Last week we covered NVIDIA: The Leather Jacket Monopoly — the company that now owns the entire AI GPU market, coined a term they did not invent, and acquired their primary 3D graphics rival for the price of a modest office building. That rival deserves its own entry.

3dfx Interactive.

3dfx did not survive. But for roughly four years — 1996 to 1999 — they were the most important graphics company in consumer computing. They owned 73% of the retail gaming GPU market. They invented SLI. They made games run in three dimensions when the rest of the industry was still arguing about the best way to fake it. They shipped a card so ahead of its time that it required a second card merely to perform 2D desktop operations, and consumers bought it anyway.

Then they bought a factory in Juarez, Mexico. Their partners left. NVIDIA hired their engineers, acquired their patents, renamed their invention, and sold it back to the industry at a premium. The company was liquidated for $112 million. The creditors disputed that price for thirteen years.

This is a story about how to win a technology market, and then lose it with a single acquisition announcement.

The SGI Refugees:

On August 24, 1994, three engineers from Silicon Graphics Inc. incorporated 3dfx Interactive in San Jose, California. Their names were Gary Tarolli (Chief Scientist), Scott Sellers (hardware), and Ross Smith (marketing). They were not subtle about their origin story. Their first chip was named the SST-1 — for Sellers, Smith, Tarolli — in that order. The chip was named after themselves.

The Supreme Leader notes that naming your chip after yourself is acceptable under two conditions: first, the chip must succeed; second, you must still be around when it does. 3dfx satisfied the first condition. They would not satisfy the second.

Their seed capital came from Gordie Campbell of TechFarm, who invested $5.5 million. Their fifth employee was Brian Hook, a programmer who designed the first version of the Glide API — the proprietary graphics interface that would simultaneously make 3dfx dominant and irrelevant. Hook wrote Glide v1, saw what was coming, and left 3dfx to take a job at id Software — the studio making the very games that would prove Glide was unnecessary. This is foreshadowing. Hook recognized it. Management apparently did not.

The Voodoo 1 and the Click:

In 1996, 3dfx shipped the Voodoo Graphics card — known universally as the Voodoo 1. It rendered 3D. It did not render 2D. It did not have a 2D output. It was a dedicated 3D accelerator only. To use it, you installed it alongside your existing 2D VGA card, connected a passthrough cable from the VGA card into the Voodoo card, and ran both simultaneously.

When software switched from 2D to 3D mode, a physical relay inside the card clicked — audibly, mechanically — toggling the signal from the 2D card to the Voodoo’s output. You could hear 3D acceleration begin. The click of the relay was the sound of a company announcing that the future had arrived and it would not be quiet about it.

The Republic of Derails considers this approach correct. Every other company in computing history hides hardware mode transitions behind silent driver calls and status LEDs. 3dfx announced theirs with a mechanical relay. When the Supreme Leader’s motorcade activates its jamming array, there is also a sound. This is the appropriate protocol. Silence suggests malfunction.

The Voodoo 1 sold at $199-$299 and transformed the gaming market. Games that ran on it — Quake, Tomb Raider, Turok — were demonstrably superior to the software-rendered versions on the same hardware. The install base grew faster than most gaming peripherals had grown in the industry’s history. 3dfx was not a GPU company that became a gaming company. They were the gaming company that happened to make GPUs.

Glide, Carmack, and the MiniGL Problem:

The Voodoo’s proprietary API — Glide — was a free SDK for developers. It was fast, well-documented, and accessible. It was also exclusively usable on 3dfx hardware, which meant that any game that shipped with a Glide renderer could only use that renderer if the player owned a Voodoo card.

John Carmack — id Software’s lead programmer and the man whose engines drove most of the high-profile 3D games of the era — refused to support Glide. He had been burned previously by Rendition’s proprietary Verite API on Quake, investing development time in a hardware-specific renderer that produced limited return. He did not intend to repeat the experience.

The irony was severe. 3dfx had shipped the hardware that made Quake look best. But Carmack would not write a Glide renderer for Quake because Carmack had a principled objection to proprietary APIs. So 3dfx built MiniGL — a translation layer that converted OpenGL calls into Glide commands — specifically to make their own card run Quake acceptably. The company that invented the API had to build a compatibility shim to their own API so the most important game of the era would run on their hardware.

Brian Hook, who wrote Glide v1, was now at id Software. He was on the other side of the negotiation. The man who designed the API was now working for the company that refused to use it.

This should have been a warning. It was not processed as one.

The Voodoo 2 and the Voluntary Purchase of Two Cards:

In 1998, 3dfx shipped the Voodoo 2. It improved performance substantially over the Voodoo 1. It also introduced a feature that would define 3dfx’s legacy and eventually become NVIDIA’s premium upsell for two console generations:

SLI — Scan-Line Interleave.

