Core Points

• Google launched its own chip development project in 2013, but it wasn't until the second quarter of this year that it began selling Tensor Processing Units (TPU) to external customers.
• Now, Google is leveraging Oracle's cloud platform as another channel to sell TPU, in an attempt to compete more fiercely in the AI chip market and against NVIDIA.
• Before joining Google, Thomas Kurian, the CEO of Google Cloud, worked at Oracle for over 20 years.

On April 9, 2019, at the Google Cloud Next conference in San Francisco, Thomas Kurian, the CEO of Google Cloud Business, delivered a speech.

For many years, Google has been promoting to its customers the advantages of "one-stop procurement of its own chips, large models, and cloud services," creating a complete technology stack. Amazon and Microsoft are also attempting to replicate this model.

And now, Google has chosen to extend its hardware beyond its own cloud ecosystem in order to reach more customers and expand the scale of its AI chip business. Currently, this market is dominated by NVIDIA.

Just one day before Alphabet released its third-quarter financial report, Google announced on Tuesday that Oracle Cloud would make available Google's Tensor Processing Unit ( TPU ) – a custom chip that Google has been developing for many years. This is the first time that customers can use TPU outside of Google's own infrastructure, and it also indicates that Curian, who has been in charge of Google Cloud for over twenty years at Oracle, is looking for new areas of business growth.

Check Chris Ballard, the Managing Director of Capital Management, said regarding Google's chip business: 'They are very strong, and it is entirely possible in the future that they could snatch market share from NVIDIA.'

Currihan, who joined Google at the end of 2018, now has a promising career ahead of him. Analysts expect that following a 82% growth in the second quarter, Google Cloud's revenue in the third quarter will increase by about 90% year-on-year, with the TPU business being an important driving force for this growth.

Uday Cheruvu, the portfolio manager of Harding Lovner Fund, which holds Alphabet shares, stated: "In the past, Google was good at developing technology behind closed doors, but not at serving external customers. Now, that mindset has changed."

Before Oracle officially announced its collaboration this week, Google made TPU available to the public in the second quarter of this year, allowing external data centers to use it.

There are significant differences in the market's forecasts for TPU's business, but all point to high growth. Cowen predicts that TPU's sales this second half of the year could reach $9.4 billion, and by 2027, they are expected to reach $62.5 billion; Wells Fargo estimates that TPU's revenue in the third quarter will be around $9 billion, with next year's revenue forecasted at $99 billion.

Chief Financial Officer Anat Ashkenazi stated during the July earnings conference call that only a very small portion of the TPU revenue generated by the existing agreements can be recognized this year, with "the majority of the revenue to be attributed to the fiscal year 2027." She also mentioned that while building its own infrastructure, Google plans to make greater use of external computing resources in the third quarter.

In the second quarter, Google Cloud's revenue was $25 billion, which is lower than Amazon Web Services' $42 billion and Microsoft Azure's $29 billion. However, Google Cloud has the highest growth rate among the three.

Parad said that as of June 30, its institution held approximately $236 million in Alphabet shares: "The larger the 'cake' of computing power becomes, the faster Google's expansion speed is."

"Optimize the entire full-stack system"

Curian hopes to expand the market share of TPU with external partners, but he also needs to weigh the benefits of increasing customer base against Google's own internal computing power requirements.

At a Goldman Sachs conference in September, Currihan stated that Google's AI servers have a payback period of less than two years; servers equipped with self-developed chips can halve that payback time. "This is thanks to our ability to deeply optimize the entire stack of technologies," Currihan said.

According to industry research firm SemiAnalysis, the total ownership and operating costs of Google's latest generation of TPU servers are approximately 44% lower than those of NVIDIA GB200 servers.

External sales of TPU will also have an impact on profit margins. TD Kohn estimates that the business operating margin is about 20%, while Google Cloud's overall operating margin in the last quarter was about 35%. Cheluwu believes that it is difficult for Google Cloud's operating margin to exceed the mid-35% range. Google has refused to comment on this.

Google's largest customer is also one of Google's most important AI competitors – Anthropic. Its Claude large model competes directly with Gemini, and the two parties have reached a cooperation worth tens of billions of dollars. Anthropic can call up to 1 million TPU at most. Mizuho said that this cooperation will release nearly 1 gigawatt of computing power this year, and described this deal as Google "arming its most direct model competitors".

OpenAI also plans to purchase Google TPU. The relevant news was first disclosed by The Information on Tuesday. This company, which develops ChatGPT, has long been a customer of NVIDIA.

Anthropic is currently seeking external financing to support the substantial expenses of purchasing Google chips. This week, Broadcom, which collaborates with Google on TPU business, announced a $63 billion financing facility funded by Blackstone and three major banks; this facility will purchase TPU and network hardware from Broadcom, which will then be leased to Anthropic for a period of five years. Analyst William Blair stated that this financing can only cover about half of Anthropic's computing power needs by 2027, indicating that there will be additional financing arrangements in the future.

Anthropic Submitted on Monday. IPO The filing states that the company simultaneously purchases chips from Google, Amazon, NVIDIA, and AMD. "We can select the chip products with the best cost-performance ratio and energy efficiency according to different business scenarios." Speaking of cooperation with Google and Broadcom, Anthropic said: "The cooperation allows us to access the computing power resources provided by multiple generations of TPU platforms."

Responsible Editor: Guo Mingyu