Can anyone beat Nvidia in AI? Analysts say it's the wrong question

The question many are asking now is who can beat Nvidia, which currently controls 80% of the AI ​​chip market. However, several analysts have recently luck That's the wrong question, because, as one of them, Futurum Group CEO Daniel Neumann, put it, “There are no natural predators for Nvidia right now.”

That's because Nvidia's graphics processing units (GPUs), developed in 1999 for super-fast 3D graphics in PC video games, have turned out to be ideal for training the increasingly large generative AI models being developed by companies like OpenAI, Google, Meta, Anthropic and Cohere. That training now requires access to vast numbers of AI chips. For years now, Nvidia's GPUs have been considered the most powerful and most sought-after.

These don't come cheap: Training the best generative AI models requires tens of thousands of top-of-the-line GPUs, costing $30,000-40,000 each. For example, Elon Musk recently said that to be “special,” his company xAI's Grok 3 model would need to be trained on 100,000 of Nvidia's top-of-the-line GPUs, which would generate more than $3 billion in chip revenue for Nvidia.

Nvidia's success isn't just a product of its chips, but also of the software that makes them accessible and easy to use. Nvidia's software ecosystem has made it a go-to for a large cohort of AI-centric developers who have little incentive to switch. At the company's annual shareholder meeting last week, Nvidia CEO Jensen Huang called the company's software platform, called CUDA (Compute Unified Device Architecture), a “virtuous cycle.” More users means Nvidia can invest more in upgrading its ecosystem, which in turn attracts even more users.

In contrast, Nvidia's semiconductor rival AMD, which accounts for about 12% of the global GPU market, has competitive GPUs and is improving its software, Newman said, but while it could provide an alternative for companies that don't want to be tied to Nvidia, it doesn't have the existing user base of developers who find CUDA easy to use.

Moreover, major cloud service providers like Amazon's AWS, Microsoft Azure and Google Cloud each make their own chips, but they're not looking to replace Nvidia. Rather, they want to offer a range of AI chips to choose from in order to optimize their own data center infrastructure, keep prices down and sell their cloud services to the widest base of potential customers.

“NVIDIA has early momentum and once they establish a fast-growing market it's hard for others to catch up,” explained Jack Gold, an analyst at J. Gold Associates, adding that Nvidia has managed to build a unique ecosystem that sets it apart.

Matt Bryson, senior vice president of equity research at Wedbush, added that Nvidia's chips will be especially hard to replace for training large-scale AI models, where he explained most of today's spending on computing power is directed. “I don't see that dynamic changing any time soon,” he said.

But a growing number of AI chip startups, including Cerebras, SambaNova, Groq and now Etched, see an opportunity to grab a piece of Nvidia's AI chip business. These startups are focused on the specialized needs of AI companies, specifically something called “inference” — running data through already-trained AI models and getting the models to spit out information (every answer from ChatGPT requires inference, for example).

For example, last week Etched raised $120 million to develop a specialized chip designed specifically to run Transformer models, a type of AI model architecture used by OpenAI's ChatGPT, Google's Gemini, and Anthropic's Claude. Meanwhile, Groq, which focuses on running models at super-fast speeds, is reportedly raising new funding at a $2.5 billion valuation, and Cerebras has reportedly filed for a privately disclosed initial public offering just months after unveiling its latest chip that it claims can train AI models 10 times the size of GPT-4 and Gemini.

All of these startups will likely focus initially on smaller markets, such as offering more efficient, faster or cheaper chips for specific tasks. They could also focus on specialized chips for specific industries or AI-powered devices like PCs and smartphones. “The best strategy is to carve out a niche and not try to take over the world, which is what most are trying to do,” said Jim McGregor, principal analyst at Tirias Research.

So perhaps a more pertinent question is: How much market share can these startups grab alongside cloud providers and semiconductor giants like AMD and Intel?

That remains to be seen, since the market for chips to run AI models and inference is still very new, but at Nvidia's annual shareholder meeting last week, Huang told investors that the company plans to keep innovating and remain the gold standard for AI training chips, even as competitors try to take away Nvidia's market share.

He argued that when Nvidia launches its Blackwell systems later this year, they will surpass the performance of Nvidia's current H100 chips and further solidify its lead. “The Blackwell architecture platform will probably be our most successful product ever,” Huang said.

But in the long term, analysts see plenty of areas where competitors can thrive, especially when it comes to power usage, which is a big cost for companies training and running AI models. “I think the power issue could create competitive alternatives,” Bryson said. “GPU data centers are going to become increasingly power hungry as models get larger, and a solution that significantly reduces power requirements could get a lot of traction.”

Nvidia also has to watch out for antitrust troubles. Reuters reported on Monday that France's antitrust regulator will accuse Nvidia of “alleged anti-competitive conduct,” which would make it the first regulator to do so. In the U.S., the Department of Justice is currently leading the investigation into Nvidia. An antitrust case could slow Nvidia's momentum, leaving room for smaller rivals to close the gap.

Nvidia's competitors will surely be able to take advantage of any advantage they can get, because for now, at least, Nvidia appears to be in a favorable position.

“Certainly, you could make the argument that Nvidia is on its way to becoming almost worthless unless they inflict injury on themselves and the market shifts,” Neumann said, but he argued that the “huge” market for specialty chips certainly leaves room for many other players.

“If AI is as big a trend as we all believe it is, we're going to see a very healthy ecosystem of chipmakers and software creators born to solve application-specific problems,” he said.