technical ambition was real. The Cray-3 was one of the first commercial machines to use gallium arsenide logic at its core, and it achieved speeds that impressed the national laboratories. But the machine arrived late, and by the time it shipped, silicon-based competitors had closed much of the gap. The Cray-3’s price and its specialized software stack limited its appeal to customers who already trusted Cray’s architecture.

That last point deserves emphasis, because it is a mechanism rather than a complaint. A machine with a novel logic family is not just a faster box; it is a new target for every piece of software that has to run on it. Compilers, libraries, and the numerical codes that government labs depend on have to be brought over, tuned, and validated. Customers weighing a purchase are therefore not comparing clock rates alone. They are comparing a proven software environment against one that has to earn trust, and they are doing so while the proven environment keeps getting faster on a schedule they can predict.

One customer and a public listing

The Cray-3 found a home at the National Center for Atmospheric Research in Boulder, Colorado, which used it for weather and climate modeling. That was a prestigious win, but it was not enough to sustain a company. Cray Computer had gone public in 1989, raising money on the strength of Seymour Cray’s name and the promise of gallium arsenide. Public shareholders expected revenue growth, and the company had only a few machines to sell.

The mismatch between the listing and the product is the crux of the commercial failure. Public markets price a company on expectations of growth, and growth in this segment comes from shipping units. But a company building one machine at a time cannot ship units faster than it can design and manufacture them, and each design cycle consumes years. The capital raised at the listing bought time, not a business. When the first machine slipped, the money that had been raised to fund the second machine was already being spent keeping the first one alive.

The economics of supercomputing are brutal for a small vendor. Each machine takes years to design, costs tens of millions to develop and sells to a tiny pool of government labs, defense contractors and research universities. A single delayed product can wipe out a year of cash. Cray Computer had no mainstream business to fall back on, unlike Cray Research, which sold smaller systems and service contracts.

This is the asymmetry that decides which vendors survive. A larger company can absorb a delayed flagship because the rest of its product line keeps revenue coming in; the flagship becomes a marketing asset and a technology incubator rather than the sole source of survival. A small company has the opposite structure: the flagship is the company, and every delay is a direct draw on reserves. Cray Research could treat an ambitious machine as one bet among several. Cray Computer could treat it only as the bet.

The Cray-4 and the end of the line

The company’s second machine, the Cray-4, was designed to be faster and cheaper than the Cray-3. It used the same gallium arsenide approach but with a more manufacturable cooling design. Cray Computer built prototypes and ran them in-house, and the early results were promising. But the company ran out of money before it could turn the Cray-4 into a product line.

Note what the Cray-4 represented: not a new gamble but a correction. The second machine kept the technology bet and fixed the manufacturing problem, which is exactly the sequence a small vendor needs — prove the concept, then make it buildable. The prototypes existed and the results were good. What was missing was the one input that cannot be engineered: time paid for in cash. A design that is nearly finished is worth nothing to a creditor if the next payroll is due first.

In 1995 Cray Computer filed for Chapter 11 bankruptcy protection. The assets were sold, the Colorado Springs facility wound down, and Seymour Cray moved on. He died in a car accident in 1996, shortly after starting a new venture called SRC Computers. The Cray-4 never reached the market in volume, and the gallium arsenide bet that had animated the company was quietly abandoned by much of the industry as silicon manufacturing improved.

What the failure revealed

Cray Computer’s collapse was not a story of bad engineering. The Cray-3 and Cray-4 were technically advanced machines, and the gallium arsenide work pushed packaging and cooling designs that later systems would adopt. The failure was commercial. A company built around one brilliant engineer and one risky technology had no margin for delay. When the Cray-3 slipped, the company had no second product to sell, and when the Cray-4 needed more time, the cash was gone.

The episode also showed how quickly the economics of supercomputing changed. Silicon-based parallel machines from competitors used commodity processors and standard software, and they improved on a predictable schedule. Cray’s custom gallium arsenide circuits were faster per processor, but the industry was moving toward clusters of standard chips. The Cray name survived through Cray Research and later through SGI and Cray Inc., but Seymour Cray’s second company did not.

The distinction between faster per processor and faster in aggregate is the heart of it. A custom circuit can win a single-thread comparison and still lose the workload, because a machine built from standard chips can add processors and ride the manufacturing curve that the whole industry is paying to advance. The custom designer has to beat that curve with engineering skill alone, on every generation, with no shared cost base. That is a race a small team can win occasionally and cannot win indefinitely.

What this means for the reader

The lesson is not that ambition is punished. It is that technical advantage and commercial durability are separate problems, and solving the first does not solve the second. Anyone building a product around a genuine engineering edge should ask the same questions Cray Computer faced: what carries the company if the flagship slips, how many customers does the model require, and does the funding structure match the length of the design cycle? Cray had answers to the engineering questions and none to the commercial ones, and the engineering was good enough that the answers mattered.

Frequently asked questions

Why did Seymour Cray leave Cray Research? He wanted to build machines his own way, without the management layers and product cycles that had grown around Cray Research. He founded Cray Computer Corporation in 1989 to pursue gallium arsenide supercomputing.

What was the Cray-3? The Cray-3 was Cray Computer’s first gallium arsenide supercomputer. It used liquid cooling and achieved high clock speeds, but it shipped late and sold to only a small number of customers, including the National Center for Atmospheric Research.

When did Cray Computer file for bankruptcy? Cray Computer Corporation filed for Chapter 11 bankruptcy protection in 1995, after the Cray-4 failed to reach the market in time to save the company.

Was the gallium arsenide approach wrong? Not on the physics. The machines were fast and the packaging and cooling work influenced later systems. The approach lost commercially because silicon manufacturing kept improving on a predictable schedule, letting competitors close the gap with standard parts and standard software.

What happened to Seymour Cray after the bankruptcy? He moved on after Cray Computer’s assets were sold and started a new venture, SRC Computers. He died in a car accident in 1996.