NVIDIA is under investigation by the French Competition Authority, which alleges that the company has indulged in anti-competitive practices. NVIDIA is a dominant player in the market of computing chips for artificial intelligence. The French CA views this dominance with suspicion, claiming that there is a risk of abuse by NVIDIA. There are two issues with this assertion. First, NVIDIA has reached this market position through early investment, and legitimate innovation targeted towards AI - which set them apart from their competitors. Second, there is plenty of market competition, with old competitors, big technology companies, and new entrants looking to displace the AI sector's dependence on NVIDIA. Antitrust authorities must ensure that they sufficiently separate anti-competitive behaviour and innovation. Overregulation can often discourage innovation, so it’s important to be cautious when regulating emerging technologies.
Graphics card maker NVIDIA is facing increasing antitrust scrutiny, with the French Competition Authority[1] and the US Department of Justice[2] (DOJ) announcing investigations into its alleged anti-competitive practices in July. This blog will not comment on the DOJ’s investigation because details of the investigation are not publicly available. Instead, it focuses on the French CA’s inquiry, which is primarily concerned with the AI sector’s growing reliance on NVIDIA for GPUs (graphics processing units) and CUDA (Compute Unified Device Architecture). CUDA is NVIDIA’s proprietary software that enables the use of GPUs for artificial intelligence (AI).
The French CA has not revealed specific charges that will form part of the investigation. However, its opinion on the competitive functioning of the AI sector reveals the authority’s likely approach. In the opinion, it claims that there is a risk of abuse by NVIDIA due to its market dominance in the sector of computer components for AI. Specifically, it identified potential risks of:
Price fixing: agreements between competitors to maintain a fixed price,
Supply restrictions: agreements between competitors to limit production levels,
Unfair contractual conditions: contract terms that create an imbalance in the rights and obligations of the parties involved, and
Discriminatory behavior: where the company applies different conditions on equivalent transactions with trading partners - placing them at a disadvantage.[3]
To ascertain these potential risks, the French CA has referenced a market study where 40 companies within the sector have been interviewed. The study acknowledges that none of these practices have materialised so far based on their interviews.[4]
NVIDIA’s market dominance stems from the AI industry's heavy reliance on its GPUs and proprietary software, CUDA. NVIDIA makes GPUs, which are specialised computer chips designed to handle multiple tasks simultaneously through a feature known as parallel processing. Parallel processing is useful for AI development, to break down complex AI-related tasks such as pattern recognition and understanding natural language.[5] However, GPUs were traditionally designed to render graphics and were not inherently compatible with AI development. To remedy this, NVIDIA launched CUDA in 2006, a software that allows developers to harness the parallel processing power of GPUs for general-purpose computing.[6] CUDA’s parallel processing functionality, paired with NVIDIA’s GPUs, revolutionised AI development by exponentially reducing the time it took to develop AI models.[7] This made NVIDIA the backbone of the AI industry,[8] with the company accounting for an 80% market share in AI chips in 2024.[9]
Unlike open-source software, which has its source code available for others to view, modify and distribute, CUDA is proprietary and controlled exclusively by NVIDIA. Consequently, other GPU manufacturers cannot use CUDA. Further, NVIDIA’s GPUs are fully compatible only with CUDA, limiting their interoperability with other software. The close integration between CUDA and NVIDIA’s GPUs played a significant role in cementing the company’s leading position in the AI sector.[10] The French CA flagged this bundling of chips with CUDA as potentially anti-competitive.
There are, however, two primary concerns with the French CA’s assertions. First, NVIDIA’s competitive advantage results from years of innovation and risk-taking. NVIDIA introduced CUDA at a time when AI and machine learning were primarily limited to academic research. There was initial market skepticism in 2008 which saw NVIDIA’s investment as a risky bet on an uncertain technology.[11] Despite the skepticism, the company continued to invest heavily in CUDA, spending $12 billion on research and development from 2006 to 2017.[12] NVIDIA’s research and development strategy allowed it to benefit from the AI boom.[13]
Moreover, competitors may have failed to see the opportunity that NVIDIA recognized. For instance, Intel, which was dominant in chip-making until 2020[14], had the resources and capability to make similar investments geared toward AI and GPUs but chose not to.[15] Intel believed that the CPU (Central Processing Unit), a computer chip that focuses on sequential computation, would handle AI processing more efficiently than a GPU.[16] However, the growth in GPU usage proved Intel wrong,[17] and it relented to begin offering competitive GPU alternatives such as its Gaudi chips.[18]