Two Voodoo 2 cards connected by a ribbon cable. Each card rendered alternating horizontal scan lines — card one renders odd lines, card two renders even lines. Output was merged at a scanout chip and sent to the display. The result was near-doubled performance in supported games.

Consumers bought two cards. Voluntarily. In 1998. For a single gaming PC. At approximately $300 per card. The marketing did not need to be aggressive. The performance delta was visible. The install base of dual-Voodoo 2 systems was large enough to be a market segment of its own.

By the end of 1998, 3dfx held 73% of the retail gaming GPU market. They were the bestselling retail gaming card for 12 consecutive months — all twelve months of 1998. Every major game had a Glide renderer. Every gaming publication recommended the Voodoo 2. The company had achieved a kind of cultural gravity that most technology companies never reach: their product was considered necessary by the people who cared most about the application it ran.

The Voodoo Card Lineup:

Card	Year	Notable Feature	Fatal Limitation
Voodoo Graphics (V1)	1996	First dominant consumer 3D accelerator; audible relay	3D-only; required separate 2D card
Voodoo 2	1998	Invented SLI (Scan-Line Interleave); dual-card configs	Still 16-bit color; add-on only
Voodoo 3	1999	First integrated 2D+3D from 3dfx; competitive at launch	16-bit color only; no hardware T&L
Voodoo 4 4500	2000	32-bit color support finally; single-chip	Arrived during the company’s death spiral
Voodoo 5 5500	June 2000	First 32-bit; two VSA-100 chips; SLI on one card	Required external 250W power brick; GeForce 2 GTS already faster
Voodoo 5 6000	Never shipped	Four VSA-100 chips; external 250W PSU; one board	~1,000 prototypes. No retail units. Prototype cards sell for $9,500-$12,600 today.

The Voodoo 5 6000 is the most extreme artifact in consumer GPU history. Four GPUs on a single board. An external 250W power supply required to run it — in the year 2000, for a consumer gaming card. Approximately one thousand prototypes were manufactured. Not one reached retail. Those prototypes now exchange hands among collectors for prices that exceed the original MSRP of an entire gaming PC of that era.

The Supreme Leader has presided over development programs that also did not result in shipped units. These are called development programs, not failures. The distinction is administrative. The difference between 3dfx and the Republic of Derails is that the Republic can invoice participating entities for prototype costs. 3dfx could not invoice their own creditors. This is why vertical integration requires leverage, not merely a factory.

The STB Acquisition: The Sound Before the Silence:

On December 14, 1998 — at the peak of 3dfx’s dominance — the company announced the acquisition of STB Systems, a Texas-based board manufacturer, for $141 million in stock.

The strategic logic was vertical integration. Instead of designing chips and licensing them to board partners like Diamond Multimedia and Creative Labs, 3dfx would manufacture their own cards, capture the retail margin, and control the full product stack.

The execution logic was a factory in Juarez, Mexico.

The moment the acquisition was announced, 3dfx’s OEM partners understood the situation with clarity. Diamond Multimedia, Creative Labs, and their peers had been building and selling 3dfx-powered boards. They were now competing against their own chip supplier. Their supplier was now also their direct retail competitor.

They called NVIDIA immediately.

NVIDIA at this point was competitive but not dominant. Their RIVA TNT2 was a capable chip. They were not 3dfx. But they were willing to sell chips to board partners without also selling boards at retail. This is called a value proposition. Diamond Multimedia listed. Creative Labs followed. Virtually every major board partner that had built the Voodoo channel defected to NVIDIA within a quarter.

STB’s management had reportedly warned 3dfx that the Juarez manufacturing operation could not produce cards to the quality standards that their OEM partners — the ones 3dfx had just made into competitors — required. 3dfx management proceeded with the acquisition regardless. The OEM partners who had given those warnings were no longer 3dfx’s customers. They were NVIDIA’s.

October 11, 1999: The GeForce 256:

On October 11, 1999, NVIDIA launched the GeForce 256. It included hardware Transform and Lighting (T&L) — the ability to offload geometry transformation and scene lighting calculations from the CPU to the GPU. NVIDIA coined the term “GPU” for this card specifically. They did not invent the term, as covered in The Leather Jacket Monopoly, but they owned the narrative.

The Voodoo 3 — 3dfx’s current product — had no hardware T&L. It also remained 16-bit color only, a limitation that had been acceptable in 1997 and was embarrassing in late 1999. The Voodoo 5, with 32-bit color and improved performance, was months away.

In 1998, 3dfx had 73% of the retail market. NVIDIA surpassed 3dfx in total GPU shipments for the first time in 1999. Not market share. Not revenue. Raw units shipped. The company that had been the bestselling gaming GPU for all twelve months of the prior year could no longer claim to be the largest GPU shipper in the industry.

