In recent weeks, some of the best-known players in artificial intelligence have intensified their efforts to gain dominance in the AI space. Anthropic, a startup backed by Google and Amazon, launched Claude 3 Opus. Google launched Gemini, but is back to the drawing board after facing complaints. Meta is spending billions to create AGI, or artificial general intelligence. Elon Musk, who owns X that offers an AI product called Grok, has sued OpenAI for betraying its original mission.
All these have sparked an equally intense war in the hardware layer of AI, specifically GPUs (graphic processing units). GPUs are specialized chips that provide the computing power needed for models such as ChatGPT, Claude and Gemini. Currently, the AI chips segment is dominated by Nvidia, with an 80% market share, according to Jon Peddie Research. It is followed by AMD and Intel—better known for CPUs (central processing units), which can handle a broad range of computing tasks but are less suited for AI.
A good portion of Meta’s spending on AI went into purchasing GPUs. It bought 150,000 of these GPUs in 2023, according to Omdia Research. Microsoft bought a similar number last year. The demand for GPUs is not limited to US-based tech companies. China is offering its startups GPU coupons to subsidize computing costs. Last year, India’s information technology minister said the country would build a GPU cluster as a part of its AI programme. The demand for AI chips has pushed the share prices of all three GPU manufacturers, especially Nvidia, which became the first semiconductor company to cross $2 trillion in market capitalization.
Secret sauce
Nvidia achieved its dominant position by a combination of luck and strategic moves. It developed GPUs for the gaming market, foreseeing correctly that the segment would grow big and demand specialized chips to render images. Once it found that GPUs were also suited for training large language models used by AI researchers, it doubled down on that market. Similarly, during the crypto boom, demand for its chips from crypto miners grew.
One reason behind such demand from different segments is that Nvidia had invested in a programming platform, Cuda, which allows developers to finetune its chips. Similarly, its acquisition of Mellanox Technologies, its biggest so far, gave it networking technology at a chip level, which as Nvidia CEO Jensen Huang puts it, “enabled the modern AI supercomputer”. The complete package gives Nvidia an edge. Two years ago, it wanted to buy British semiconductor company ARM, which could have made it even bigger, but it ran into regulatory issues.
Dry powder
Nvidia is keen to pursue this kind of scope and scale because the biggest threat to its dominance comes from large players. Tech giants have billions of dollars in cash, nurture AI ambitions, and see dependence on Nvidia as a weakness. Building their own chips can bring down costs. Last August, Google unveiled its full-fledged fifth-generation tensor processing unit designed to train and serve content from LLM. Similarly, Microsoft, a key customer of Nvidia, has been working on its AI chips, the Azure Maia AI Accelerator. Anthropic, which runs Claude, will use Amazon’s AI chips.
Nvidia is also facing increasing competition from its traditional rivals, Intel and AMD. AMD said its Instinct MI300X data centre GPU has more memory and performs better than Nvidia’s flagship H100 chip. Intel has also redrawn its strategy around AI, which includes Gaudi 3 as a rival to Nvidia’s H100.
New challengers
Technology incumbents are seen as the biggest threat to Nvidia’s dominance. Some like Amazon, Microsoft and Google run large data centres, and have the expertise or the cash to buy it. The GPU landscape is also seeing new players, with companies such as Graphcore and Cerebras having raised over $700 million each. Last December, Moneycontrol reported that Cerebras was in talks with the Indian government to help build the country’s compute infrastructure.
As the incumbents and challengers make the GPU market more competitive, there are also questions about how long the AI boom will last. The industry is overflowing with optimists, represented by Sam Altman, who is looking to raise $7 trillion to boost AI chip production. Some believe that even if demand for computing power needed for training goes down, there will still be demand for serving the content to users. Whichever way it goes, the GPU landscape promises change.
www.howindialives.com is a database and search engine for public data