The Game Chip and the Scientists
You run a company that designs graphics chips for personal computers. Most of your revenue comes from people who play video games. Your newest chip, about to launch, could also be made to run ordinary programs for scientists and engineers, if you pay for it.
Do you build general-purpose computing into every graphics chip you sell, including the cheap ones for gamers, and give away the software to program it?
The situation
It is late 2006. Your company designs graphics processors, the chips that draw three-dimensional images on a computer screen, and has them manufactured by a contract foundry. You sell them to makers of PCs and graphics cards. Most of your money comes from people who play games and want faster, more realistic graphics. Your competition is fierce: one main rival in graphics, which has just been acquired by a maker of general-purpose processors, and constant pressure from PC makers on price.
Over the past decade graphics chips have changed. They used to perform fixed steps in a drawing pipeline. Now they contain hundreds of small programmable arithmetic units, working in parallel, that run short programs to shade each pixel. Your next chip, about to launch, takes this further. Instead of separate units for different stages of drawing, it has a single large array of general-purpose processors that can be assigned to any task. In principle, those processors could run calculations that have nothing to do with graphics: physics simulations, financial models, medical imaging, signal processing.
What you know
Researchers have noticed. For several years university groups have been using graphics chips for scientific work, by an awkward trick: they disguise their data as images and their calculations as shading programs, then read the results back as if they were pixels. It works, and for some problems it is many times faster than a conventional processor. But it is so difficult that only specialists bother.
In 2004 a team at Stanford published a language that let programmers write parallel programs for graphics chips without pretending to draw anything. Its lead author has since joined your company, and a team has been building a programming model and toolkit around a version of the C language that most scientists and engineers already know.
To make the new chip run such programs well, you would need to add features that games do not use: ways for the processors to share memory and coordinate, support for general programs, and a software stack that you would maintain and give away. That means more transistors, more engineering and more support, on every chip that includes it. Your chips are already among the largest ever made, at about 680 million transistors for the flagship.
What you don’t know
You do not know whether there is a large market for computing on graphics chips. The researchers using them are a few thousand at most. You do not know whether programmers will adopt a toolkit tied to one company’s hardware, or wait for an open standard. You do not know whether the world’s largest processor maker, which has begun designing chips with many simple cores of its own, will make your approach irrelevant. And you do not know how long it would take for any new market to pay for the cost.
The voices in the room
The gaming business argues that every extra transistor makes your chips more expensive or less powerful for the customers who actually pay. If scientists want computing, sell them a special product at a special price. Others argue that a programming platform lives or dies by its install base: a researcher will try your toolkit only if the hardware is already in the PC on their desk, or in the gaming machine in their student’s bedroom. A dedicated product for a few laboratories will never build the community of programmers that makes a platform valuable. A third group says the company is not big enough to carry a new industry on its back, and should wait to see whether one emerges.
What happened
The company was Nvidia. Its GeForce 8800, launched on 8 November 2006, was its first chip built on the unified architecture and its first to support CUDA, a platform for general-purpose computing on graphics processors. A public beta of the CUDA toolkit followed in February 2007, free to download. Nvidia did not confine the capability to expensive products: CUDA ran on its ordinary consumer graphics chips as well. In mid-2007 it also launched Tesla, a line of computing cards with no display outputs, priced from about $1,499, aimed at scientific users.
The CUDA programming model drew on the Stanford work of Ian Buck, the lead author of the Brook language, who joined Nvidia in 2004 and led the effort with the architect John Nickolls and others.
The cost was real. Jensen Huang, Nvidia’s chief executive, later recalled that putting CUDA on every consumer GPU increased the cost of the chips so much that it consumed all of the company’s gross profit, and that investors did not value the move. His numbers for the fall in the company’s value are recollections, and the period was complicated by other events. In July 2008 Nvidia took a $196 million charge to cover failures of some notebook chips caused by weak packaging materials, and reported a quarterly loss. Then came the financial crisis. Nvidia’s share price fell 76 percent in calendar 2008. It is impossible to separate how much of that fall came from CUDA’s cost, how much from the defect and how much from the crash.
The competitive threats did not materialise as feared. An open standard for programming graphics chips, OpenCL, was published at the end of 2008 with support from Apple, AMD and others, but CUDA’s head start in tools, libraries and teaching kept most developers on Nvidia’s platform. Intel cancelled the first version of its many-core graphics chip, Larrabee, as a consumer product in December 2009.
The market that justified the bet arrived from an unexpected direction. In 2012 Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton at the University of Toronto trained a deep neural network for image recognition, later called AlexNet, on two Nvidia GTX 580 gaming cards using CUDA. It won the ImageNet competition with a top-five error of 15.3 percent, against 26.2 percent for the runner-up, and set off the modern era of deep learning. Because CUDA ran on cheap gaming hardware, students and researchers had been able to experiment for years without needing a supercomputer. In fiscal 2026 Nvidia’s revenue was $215.9 billion, of which $193.7 billion came from data centres, most of it from chips used to train and run AI models, programmed through CUDA.
Judging the decision apart from the outcome
The reasoning was coherent with what could be known. Researchers were already using graphics chips for computation, the raw throughput advantage was clear, and the main obstacle was ease of programming, which Nvidia was in a position to solve. The decisive insight was about two-sided markets: a programming platform needs an install base before developers will commit, and the cheapest way to create one was to ship the capability inside a product people were already buying for another reason. A dedicated scientific product line would have been more profitable per chip, and would have built a much smaller community.