The relay that had clicked so confidently in 1996 was no longer audible over the sound of the balance sheet.

The Collapse:

• November 16, 2000: 3dfx discontinues production of all graphics cards.
• December 15, 2000: 3dfx’s creditors initiate bankruptcy proceedings. On the same day, 3dfx agrees to sell its assets to NVIDIA.

The simultaneity is worth noting. The creditors filed and the asset sale was agreed on the same calendar day. The company did not navigate bankruptcy and then sell. It sold at the moment bankruptcy began. This is called efficient timing by people who were not the creditors.

NVIDIA’s payment structure:

Component	Amount
Cash	$70 million
NVIDIA stock (1 million shares)	~$27 million at the time
Loan forgiveness	$15 million
Total	~$112 million

3dfx’s creditors disputed this price for thirteen years, arguing the assets had been undervalued in the sale. They lost. The $112 million was the number.

NVIDIA acquired 3dfx’s full patent portfolio — including all SLI patents — and hired approximately 100 of 3dfx’s engineers. The Voodoo brand, the Glide API, the relay click, the ribbon cable, the Juarez factory, the creditor disputes — all of it absorbed into the company that had been the second-largest gaming GPU vendor the year before.

NVIDIA later renamed SLI from “Scan-Line Interleave” to “Scalable Link Interface.” They launched it as a premium consumer feature with the GeForce 6 series in 2004. They sold it back to the same gaming market that had already bought it voluntarily in 1998 under the original name, for two cards, connected by a ribbon cable. The patent was now theirs. The branding was new. The concept was unchanged.

The Verdict:

What 3dfx Did	What It Cost
Named chip after themselves	Only works if you ship
Invented SLI	NVIDIA owns the patent
Invented proprietary API (Glide)	Had to build a compatibility layer for their own hardware
Acquired STB Systems at peak market share	Turned every OEM partner into a competitor
Built factory in Juarez	OEM partners had warned them; they proceeded
Shipped Voodoo 5 5500 requiring 250W external power brick	GeForce 2 GTS was already faster at launch
Produced ~1,000 Voodoo 5 6000 prototypes	None shipped; prototypes now cost more than a new GPU workstation
Sold company for $112 million	Creditors disputed the price for thirteen years

3dfx’s sin is not incompetence. Their engineers were exceptional. The Voodoo architecture was genuinely innovative. The SST-1 chip worked. The Glide API was fast. SLI was real performance improvement that consumers paid for willingly. Brian Hook wrote a clean API on the first attempt, recognized the strategic problem with it, and left before it became a liability. These are not the actions of an incompetent organization.

3dfx’s sin is the STB acquisition. One decision — made at maximum leverage, against the advice of the partners most directly affected by it, for the stated goal of capturing more margin — converted every OEM ally into a NVIDIA customer within a single quarter. The factory that was supposed to be a competitive advantage was a Juarez card plant that produced units their former partners would not accept in their systems.

The chip named after its own engineers shipped and succeeded. The company named after nothing in particular — some have suggested it was a sound the founders liked, others that it was arbitrary — was acquired for $112 million and dissolved.

The Lesson:

The Voodoo 1 clicked when it switched modes. That sound was correct. Every hardware mode transition should be announced. The industry disagreed and spent the next three decades hiding everything behind silent driver abstractions, firmware blobs, and status LEDs that nobody reads. 3dfx announced the future with a mechanical relay. They were right about that.

They were wrong about Juarez.

The Republic of Derails operates several manufacturing facilities. None of them are managed by our former distribution partners. The distinction between a chip company and a board company is not merely organizational — it is a declaration of intent toward your supply chain. When 3dfx acquired STB, they declared themselves competitors to their own partners. The partners responded rationally and immediately.

In the Republic, we do not acquire our distributors. We audit them. Annually. With forms they cannot appeal. This keeps the relationship productive and the supply chain loyal. 3dfx could have audited Diamond Multimedia. Instead they bought STB. One of these options costs $141 million and ends with creditors disputing the bankruptcy valuation for thirteen years.

The Voodoo 5 6000 never shipped. The Supreme Leader has released development programs that also did not result in shipped units. The difference is that the Republic’s development programs are eventually reclassified as strategic deterrents, which allows for regulatory invoice issuance. 3dfx’s Voodoo 5 6000 prototypes were reclassified as collector’s items worth $12,000. This is not the same thing, but the collectors appear satisfied.

The SST-1 chip name — Sellers, Smith, Tarolli — was vanity that worked. The company that bore their chip’s success did not survive to see the collectors paying $12,000 for the hardware they never shipped.

NVIDIA wore the leather jacket to every keynote after 1999. 3dfx made the technology possible. The click of the relay in 1996 started the GPU era. The creditors’ filing in 2000 ended 3dfx’s participation in it.

The relay clicked. Then it stopped.

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