Get ready: on Wednesday, Alphabet (ASX: GOOG) will heavily promote its own TPU chips. The company will release its third-quarter earnings tomorrow, and its self-developed AI chip, the Tensor Processing Unit (TPU), is certain to become a hot topic among Google executives and Wall Street analysts.
Several new signals this week indicate that customer demand for TPUs is heating up sharply. Earlier today, Google officially announced that Oracle will become the first major cloud provider outside the Google ecosystem to offer TPU computing power to customers in its own data centers. Not long before that, a reporter disclosed that OpenAI, a major customer of Nvidia AI chips, plans to purchase Google TPU chips.
The most weighty detail about TPU demand came from Anthropic's IPO prospectus released on Monday. The filing documented the AI company's huge financial commitment to Google chips: Anthropic plans to invest $114.5 billion in Google Cloud, though the document did not specify what proportion will be used for TPU computing power. The company also estimated that it may spend an additional $125 billion to $150 billion to purchase TPU computing power from Google and other suppliers.
In addition, Anthropic will pay Broadcom, which co-designed the TPU with Google, $161.2 billion for TPU computing power procurement. Historically, when TPUs were first created more than a decade ago, they were used only internally by Google. It was not until 2018 that Google allowed external customers to access the chip through Google Cloud.
Recently, Google has successively formed partnerships with Fluidstack and now Oracle to deploy TPUs in partners' data centers. Google only began recognizing revenue from selling TPUs to third-party data centers last quarter. However, Chief Financial Officer Anat Ashkenazi said on the last earnings call that only a small portion of revenue from TPU cooperation agreements was recognized in that period, and the vast majority of revenue from such agreements will be realized starting next year.
Analysts predict that Google will reap considerable gains by then. TD Cowen analyst John Blackledge said in a research note earlier this month that TPUs will generate $9.4 billion in sales revenue in the second half of this year, accounting for about 13% of Google Cloud's estimated $70 billion in revenue for the same period. Blackledge expects the TPU business to reach $138 billion by 2028, accounting for 34% of Google Cloud's estimated total revenue that year.
Although that scale is already quite substantial, there is still a considerable gap compared with AI chip leader Nvidia. Analysts generally predict that Nvidia's total revenue for the fiscal year ending January 2028 will reach $692 billion. Even so, Nvidia has already noticed the rapid rise of Google's TPU business.
Today brought further signs that Nvidia is striving to retain its top core customers. The chipmaker officially announced a deepened partnership with Meta, and Meta said it will further expand its deployment of Nvidia chips. Under the terms of the cooperation, Nvidia will contribute $3 billion to participate in Meta's nearly $100 billion stock offering announced this week. Earlier this year, this newspaper reported that Meta was also in talks to rent Google TPUs. Facing Google's incursion into its core territory, Nvidia will not sit idly by.
Comments