Comment: Nvidia's dominance - complaining doesn't help
Yes, Nvidia dominates the GPU market thanks to CUDA, to the detriment of many customers. But the hammer of antitrust law is the wrong way to go, says Wolf.
![Nvidia logo on graphics card](https://heise.cloudimg.io/width/610/q85.png-lossy-85.webp-lossy-85.foil1/_www-heise-de_/imgs/18/4/6/3/8/7/3/6/shutterstock_1606529806-c4a9db08fcdc7e40.jpeg)
(Image: Konstantin Savusia/Shutterstock.com)
In the gold rush, it's not the gold diggers who get rich, but the manufacturers of shovels. The AI frenzy of the last two years has mainly benefited GPU manufacturer Nvidia. The AI gold diggers do not want to or cannot work with the competitors' shovels. Nvidia briefly became the most valuable company in the world, raking in 15 billion dollars in profits in the last three months alone, with incredible margins of around 70 percent. There is no doubt that the company is exploiting its quasi-monopoly to the full.
This has brought the competition authorities onto the scene. The French antitrust authority Autorité de la concurrence has now launched an official investigation, may decide to file a lawsuit and has identified the CUDA platform as the main cause of market dominance. And rightly so. Although Nvidia builds fabulous accelerators, it is the software interface that is without alternative for many developers and binds them to the platform. Nevertheless, it is the wrong way to hit this point with the hammer of competition law. After all, who can blame Nvidia for developing interfaces and ecosystems for its own GPUs so that they work perfectly?
Do you already know the free iX newsletter? Register now and don't miss anything on the monthly publication date: heise.de/s/NY1E The next issue will be about the cover topic of the August iX: fresh Rust tools for Linux.
GPUs for machine learning
There are several reasons for Nvidia's dominance. The company has simply laid the foundation with convincing graphics cards. It recognized early on how important GPUs and their interfaces are for machine learning. And when the boom started, CUDA was already there. However, another very significant reason is the complete failure of other market players to really help an open standard achieve a breakthrough. There is no shortage of attempts, from OpenCL, ROCm, oneAPI, Triton to the UXL Foundation under the umbrella of the Linux Foundation. These attempts will not be any more successful if Nvidia is fined.
However, the documents published by the French Autorité indicate that it does not (only) want to reach into the company's coffers, but has more in mind: it wants to use the Digital Markets Act (DMA) against the overpowering manufacturer. The authority would then have to convince a court that Nvidia is a central platform service and has a gatekeeper role. Extending the DMA in this way is adventurous. After all, Nvidia does not fit into the categories listed in the DMA. If it does succeed, the company could be forced to open up CUDA further and change its license conditions. But what would be the point?
Nvidia can do without Europe
At the moment, Nvidia could even afford to disappear from the European market. Global GPU demand far exceeds supply, and other countries also have beautiful data centers. This would be a disservice to Europe's competitiveness and France, of all places, is home to two of the few real European AI lighthouses, Mistral and Hugging Face.
But even if it doesn't come to that: Nvidia's intertwining of hardware and software development and its unlimited resources alone would continue to determine where things go - and competitors would be forced to scurry after it. They are unlikely to like that. AI framework and library providers, on the other hand, would have even less reason to consider other architectures. None of this will help: the dominance will continue until an open, competitive CUDA alternative has prevailed. With or without a cartel process.
This commentary is the editorial of the new iX 8/2024, which will be published in the heise Shop on July 25 and on newsstands on July 26.
(ulw)