Google accelerates TPU deployment pace, competing with NVIDIA in the AI chip field

Google is one of the largest buyers of NVIDIA's AI chips and leases these chips to customers of Google Cloud, such as OpenAI and Meta Platforms. However, Google's ambition to develop its own AI chips has not diminished.

According to seven people involved in the negotiations, Google has recently approached several small cloud service providers that primarily lease NVIDIA chips, expressing a desire for their data centers to also deploy Google's AI chips.

Representatives from companies involved in the deal privately disclosed to the media that Google has reached an agreement with at least one cloud service provider, including Fluidstack, which is headquartered in London and will deploy Google's Tensor Processing Units (TPUs) in its New York data center.

Additionally, Google has attempted to reach similar agreements with other cloud service providers focused on NVIDIA chips, such as Crusoe, which is building a data center for OpenAI that will deploy a large number of NVIDIA chips, and CoreWeave, which leases NVIDIA chips to Microsoft and OpenAI.

The media reports that it is currently unclear why Google has chosen to deploy TPUs in the data centers of other cloud service providers for the first time. Analysts believe this may be because the pace of Google's own data center construction cannot keep up with the growing demand for chips, or it may be looking to find more new customers for its TPUs through other cloud service providers, such as AI application developers. This approach is similar to the model of cloud service providers leasing NVIDIA graphics cards.

Analysts suggest that if it is the latter case, Google's actions would mean a more direct competition with NVIDIA, as NVIDIA primarily sells chips to these cloud service providers. Regardless of the purpose, deploying TPUs in the data centers of other cloud service providers will mean a reduction in the number of NVIDIA GPUs used in those facilities.

Gil Luria's team of stock research analysts at investment firm D.A. Davidson told the media that an increasing number of cloud service providers and large AI developers are interested in TPUs, hoping to reduce their reliance on NVIDIA. After communicating with researchers and engineers from several leading AI laboratories, they found that the industry has a positive view of Google's custom acceleration chips for machine learning and AI.

Therefore, the analyst team believes that if Google merges its TPU business with its AI research institution DeepMind and spins it off for public listing, there will be strong market demand. According to Luria's team's estimates, the potential valuation of this business is about $900 billion, whereas earlier this year, their valuation was $717 billion.

"No one wants to be completely dependent on a single supplier for critical components."

"If this business is indeed spun off, investors will gain both a leading AI acceleration chip supplier and a top AI laboratory, which could become one of Alphabet's most valuable assets."

NVIDIA CEO Jensen Huang, however, poured cold water on this competitive chip project. He stated to the media that AI application developers prefer GPUs because they are more versatile and have stronger software support

Bringing in Friends of Nvidia

Media reports indicate that Google's negotiations suggest it is trying to get closer to emerging cloud service providers that Nvidia is focusing on. Unlike Google Cloud and Amazon Web Services, these companies almost exclusively rent Nvidia chips and are more willing to procure a diverse range of Nvidia products than traditional cloud service providers. Nvidia has also invested funds in these companies and prioritized supplying the hottest chips.

Google primarily uses TPUs to develop its own AI models, such as the Gemini series, and the internal demand for TPUs has surged in recent years.

However, Google has long rented TPUs to other companies. For example, Apple and Midjourney rent TPUs through Google Cloud. Earlier this summer, Google even briefly piqued OpenAI's interest in renting TPUs, but OpenAI suddenly changed its mind.

Internally, Google has discussed expanding its TPU business to increase revenue and reduce the cloud computing department's reliance on expensive Nvidia chips. According to two former executives, senior management has also explored selling TPUs directly to customers outside of Google Cloud.

Analysts believe that small cloud service providers like CoreWeave and Fluidstack, which provides Nvidia GPUs to startups like Mistral, have a strong commercial incentive to prioritize offering Nvidia chip servers, as AI developers generally prefer Nvidia products.

However, Google seems to have found a way to encourage Fluidstack to support its TPU expansion plan: if Fluidstack cannot bear the rental costs of the upcoming New York data center, Google will provide up to $3.2 billion in "backing" support. This commitment helps Fluidstack and its data center partners raise debt financing to build facilities.

TPU Demand is Rising

Media reports state that demand for Google's sixth-generation Trillium TPU chips has been strong since they were opened to external customers last December. Analysts expect that demand for the seventh-generation Ironwood TPU will "significantly increase." Ironwood is Google's first chip designed specifically for large-scale AI inference tasks (i.e., deployment operations after model training is completed).

Analysts point out that Google's TPU chips can achieve a computing power of up to 42.5 exaflops (quintillion floating-point operations) and have significantly increased high-bandwidth memory capacity. These chips "also offer significantly improved cost efficiency," which is one of the main reasons attracting more cutting-edge laboratories.

Although the startup Anthropic has previously used TPUs on a small scale, analysts note that the company is currently hiring TPU kernel engineers, which may indicate they are considering switching from using Amazon Web Services' Trainium chips to TPUs. Trainium is a chip designed by Amazon for AI training, and the company has invested $8 billion in Anthropic Analysts also pointed out that Musk's xAI company has shown interest in purchasing TPUs, partly due to the "significant improvements in JAX-TPU tool support" this year. JAX is a high-performance computing Python library developed by Google that allows programs to run efficiently on TPUs. Analysts noted that until recently, the JAX ecosystem had limited the possibility of large-scale deployment of TPUs outside of Google.

According to D.A. Davidson's DaVinci developer dataset, the developer activity related to TPUs on Google Cloud grew by approximately 96% during the six months from February to August 2025