Google was exposed to increase the promotion of self-developed chips: actively contact NVIDIA chip cloud service providers

Google has recently been in contact with a group of small cloud service providers that mainly purchase Nvidia chips to discuss deploying Google's self-developed chips in their data centers. According to people familiar with the matter, negotiations have made preliminary progress: an agreement has been reached with London-based Fluidstack to host Google TPU at the latter's data center in New York.

In addition to Fluidstack, Google has also contacted other cloud service providers including CoreWeave, known as "Nvidia's son".This series of moves shows that Google is trying to close the relationship with emerging cloud service companies that are "supported by NVIDIA."

Unlike large cloud vendors such as Google and Amazon, these emerging companies rely almost entirely on NVIDIA chips and are more active in purchasing a variety of NVIDIA products. Nvidia not only invests in many of these companies, but also prioritizes supplying them with the most in-demand chips.

The way Google reached cooperation with Fluidstack this time is particularly direct: If Fluidstack cannot afford the construction cost of the New York data center, Google is willing to step in as a "backup guarantee" and provide financial support of up to US$3.2 billion.

It is unclear why Google actively promotes its self-developed chips.Previously, Google almost never sold self-developed TPU to external parties. According to Capvision data, 70%–80% of Google TPU’s computing power is used for internal business, and the remaining 20%–30% is open to the outside world through self-construction and leasing.

Some analysts believe that this strategy may stem from the fact that Google's progress in building its own data center is difficult to match the rapid growth of its chip demand, or it may be due to the hope of gaining more customers for TPU through third-party cloud service providers. If it is the latter, it means that Google is competing more directly with Nvidia.

The analysis further pointed out that if Google expands the deployment of TPUs in a rental model, it will be similar to the current practice of cloud service providers using Nvidia GPUs. As the penetration rate of Google TPU increases in these data centers, the purchase demand for Nvidia GPUs in corresponding scenarios may decrease.