This week we have been examining the companies that built the physical and visual substrate of modern computing: 3dfx, which invented the consumer GPU and destroyed itself with a single acquisition; and the broader arc of how SGI’s refugees founded it. Today we examine the origin. Today we examine the cathedral.

Silicon Graphics Inc.

Silicon Graphics built a company around the idea that geometry was computation. Not text. Not spreadsheets. Not database records. The meaningful computational problem was taking a three-dimensional object — a dinosaur, a submarine, a spaceship — and transforming it into pixels at a rate the human eye could tolerate. They built custom silicon to do this. They built a proprietary operating system to manage it. They sold the result to Hollywood, to aerospace, to the government, and to anyone willing to spend $25,500 on a workstation that sat on a desk and glowed teal.

Then they gave away the core API that made the workstation worth owning, watched commodity graphics cards absorb their entire market, acquired Cray Research as a catastrophic distraction, and filed for bankruptcy on April Fools’ Day.

The Supreme Leader considers this a complete story. Every element is present. Technical excellence, genuine innovation, a moment of strategic generosity that became an existential error, management decisions of increasing implausibility, and a finale with perfect comedic timing. Silicon Graphics did not fail by accident. They failed with thoroughness.

I. The Geometry Engine

Jim Clark was a Stanford associate professor in 1982. He was not satisfied being a professor. He was building hardware. He had designed, with Marc Hannah, a chip called the Geometry Engine — the first VLSI implementation of a dedicated 3D vertex processor. It performed approximately 6 million operations per second. It was, in direct technical lineage, the ancestor of every GPU shipped between then and now.

Clark recruited seven graduate students. They incorporated Silicon Graphics in Mountain View, California. Their first commercial product shipped to Carnegie Mellon University’s Electronic Imaging Laboratory in 1984. Not to Hollywood. Not to the government. To a university laboratory, which is the correct first customer for technology that does not yet exist at commercial scale.

The Geometry Engine was not a graphics accelerator in the modern consumer sense. It was a dedicated vertex processor. It transformed 3D coordinates — the geometry of a scene — into the projection that a viewer would see from a specific position. This is the fundamental operation in all 3D rendering. Every 3D pipeline since has done this. The question is only whether the silicon dedicated to it is inside a $38,995 SGI workstation or a $249 consumer GPU running on commodity x86 hardware. SGI answered that question one way in 1984. The market answered it differently in 1999.

The co-founder who designed this chip with Jim Clark was Marc Hannah, who handled the VLSI architecture. Hannah was one of the original seven. The chip he co-designed begat the GPU industry. This is not a metaphor. The ArtX engineers who later founded their company, whose GPU lineage is now inside every AMD Radeon card sold today — they came from SGI. The chain of custody is direct and documented. We will return to this in the final section.

II. The Hollywood Credential

SGI’s workstations ran IRIX — their proprietary Unix, based on System V with BSD extensions, running on MIPS RISC processors. The software running on IRIX was what Hollywood needed, and IRIX was the only platform that could run it at sufficient speed.

The credential was built across three films in four years, by Industrial Light & Magic:

The Abyss (1989): the first photorealistic CG creature in a feature film — a pseudopod of water that moved with physical plausibility and reflected its environment. ILM rendered it on SGI workstations.

Terminator 2 (1991): the T-1000, a liquid metal figure that morphed, shattered, and reconstituted itself. Alias PowerAnimator running on SGI IRIX. The visual effects that made this film culturally significant were computed on Silicon Graphics hardware.

Jurassic Park (1993): ILM modeled and animated on SGI workstations using Alias and Softimage. Fully articulated dinosaurs. Wet skin. Motion through an environment. The sequence where the T-Rex attacks the tour vehicles in the rain — that rain interacted physically with a CG animal. This was not possible on anything else available in 1993.

The critics who called the Jurassic Park computer sequence “unrealistic” had an objection the Supreme Leader will now address permanently.