Second, the market for AI chips and software is marked by dynamic competition and innovation. A coalition of tech companies, including Qualcomm, Google, and Intel, are working together to develop alternatives to CUDA.[19] AMD released its ROCm software stack in 2016. It’s an open-source alternative to CUDA, which enables parallel processing for AI development on GPUs.[20] OpenAI in July 2021 released Triton, a new programming language that enables researchers without any experience in using CUDA to write code for GPUs.[21] Similarly, on the hardware front, established players like AMD and Intel are offering competitive options,[22] while Google developed custom chips for AI, which Apple chose over NVIDIA’s hardware for its AI projects.[23]
Additionally, newer market segments for AI chips are emerging in the AI sector, increasing opportunities for new players to enter the market. These segments are shifting the focus away from NVIDIA’s GPUs. For instance, AI models undergo a ‘training’ stage - where large datasets are algorithmically processed to develop capabilities like natural speech comprehension. However, several models have now progressed to the ‘inference’ stage, where they no longer require training. Instead, they need to process inputs and generate responses. This means they require less processing power, which reduces the need for high-end NVIDIA GPUs.[24] Specifically, smaller specialised AI chips are more suitable for the majority of inference-based use cases. An increasing number of AI startups, like Groq, are developing specialised chips that provide inference at a lower cost and more efficiently.[25]
As antitrust authorities continue their investigations, distinguishing between market advantages due to innovation, and market advantages due to anti-competitive behaviour will be crucial. While it’s important to prevent abuses of market power, regulators must also be careful not to penalize companies for succeeding through innovation. Overregulation could stifle innovation that drives progress in industries like AI. For instance, the UK’s ex-ante competition regulation, the Digital Markets, Competition, and Consumers Act (DMCC), imposes certain conduct requirements on digital companies that are designated as holding ‘Strategic Market Status’.[26] Research by the Computers & Communication Industry Association (CCIA) shows that the law’s obligations will cause firms to divert resources allocated for innovation to ensure compliance, resulting in innovation delays for users and an investment reduction of between 4 to 8 percent in digital services.[27]
The outcome of the French investigation could influence other regulators, including the Competition Commission of India (CCI). The CCI on 22nd April 2024 announced a market study to understand the potential competition issues’ in AI. The study will specifically focus on the advantages enjoyed by incumbent players in the AI sector.[28] In light of these developments, competition authorities must balance competition concerns and innovation to avoid hindering India’s emerging AI ecosystem.
—
[1] https://www.reuters.com/technology/french-competition-authority-confirms-investigation-into-nvidia-2024-07-15/
[2] https://www.theinformation.com/articles/nvidia-faces-doj-antitrust-probe-over-complaints-from-rivals
[3] https://www.autoritedelaconcurrence.fr/en/press-release/generative-artificial-intelligence-autorite-issues-its-opinion-competitive
[4] https://francedigitale.org/en/posts/report-generative-ai
[5] https://www.polymersearch.com/glossary/parallel-processing
[6] https://blogs.nvidia.com/blog/what-is-cuda-2/
[7] https://www.ibm.com/think/topics/parallel-computing
[8] https://www.cambridge.org/core/books/abs/programming-in-parallel-with-cuda/brief-history-of-cuda/4999381F4B8B11F949212B1CADB8CBD4
[9] https://www.nasdaq.com/articles/nvidia-dominating-artificial-intelligence-chip-market-apple-has-been-securing-supply
[10] https://analyticsindiamag.com/ai-origins-evolution/nvidias-ai-supremacy-is-all-about-cuda/
[11] https://www.newyorker.com/magazine/2023/12/04/how-jensen-huangs-nvidia-is-powering-the-ai-revolution
[12] https://www.aol.com/finance/going-nvidia-jensen-huang-high-113053626.html
[13] https://www.vox.com/business-and-finance/357250/nvidia-stock-shareholder-ai-investor-market-capitalization-bubble
[14] https://www.nbcnews.com/business/business-news/what-is-nvidia-what-do-they-make-ai-artificial-intelligence-rcna140171
[15] https://www.nytimes.com/2023/08/21/technology/nvidia-ai-chips-gpu.html
[16] https://www.tbsnews.net/tech/how-chip-giant-intel-spurned-openai-and-fell-behind-times-911821
[17] https://www.tomshardware.com/pc-components/gpus/intel-ceo-details-company-s-three-biggest-mistakes
[18] https://www.vox.com/technology/364745/intel-nvidia-ai-silicon-valley-layoffs-stock-semiconductor
[19] https://www.reuters.com/technology/behind-plot-break-nvidias-grip-ai-by-targeting-software-2024-03-25/
[20] https://analyticsindiamag.com/ai-origins-evolution/amds-rocm-is-ready-to-challenge-nvidias-cuda/
[21] https://analyticsindiamag.com/ai-origins-evolution/how-is-openais-triton-different-from-nvidia-cuda/
[22] https://www.cnbc.com/2024/06/03/amd-unveils-new-ai-chips-amid-rising-competition-with-nvidia-intel.html
[23] https://analyticsindiamag.com/ai-news-updates/apple-chooses-google-tpus-over-nvidia-gpus-to-build-apple-intelligence/
[24] https://www.wsj.com/tech/ai/how-a-shifting-ai-chip-market-will-shape-nvidias-future-f0c256b1
[25] https://fortune.com/2024/07/02/nvidia-competition-ai-chip-gpu-startups-analysts/
[26] https://bills.parliament.uk/bills/3453
[27] https://ccianet.org/news/2024/01/proposed-uk-tech-regulations-under-dmcc-could-cost-consumers-up-to-160-billion-new-research-finds/
[28] https://www.cci.gov.in/images/whatsnew/en/tendernotice-1-11713759672.pdf