The grade here is Sound rather than Strong because the bet depended on things that were not knowable and on the company’s ability to survive a long wait. No one in 2006 could foresee that deep learning would supply the market. The bet was justified by a general belief in parallel computing and a willingness to carry costs for years, made possible by a founder-led company with a strong gaming business underneath. That combination is not available to most organisations, and the same reasoning, applied by a weaker company, could have ended in failure. Our knowledge of the outcome makes the logic look more certain than it was; this is exactly the case where hindsight and survivorship inflate our sense of the decision’s quality.
What a good process would have weighed
The separate product line protected margins but would have kept the platform small. Waiting for a standard avoided cost but surrendered the chance to set the programming model, and when the standard arrived it proved weaker than the one Nvidia had built. Staying in graphics would have left Nvidia exposed to processor makers moving into its territory. The chosen path accepted a long, uncertain cost in exchange for owning the platform if a market emerged. A good process would also have asked how long the company could afford to wait, and what would have to be true, and by when, for the bet to be scaled back. The record does not show whether Nvidia set such limits.
Sources and limitations
The cost of CUDA to Nvidia’s margins and valuation is described mainly through the chief executive’s later recollection, not through figures the company disclosed at the time, and his numbers should be treated as approximate. The share-price decline of 2008 had several causes that cannot be separated. The account of CUDA’s origins comes partly from Nvidia’s own published history. The claim that cheap gaming hardware made academic experimentation possible is supported by AlexNet’s use of consumer cards but is the author’s interpretation of its wider importance. The assessment of process, including the absence of documented stopping rules, is the author’s judgement.
The reveal
Nvidia and CUDA, 2006–2008
Put the general-purpose computing features into every chip, from the cheapest to the most expensive, and give programmers a free toolkit in a familiar language.
Paid off
The company put its computing architecture into every graphics chip from late 2006 and released a free programming toolkit in 2007. The cost depressed margins during a period that also brought a costly product defect and the financial crisis. Six years later, researchers trained the network that launched modern deep learning on two of its gaming cards, and by fiscal 2026 its data-centre revenue was $193.7 billion.
Sound
Judged on what could be known at the time, separately from the outcome. See moral luck.
The forces pulling at the people deciding
-
Programmers would write software for the chips only if many people owned them, and buyers would value the capability only if useful software existed. Putting the hardware in every gaming PC was a way to build one side of the market before the other existed.
-
The gaming business was large, competitive and profitable. Every square millimetre of chip spent on computing for unknown customers was taken from a known market, which is the classic tension between exploiting what works and exploring what might.
-
The first programming model to win developers' habits, libraries and code would be very hard to dislodge later. That made being early valuable, and made the cost of being early worth bearing.
-
There was no data on how large general-purpose computing on graphics chips might become. The decision could not be settled by forecasting, only by judgement about what an install base might make possible.
-
This bet is famous because it worked. Many companies made costly platform bets that never found their market, and they are rarely studied.
Signals that were visible at the time
- University researchers were already contorting graphics chips into doing scientific calculations by disguising their data as images, which showed demand and also how hard the existing tools were to use.
- A Stanford doctoral project had shown a language that let programmers treat a graphics chip as a parallel processor, and its lead researcher had joined the company.
- Graphics chips had far more parallel arithmetic units than general processors, and their raw throughput was growing faster.
- The company's main graphics rival had just been bought by a processor maker, and the largest processor maker was preparing many-core designs of its own.
- Gaming customers were price-sensitive, and the extra circuits would raise the cost of every chip.
A capability that might become a platform has to be put where people already are, and paid for long before anyone can prove it was worth it.
Follow the thread
Connections
Where this comes from
Sources
- Ian Buck and others, “Brook for GPUs. Stream Computing on Graphics Hardware,” ACM Transactions on Graphics 23, no. 3 (2004), 777–786.
- “Inside the programming evolution of GPU computing,” NVIDIA Developer Blog and The Next Platform, October 2015. Ian Buck's account of Brook and the origins of CUDA.
- “Nvidia's CUDA. The End of the CPU?” Tom's Hardware, 2008. CUDA introduced with the GeForce 8800; beta SDK in early 2007.
- “Nvidia launches Tesla,” The Register, June 2007.
- NVIDIA, Form 10-Q for the quarter ended 27 July 2008. The $196.0 million charge for weak die and packaging material in notebook products.
- NVIDIA, press release on second-quarter fiscal 2009 results, August 2008.
- Macrotrends, NVIDIA stock price history. Share price change of −76 percent in calendar 2008.
- Jensen Huang, interview with Lex Fridman, transcript. Huang's recollection of the cost of CUDA and its effect on gross profit and market value.
- “Intel Larrabee cancelled,” bit-tech, December 2009.
- Alex Krizhevsky, Ilya Sutskever and Geoffrey E. Hinton, “ImageNet Classification with Deep Convolutional Neural Networks,” Advances in Neural Information Processing Systems 25 (2012).
- NVIDIA, “NVIDIA Announces Financial Results for Fourth Quarter and Fiscal 2026,” 25 February 2026.
- Tae Kim, The Nvidia Way (W. W. Norton, 2024).
Teaching this case? The teaching note has learning objectives, a 60-minute session plan, discussion questions and the key facts on one page.
Shrikant Kumar. “The Game Chip and the Scientists.” Decide case, Telos. https://shrikantk.com/decide/the-game-chip/