The Unix Scene:

In Jurassic Park, Lex Murphy — a child — sits at a Silicon Graphics workstation, navigates a three-dimensional rendering of a filesystem, and says: “It’s a UNIX system! I know this!” She then uses this system to lock the doors against the raptors.

The application she was using was fsn — the File System Navigator, a real IRIX application from Silicon Graphics that visualized a filesystem as a navigable 3D landscape. It existed. It shipped with IRIX. It rendered directories as towers of varying heights representing storage usage, which the user could fly through.

The critics who called this scene unrealistic were wrong. The Unix system was real. The IRIX application was real. The SGI workstation was real. The only fictional element was that a child could navigate it quickly enough to lock security doors before velociraptors arrived. The filmmakers used real SGI hardware on the set. The production designer put a real IRIX application in frame. The critics saw a 3D filesystem visualization in 1993 and concluded it was invented because it looked too sophisticated. It was not invented. It was just SGI’s filesystem tool, which looked like something from the future because SGI’s filesystem tools were.

Alias and Wavefront:

In 1995, SGI acquired Alias Research and Wavefront Technologies for $500 million, combined. The merger produced Alias|Wavefront. That company eventually became Maya. Maya is still the dominant 3D animation software in film production. Every animated feature, every visual effects film with significant CG, every game cinematic that reaches a certain budget level — Maya. The lineage from SGI’s 1995 acquisition to the tool that animated the most recent Avatar sequel is unbroken.

SGI spent $500 million on software that is still the industry standard thirty-one years later. As acquisitions go, this one was technically correct. The issue was the other acquisition. We will get there.

III. The Color Question

SGI’s industrial design was deliberately anti-beige. This is important enough to state clearly. In the early 1990s, every workstation from every competitor was beige. Off-white. The color of institutional surrender.

SGI chose differently. The Indigo was teal. The Indigo2 Impact was deep purple. The O2 was cobalt and steel blue. Each generation had a designed color identity. This was not accidental. SGI was selling into environments — film production companies, animation studios, research laboratories — where the machine was visible, where it sat in rooms with people who cared about visual artifacts, where a beige box was not adequate signaling for a $25,500 workstation.

The Supreme Leader notes that this approach is correct. Equipment should announce its purpose through its appearance. The Republic of Derails operates several categories of infrastructure where the hardware enclosure color indicates security classification. Beige is used for nothing. Beige is the color of a decision that has not been made.

SGI’s engineers did not merely build faster geometry engines. They built them in purple enclosures that sat on desks in Skywalker Ranch and told everyone in the room what they were. The color strategy was a functional choice dressed as aesthetics. When an animator is working at 02:00 and reaches for the machine producing the impossible image, that machine should not look like a filing cabinet.

IV. The Nintendo Footnote

In August 1993 — the same year Jurassic Park was in theaters on SGI-rendered dinosaurs — SGI and Nintendo announced Project Reality. The goal was a dedicated gaming system using SGI’s graphics architecture at consumer price points.

SGI designed the Reality Coprocessor (RCP) for what became the Nintendo 64 — a combined geometry processor and rasterizer running at 62.5 MHz. One of the first fully programmable GPU architectures at consumer scale. Not a fixed-function chip. Programmable.

The Nintendo 64 launched in Japan in June 1996. Every N64 cartridge ever manufactured — every game of Super Mario 64, Ocarina of Time, GoldenEye 007, and the rest of the catalog — ran on SGI’s graphics architecture at $199.99. The same company selling workstations at $38,995 to ILM was selling their graphics DNA inside children’s toys at the same time.

The Supreme Leader finds this an admirable two-channel strategy. The same core technology, priced and deployed across market segments separated by a factor of two hundred. SGI’s Geometry Engine DNA was in both the machine rendering liquid metal for ILM and the machine a twelve-year-old used to rescue Princess Zelda. This is correct industrial policy. The Republic of Derails has several technologies with similar dual-use deployment profiles, for which the pricing differential is considerably larger and the customers cannot be named.

