Meta Platforms, Inc. has signed a deal to rent Alphabet’s AI chips, known as tensor processing units, to develop new AI models, according to a person involved in the talks. The multi-year deal is worth billions of dollars, said a person who was briefed about it. Meta has also been talking to Google about buying TPUs for its data centers as soon as next year, though the status of that discussion couldn’t be learned.
The deal is a victory for Google, giving it another brand-name customer that could help its efforts to build a multibillion-dollar business selling its TPUs. At the same time, it poses a threat to Nvidia, which dominates the AI chip market and currently supplies Meta with the graphics processing units to develop its AI, a process known as training.
In addition to forging the Meta deal, Google has signed an agreement with an unidentified large investment firm to fund a joint venture that would lease TPUs to other customers, according to a person involved in that arrangement. Google is in talks with other investment firms to fund other such joint ventures.
Both moves show Google is accelerating efforts to compete directly with Nvidia in the AI chip business, including in the AI training market that Nvidia dominates. Google stands to generate billions in additional revenue from selling TPUs.
News of the Google-Meta deal comes days after Nvidia announced a new deal with Meta, in which the Facebook owner said it will buy millions of GPUs for its data centers in the coming years. At the time, the Nvidia deal raised questions about the future of Google’s talks with Meta, which The Information reported last year, but it isn’t clear that it had any impact. Another factor in the Meta-TPU deal could be Meta’s own struggles to develop and AI training chip, details of which The Information reported earlier Thursday.
Some leaders of the Google Cloud unit previously suggested internally that supercharging the TPU business could help the company grab as much as 10% of Nvidia’s annual revenue—around $200 billion in the past 12 months, according to a person who heard the remarks.
Google is looking at a variety of ways to get TPUs into the hands of customers, according to a person who has been involved in its TPU strategy. That’s why the company is in talks with private equity firms about launching joint ventures that would buy its TPUs and then lease them to AI customers, according to two people with direct knowledge of these conversations.
The joint ventures could also start cloud businesses and be responsible for operating the TPUs. Nvidia has taken similar steps to seed customers for its chips, including funding a wave of upstart cloud providers, known as neoclouds, that focus on renting out its GPUs to AI customers.
Google has signed at least one term sheet with a large investment firm as part of this effort, one of these people said. A small group led by Google Cloud veteran Benjamin Treynor Sloss is leading the TPU financing effort, working closely with cloud chief Thomas Kurian, one of the people said.
Google’s corporate development team is also meeting with potential finance partners to borrow money for special purpose vehicles that would buy TPUs and lease them to customers, according to someone involved in those conversations. That would look similar to how Elon Musk’s xAI has structured creative financing deals with venture capital firm Valor to get access to Nvidia GPUs. In the deals under discussion with Google, TPUs could be used as collateral for the debt, the person said.
Balancing Act
Managing the growth of the TPU business is a tricky balancing act for Google. While the tech giant is increasingly competing with Nvidia, its Google Cloud unit is also one of Nvidia’s biggest GPU customers. That’s because most AI developer customers want GPUs to develop their technology, and Google Cloud can’t afford not to offer Nvidia-powered servers. Google will need a continuous supply of Nvidia’s latest chips to stay competitive in the cloud market.
And the number of TPUs available to new customers is also a question mark. Google’s own AI team, which develops the Gemini chatbot and models, uses TPUs to develop its technology and must ensure it has enough supply to compete with the likes of OpenAI. Taiwan Semiconductor Manufacturing Co. produces both TPUs and Nvidia GPUs, meaning they effectively vie for capacity at its facilities.
Google for years has rented out TPUs to cloud customers who use them in Google Cloud data centers, but last year it began pitching some of those customers—including Meta and big financial institutions—on the idea of using TPUs in their own data centers.
One of the ways it is doing so is by pitching the cost benefit of TPUs, arguing they’re cheaper to use than pricey Nvidia chips. The high prices for state-of-the-art Nvidia chips have also made it difficult for other cloud providers, such as Oracle, to generate solid gross profit margins from renting out Nvidia chips to AI developers.
It’s not lost on Nvidia CEO Jensen Huang that two of the world’s best AI models were developed fully or partly using AI server chips made by Google rather than Nvidia GPUs.
Google’s TPU push also comes after some cloud providers and customers, including OpenAI and Meta, struggled last year to get Nvidia’s latest Blackwell AI chips up and running at the scale they wanted, partly due to technical glitches and other complexities involving the hardware.
Meta and other companies that directly purchase AI chips for their own data centers have long sought alternatives to Nvidia so they aren’t beholden to one supplier. Meta this week announced a large deal to buy AI chips from another Nvidia competitor, Advanced Micro Devices, though Meta would primarily use those chips to run its existing AI models, a process known as inference, rather than train new ones, said a person involved in the deal talks.
Indeed, the fact that Meta plans to use TPUs for AI training is notable because most analysts, doubting that anyone could compete technically with Nvidia on training, have said they believe the biggest opportunities for challenging Nvidia lie in using chips for inference, which doesn’t require large, interconnected clusters of servers.
Meta is also continuing to develop its own chips for AI inference to save on costs and diversify away from Nvidia’s chips.
Meta won’t be the first large customer of TPUs. Anthropic last year agreed to spend about $20 billion buying TPUs from Broadcom, which co-designs the chips with Google and oversees TSMC’s manufacturing of them. Anthropic plans to use the chips in data centers not run by Google, according to people who have spoken to Anthropic leaders.
TPU-Trained Models
Anthropic for years has used TPUs to both develop and run its Claude AI, and Google has invested billions of dollars in the startup. The Information last summer reported that Google was discussing a TPU deal with OpenAI, and in November it reported the deal talks between Meta and Google over TPUs.
Meanwhile, it’s not lost on Nvidia CEO Jensen Huang that two of the world’s best AI models, from Google and Anthropic, were developed fully or partly using AI server chips made by Google rather than Nvidia GPUs.
He has closely monitored Google’s technical progress with TPUs and has sprung into action to entice existing and potential TPU customers to make big commitments to Nvidia’s GPUs. Meta and Nvidia recently said they negotiated a new partnership, and last year Nvidia made an investment in Anthropic, and got a commitment from Anthropic to use its chips. Nvidia also has discussed making a $30 billion equity investment in OpenAI, The Information reported.
Those moves show that Huang and Nvidia are resisting Google’s efforts to expand its presence in the AI chip market.
Nvidia’s domination of the market for AI server chips has turned it into the world’s most valuable company, with a $4.8 trillion market capitalization. Its stranglehold on the market has also sent its revenues soaring, lifting the amount of cash it throws off—which it can in turn use to invest in customers such as OpenAI and Anthropic, as well as cloud providers like CoreWeave that rent out GPUs.
Comments