The technology world is divided into two, according to an insightful post by Dylan Patel and Daniel Nishball, researchers at SemiAnalysis. One faction, termed "GPU-poor," struggles with limited access to Nvidia’s cutting-edge AI chips, while the other, the "GPU-rich," enjoys a substantial advantage, having a plenitude of these chips at their disposal. Nvidia’s GPUs are key to training the most advanced AI models, and having a sufficient supply of these chips equates to having a significant edge over competitors. On the other hand, those with a limited supply are essentially starting from a position of disadvantage.
In the current tech landscape, mere mention of "AI" in earnings calls doesn’t cut it anymore. Now, possessing the right technical components, infrastructure and a strategic plan for deploying costly equipment have become the new table stakes. Being at the front of the line for Nvidia GPUs and maintaining a good relationship with the company are now critical to success. Among the "GPU-poor" are a number of startups and open-source researchers, as well as prominent AI firms like Hugging Face, Databricks, and Together. Meanwhile, the "GPU-rich" group is spearheaded by leading names such as Google, OpenAI, and Meta, who, as per SemiAnalysis, are likely to have the highest ratios of compute resources to researchers.
New Tech Divide: GPU-Rich vs GPU-Poor
In a recent analysis by esteemed research firm SemiAnalysis, the tech industry has been divided into two distinct categories: "GPU-rich" and "GPU-poor". This division is based on the availability and utilization of Nvidia’s latest AI chips, which are essential in training powerful AI models.
The Importance of Nvidia GPUs
Nvidia’s GPUs play a critical role in the tech world, providing a significant advantage to those who possess them. They are the powerhouse behind the most potent AI models. Companies that have a healthy supply of these chips have a significant head start over those that don’t. It’s no longer just about throwing around the term "AI" during earnings calls. Companies actually need to have the technology, infrastructure, and a smart plan to deploy the expensive equipment.
The GPU-Rich and GPU-Poor Groups
Dylan Patel and Daniel Nishball from SemiAnalysis have categorized the tech industry into the GPU-rich and GPU-poor groups. The GPU-poor category mainly consists of startups and open-source researchers who face a limited supply of GPUs. European startups and government-backed supercomputers like Jules Verne fall into this category, described by the SemiAnalysis duo as "completely uncompetitive". Renowned AI firms such as Hugging Face, Databricks, and Together also find themselves in the GPU-poor group.
There’s a smaller, intermediate group too, which has been heavily investing in purchasing GPUs from Nvidia but has not yet seen a return on their investments. This group includes Cohere, Saudi Arabia, and UAE.
On the other end of the spectrum, Patel and Nishball have identified a handful of firms with more than 20,000 A100 and/or H100 GPUs from Nvidia. The leaders in this GPU-rich group include OpenAI, Google, Anthropic, Inflection, Elon Musk’s X, and Meta.
The Top GPU-Rich Company
Google, according to SemiAnalysis, is the most compute-rich firm globally, powered by its "unbeatably efficient architecture". The internet behemoth, which transitioned to an AI-first company years ago, is preparing to launch its next mammoth AI model, Gemini. Patel and Nishball predict that Google’s pace of innovation will surpass GPT-4’s total pre-training FLOPS by 5x before the year ends, with a clear path to 20x by the end of next year.
This categorization of tech companies into GPU-rich and GPU-poor groups signifies a new divide in the tech industry. Nvidia’s GPUs have become a critical asset for tech firms, and their availability could determine a company’s potential to compete in the AI realm. As AI continues to evolve rapidly, companies will need to secure these valuable resources to stay competitive and relevant in this ever-growing field.