V. The Gift That Killed Them

SGI’s internal graphics API was called IRIS GL — the IRIS Graphics Library, where IRIS was the name of their original workstation product line. IRIS GL was the software interface through which applications instructed SGI’s hardware to render geometry. It was fast. It was capable. It was tightly coupled to SGI’s hardware. It was the reason you needed SGI hardware if you wanted to run the software.

By the early 1990s, SGI faced a problem. Sun Microsystems, HP, and IBM were pushing a competing standard called PHIGS — Programmer’s Hierarchical Interactive Graphics Standard — through standards bodies. PHIGS was slower, more complex, and less capable than IRIS GL. But PHIGS was vendor-neutral, and Sun, HP, and IBM had enough institutional weight to make it the standard if SGI did not respond.

SGI’s response was to open IRIS GL. They removed the hardware-specific components, generalized the API, and published it as an open standard. On June 30, 1992, OpenGL 1.0 was released. Authored primarily by Mark Segal and Kurt Akeley — Akeley being one of the original seven co-founders. The people who built the Geometry Engine wrote the open standard that described how to use a geometry engine.

The strategic logic was sound: if SGI could not prevent a competing standard from emerging, it was better to control the standard than to be flanked by one. OpenGL becoming dominant over PHIGS was correct. SGI succeeded at this goal completely. OpenGL became the graphics standard for workstations, for industrial applications, for scientific visualization, and eventually for consumer 3D graphics on every platform.

The outcome was not what SGI intended.

OpenGL was portable. OpenGL ran on any hardware with a conformant implementation. OpenGL could run on the Voodoo1 — as covered in 3dfx: The Click of Doom. When NVIDIA shipped the GeForce 256 in October 1999 with hardware Transform and Lighting, it ran OpenGL. When every consumer 3D card shipped with an OpenGL driver, that card could run the same software that previously required a $38,995 SGI Octane. Not as fast. Not with the same quality at maximum settings. But sufficiently well for the majority of the applications that had justified SGI’s price premiums.

SGI open-sourced their core competitive weapon. The weapon worked exactly as designed. The weapon then equipped an army of $299 consumer cards and marched back toward SGI’s market. The strategic defense against PHIGS succeeded perfectly. The defense destroyed the business.

The Supreme Leader has studied military history extensively and can confirm that this pattern is not unique to the software industry. Arming an irregular force to counter a conventional threat, then watching the irregular force turn on you, is a recurring structure in recorded history. SGI is the 1992 software equivalent. The lesson is the same: when you distribute your most effective weapon to external parties for strategic purposes, the parties receiving the weapon have different long-term interests than you.

VI. The Price of Admission

To understand what SGI sold, you must understand what they charged. This is not incidental context. The price structure is the argument.

System	Year	Configuration	Price
SGI Indigo2 XL	1993	Entry workstation	$18,000
SGI Indigo2 XZ	1993	Mid-range	$25,500
SGI Indigo2 (high-end config)	1993	Maximum spec	~$40,000
SGI Octane (R10000, 250 MHz)	1996	Workstation	$38,995
SGI Onyx RealityEngine2	1993	Entry visualization system	$179,540

For $179,540 in 1993 — approximately $374,000 in 2026 currency, adjusted at 2.08 metres of purchasing power per dollar — you received an SGI Onyx RealityEngine2. This was the machine used for real-time simulation, flight training systems, and advanced visualization. It scaled. It ran IRIX with 64-bit support (IRIX 6.0 introduced 64-bit in 1994, years before x86 64-bit existed). It ran XFS — the high-performance journaling filesystem SGI had invented in 1993 and shipped in IRIX 5.3 — a filesystem so correct that SGI open-sourced it under the GPL in 1999 and it became the default filesystem in Red Hat Enterprise Linux, where it remains today.

