• Google and Amazon have ambitions to sell AI chips to customers.
• That could make things awkward with Nvidia, which the two tech giants also rely on.
• Taking on Nvidia won't be easy, but one analyst said the process is now "irreversible."

Two of Nvidia's biggest customers might be turning into its biggest threat.

For three years, Nvidia's stock has defied gravity on the premise that the AI industry needs its chips. On Wednesday, when Google and Amazon reported their Q1 earnings, both signaled ambitions to sell their own custom AI chips directly to customers.

So far, Google's TPUs and Amazon's Trainium chips have only been available through Google and Amazon's cloud services. Customers can pay to use them, but they don't own them.

Nvidia is the undisputed leader in AI chips right now. While that shows no sign of stopping anytime soon, recent remarks from Google and Amazon suggest they are ready to challenge Nvidia's throne.

Andy Jassy, the CEO of Amazon, laid down the gauntlet to Nvidia in his annual letter to shareholders earlier this month.

"Virtually all AI thus far has been done on Nvidia chips, but a new shift has started," Jassy wrote, adding that it's "quite possible" that Amazon could start selling its chips directly to customers.

He put a timeline on this plan on the company's Wednesday earnings call, saying "there's a good chance" Amazon will start offering full racks of Trainium chips beyond its own cloud "over the next couple of years."

Google gave an even stronger commitment — and a nearer timeline.

On Wednesday, Google CEO Sundar Pichai publicly said for the first time that the company plans to deliver TPU chips to a "select group of customers" in their own data centers this year, but said the "vast majority" of revenues from those sales won't be realized until 2027.

Pichai said there was a big opportunity for Google's semiconductor business, which would also help fund the next generation of chips. That flywheel could create a mammoth business for Google. Morgan Stanley said in a December research note that selling 500,000 TPU chips could add roughly $13 billion in revenue to Google's balance sheet in 2027.

The company said in its 10-Q filing that it had so far signed a "limited number of agreements" to supply TPUs to customers.

It's also where things could get awkward. Google and Amazon are also big customers of Nvidia. They purchase Nvidia's chips to lease to their customers in their own data centers. Amazon and Google have both said that they will continue to work with Nvidia.

Nvidia shares were down more thean 4% on Thursday. The company didn't immediately respond to a Business Insider request for comment.

'Concerned but not worried'

Breaking into Nvidia's chip dominance will not be easy, analysts told Business Insider.

"Nvidia should be concerned but not worried," said Alvin Nguyen, a senior analyst at Forrester. Nvidia has built a strong ecosystem of hardware, software, and support that has made it easy for customers to choose them, he said.

"Selling products is very different than access to them," he said, adding that Amazon and Google would need to provide services like education and support to enterprises looking to buy their chips.

Google and Amazon's chip racks are also "very bespoke and proprietary" and customized for their own respective data centers, said Patrick Moorhead, CEO and chief analyst of Moor Insights & Strategy. That poses a challenge for reaching mass adoption, he said.

Plus, the chip market isn't a zero-sum game. AI companies are increasingly diversifying and using chips from multiple suppliers simultaneously. OpenAI, for example, has chip deals with Nvidia, AMD, and Broadcom, among others.

However, custom silicon is becoming an "increasingly important part of the AI story" for Google and Amazon, Bernstein analyst Mark Shmulik wrote in a note on Thursday. Both companies are pitching their chips as being more cost-effective than Nvidia's.

That opportunity is also opening up as the needs of the AI industry shifts. Earlier this month, Google announced a new TPU chip specifically for inference — the process of running the models once they're trained — which is becoming more critical as more companies bring AI agents online.

Beatriz Valle, a senior analyst for enterprise technology & services at GlobalData, called the decision by Google and Amazon an "extraordinary move" that will diversify the chip sector — and reduce cloud providers' dependence on Nvidia chips.

"This process will take years but it is irreversible now," she said.

• Arm announced its own AI chip, the AGI CPU, and is partnering with OpenAI and Meta.
• The new Arm AGI CPU aims to address energy efficiency and memory constraints in AI data centers.
• Despite strong growth prospects, Arm faces competition from established players like Nvidia and AMD.

Arm has long run its business as an architect behind the scenes, designing chips that power almost all the world's smartphone and making money off royalties from the chips it designs for customers.

Now, Arm is changing it up by announcing its own AI chip, the Arm AGI CPU.

Arm CEO Rene Haas said Tuesday at a company conference that this massive pivot wasn't just an internal strategy shift—it was a direct plea from the world's most powerful AI giants. The company name-dropped OpenAI and Meta as major partners for this chip.

"The biggest reason we're doing this is that our partners have asked for it," Haas said Tuesday.

With energy constraints and memory shortages, the AI boom has created a massive bottleneck in data centers. Faced with this demand, Arm stepped up with an AI chip that it says is more energy-efficient. Arm says it sees a $1.5 trillion market opportunity as it moves into AI chips for cloud, edge, and physical AI.

Arm stock was up by more than 18% on Wednesday. Mizuho analysts wrote that they see "strong growth opportunities" for Arm in AI infrastructure and the automotive industry. Bank of America research analyst Vivek Arya wrote in a note to investors that the company's outlook could be "too ambitious."

Meta and OpenAI partner with Arm

Meta has been building out data centers at a massive scale to power its apps and its latest superintelligence ventures. Santosh Janardhan, head of infrastructure at Meta, said Tuesday onstage that its coming "Hyperion" cluster could draw 5 gigawatts of power, enough to power 50 towns the size of Palo Alto.

"If we met the performance, we couldn't get the power. If we got the power, we wouldn't get the performance," Janardhan said.

This sparked an engineering project within Meta, where engineers were "working 'round the clock" to port its systems to Arm in three months, said Paul Saab, a Meta engineer.

"I didn't even ask my boss here for permission to buy these machines or even start the project," Saab said onstage.

While Saab says he saw major performance benefits, at the time, there wasn't an Arm chip available to buy.

OpenAI faced a similar problem. Its compute demand has grown massively as it trains and runs its ChatGPT models, its AI coding tool Codex, and more.

"That is one of the most common things I hear inside OpenAI. I need more compute," Kevin Weil, vice president of OpenAI for science, said onstage, adding that it needed chips that were energy-efficient.

Arm said it expects this chip to generate $15 billion in revenue by fiscal 2031.

The chip market is 'getting very crowded'

Arm faces the risk that the CPU market is "getting very crowded," Arya wrote in his analyst note. Other competitors, such as AMD, Nvidia, and Intel, have more CPU products and more established customers. Notably, both Meta and OpenAI also work with AMD and Nvidia, which could leave "limited" opportunity for Arm's new CPU, Arya wrote.

"Moreover, the bigger AI grows, the more pressure ARM's smartphone/consumer markets would have from limited memory supplies," Arya wrote.

That said, the increasing demand has led many customers to turn to chip companies beyond Nvidia for their computing needs. Both Meta and OpenAI also work with Broadcom to build AI chips.

The rise of AI agents has also led to greater demand for inference, or how AI models draw conclusions and make predictions. While Nvidia's core AI chips, the GPUs, dominate AI training, CPUs like Arm's AGI CPU can also help with inference. Nvidia also recently made moves into this market.

Have a tip? Contact this reporter via email at rmchan@businessinsider.com, or Signal at rosal.13. Use a personal email address, a nonwork WiFi network, and a nonwork device; here's our guide to sharing information securely.