And it could scale to 1,024 processors with a single system image. In the early 1990s. The Supreme Leader notes that running 1,024 processors under a single operating system image is still not trivial in 2026. SGI was doing this while the dominant consumer operating system was still figuring out how to handle more than one task at a time without crashing.

The Onyx was real. The price was real. The customers — Boeing, Lockheed, NASA, ILM, the United States military simulation programs — understood what they were buying. When a $374,000 machine is the reason your film has photoreal dinosaurs, the price is a production cost, not an obstacle. SGI’s market was not consumers. SGI’s market was institutions that needed the capability and could pay for it.

The market changed. The capability became available at $299. The institutions that had been paying $179,540 looked at the $299 card running the same API. They made adjustments.

VII. Jim Clark’s Exit

Jim Clark left SGI in late January 1994. The departure was not harmonious. Clark had grown frustrated with the management direction at a company he had founded. He departed with stock options and a conviction that the internet was about to become significant.

In February 1994 — thirty days after leaving the company he founded — Clark approached Marc Andreessen, who had built Mosaic at the National Center for Supercomputing Applications. Mosaic was the browser that made the World Wide Web visually navigable. Andreessen was twenty-two years old. Clark was fifty.

In April 1994, they incorporated what would become Netscape Communications Corporation — originally called Mosaic Communications Corporation until the University of Illinois objected. Clark invested $4 million of personal capital.

The Netscape IPO occurred on August 9, 1995. The company had no profits. It had been operating for sixteen months. The stock was priced at $14, opened at $75, and closed at $58.25 on the first day. It was the most significant IPO of that decade. It launched the dot-com era on Wall Street. Every subsequent internet IPO, every venture fund formation, every garage startup with a server and an idea — they were all operating downstream of the gravity well that Netscape’s August 1995 IPO created.

AOL acquired Netscape in 1999. Jim Clark exited with approximately $1.2 billion.

Return on a $4 million personal investment over five years: three hundred times the initial capital.

The Supreme Leader approves of this financial strategy unconditionally. Clark recognized that SGI had solved its problem — geometry computation for industrial customers — and that the next problem was navigation of a distributed network of hypertext documents. He moved. He moved within thirty days. He did not spend six months in transition meetings. He identified the next frontier, found the twenty-two-year-old who had already built the enabling software, wrote a $4 million check, and five years later exited with $1.2 billion.

This is correct behavior. The Supreme Leader has made similar moves. The details are classified. The multiple is different. The principle is identical: when the problem you founded a company to solve has been solved, the founder’s obligation is to find the next problem. Management’s obligation is to exploit the solved problem as long as the market permits. Clark was a founder. He found the next problem. SGI’s management was not Clark. They attempted to exploit the solved problem after commodity hardware had begun to depreciate it. These are different strategic postures for a reason.

VIII. The Collapse

In 1995, SGI’s market capitalization peaked at approximately $7 billion. They were the dominant supplier of high-performance 3D graphics workstations. They had Hollywood. They had government contracts. They had the Nintendo partnership. They had just acquired Alias and Wavefront for $500 million.

In 1996, SGI acquired Cray Research — the supercomputer company — for approximately $740 million.

The strategic logic was expansion into high-performance computing. SGI sold powerful workstations. Cray sold the most powerful single machines on earth. The markets were adjacent. The technical DNA was compatible. The acquisition would create a computing company with presence from the desktop to the top of the TOP500 list.

The execution was a catastrophe. Cray’s engineering culture, customer base, pricing model, and sales cycle had no meaningful overlap with SGI’s. Cray’s customers were national laboratories and governments with procurement processes measured in years. SGI’s customers were production companies and research institutions with procurement processes measured in quarters. The management overhead of integrating two companies with different operating tempos, different technical architectures, and different customer expectations consumed resources that SGI did not have to spare while consumer 3D hardware was beginning to eat their core market.

In 1999, SGI sold Cray Research at a loss to Tera Computer Company for approximately $35 million — having purchased it for $740 million three years prior. This is a $705 million loss on a single acquisition in three years. For context: this is the kind of loss that causes companies to be in different financial positions afterward.

The descent:

Year	Event
1995	Market cap peaks at $7 billion
1996	Cray Research acquired for ~$740 million
1999	Cray sold at a loss; consumer 3D hardware (3dfx, NVIDIA GeForce 256) fully commoditizing SGI’s market
November 2005	Delisted from NYSE; market cap ~$120 million
May 2006	Chapter 11 bankruptcy — first time
April 1, 2009	Chapter 11 bankruptcy — second time
May 11, 2009	Rackable Systems acquires SGI for $42.5 million
2009	Rackable Systems renames itself “Silicon Graphics International”
2016	Silicon Graphics International acquired by Hewlett Packard Enterprise for ~$275 million

The April 1, 2009 date requires acknowledgment. SGI filed for Chapter 11 bankruptcy protection for the second time on April Fools’ Day, 2009. The company that had built the most sophisticated 3D graphics systems in the world, that had rendered Jurassic Park and T2, that had designed the Nintendo 64’s graphics coprocessor, that had authored OpenGL, filed for bankruptcy on the one calendar day that makes every subsequent news headline read as a joke.

Whether this was deliberate comedy or administrative coincidence is unknown to the Supreme Leader. What is known is that the lawyers who filed the paperwork on that day were either unaware of the date or indifferent to it, and in either case the result was the same. A company that had peaked at $7 billion in market capitalization was acquired eighteen months later for $42.5 million by a server rack company called Rackable Systems, which then renamed itself Silicon Graphics International in an act of brand acquisition that cost less than a single Onyx RealityEngine2 had cost in 1993.

Hewlett Packard Enterprise acquired Silicon Graphics International in 2016 for approximately $275 million. This is more than the $42.5 million Rackable paid in 2009, less than the $7 billion SGI was worth in 1995, and irrelevant to the fate of everything SGI had actually built.

IX. The Lineage

SGI is dead. Its work is not.

In 1997, twenty SGI engineers departed to found ArtX, Inc. They took with them the design knowledge of SGI’s MIPS graphics pipeline — the same lineage as the Geometry Engine, the same technical tradition that had designed the N64’s RCP. ArtX designed the Flipper GPU for the Nintendo GameCube.

ATI Technologies acquired ArtX in 2000. The engineers who had left SGI were now inside ATI. Their GPU design expertise — originating from SGI’s Geometry Engine tradition — was now part of a company building consumer graphics cards.

AMD acquired ATI in 2006 for approximately $5.4 billion, as covered in AMD: The Underdog That Bit Back. The people who designed SGI’s MIPS graphics pipeline are directly in AMD’s GPU ancestry. Every Radeon GPU sold today carries this lineage. Every AMD APU’s integrated graphics unit carries this lineage. The chip design tradition that Jim Clark and Marc Hannah started in 1982 runs unbroken to the AMD Radeon cards in production in 2026.

XFS was invented by SGI in 1993, introduced in IRIX 5.3, open-sourced under the GPL in 1999, and ported to Linux. It is, today, the default filesystem in Red Hat Enterprise Linux. SGI’s filesystem invention runs on more servers than SGI ever sold. Compared to ZFS — discussed in ZFS: The Last Gift — XFS is the more widely deployed high-performance filesystem in enterprise Linux environments, by virtue of being the RHEL default. Whether it is the more technically ambitious filesystem is a different question. XFS is fast, mature, and scales to 8 exabytes. It does not checksum data. ZFS checksums data. These are different design philosophies, not competing claims of superiority.

OpenGL: the standard SGI authored in 1992 to prevent PHIGS from flanking them ran consumer graphics for a decade, ran workstation graphics for two decades, and was eventually succeeded by Vulkan and Metal for high-performance scenarios. OpenGL still ships in drivers for every GPU sold. The API that was supposed to save SGI’s market position by defeating PHIGS succeeded in that mission. It then commoditized the market. Both things are true.

Maya: the software that came from SGI’s $500 million acquisition of Alias and Wavefront in 1995 is still the primary 3D animation tool in film and game production. Alias|Wavefront became Autodesk Maya. If a feature film has a photoreal CG character in 2026, Maya was almost certainly involved. SGI’s $500 million spent on software is still generating returns — for Autodesk, not for SGI.

The Nintendo 64 is in attics. The N64’s RCP — SGI’s graphics DNA — ran on approximately 33 million units sold between 1996 and 2002. The GameCube, which used ArtX’s Flipper GPU (SGI lineage), sold approximately 21.7 million units. The GameCube’s graphical output, in 2001-2006, was routinely described as producing image quality above its price point. This is the SGI Geometry Engine tradition at $149.99. It was correct.

The Verdict

Silicon Graphics built the most expensive machine in the room. They built it better than anyone else. They open-sourced the reason to own it. They acquired a supercomputer company when they should have been defending their core market. They filed for bankruptcy twice — the second time on a date that sounds invented. They were acquired for $42.5 million by a company that immediately adopted their name.

The founding engineers scattered. The ArtX founders took the GPU design tradition to ATI, then to AMD. XFS ships in every RHEL server. Maya ships in every serious animation pipeline. OpenGL shipped in every GPU driver. The Nintendo 64 ran the Geometry Engine’s grandchild at $199.99.

SGI built the cathedral. They gave away the indulgences. The parishioners spent the indulgences at the 3dfx and NVIDIA store down the street. The cathedral closed. The theological framework it invented — OpenGL, the vertex shader, the hardware geometry pipeline — became the structure of an industry that did not have the name Silicon Graphics on any of its products.

Jim Clark left thirty days after his last management disagreement, invested $4 million in a twenty-two-year-old’s browser company, and exited with $1.2 billion. He then founded Healtheon. He then founded myCFO. He then became a person who owns very large yachts, which is a perfectly acceptable outcome for a Stanford associate professor who co-designed the first dedicated 3D vertex processor and recognized when to leave the building.

The building was purple. The machines inside it were teal, purple, and cobalt. They were the most powerful graphics workstations in the world, they rendered the dinosaurs that convinced audiences that CG was real, they ran a real UNIX system with a real 3D filesystem navigator that a real child used on screen in 1993, and they filed bankruptcy on April Fools’ Day.

Whether this was planned comedy or accidental tragedy is left as an exercise for the reader. The Supreme Leader does not exercise. The Supreme Leader observes. The observation here is that the Geometry Engine’s technical lineage is immortal, the business that built it is not, and the date chosen for the second bankruptcy filing suggests that somewhere in the management chain, someone had a sense of humor about it. Or no calendar. The outcome is the same.

The most powerful 3D graphics company in history commoditized itself, gave its API to the industry, and had its engineers found the companies that built the commodity hardware that replaced it. This is called a complete cycle. The Supreme Leader has presided over complete cycles in several industries. The lesson is always the same: technology propagates; institutions do not.

The institutions that do survive complete cycles are the ones that — unlike SGI — did not spend $740 million on Cray Research while the consumer market was beginning to render the same pixels for $299.

The purple machines are in museums now. The software they ran is still running.

T2SDE Linux — a distribution maintained by volunteers, supporting every exotic architecture that larger projects abandoned — shipped version 26.3 on March 4, 2026. The day before this decree. With SGI MIPS64 kernel variants for the Indy, Indigo2 R10000, Octane, O2, and Origin still present. Still compiling. Still booting on machines that cost $180,000 in 1995.

T2SDE has released a new version more recently than SGI filed its last bankruptcy. The Supreme Leader considers this the correct order of events. The company that built the machines is gone. The machines still run.

— Kim Jong Rails, Supreme Leader of the Republic of Derails