Google, Blackstone launch AI infrastructure joint venture

Google LLC and Blackstone Inc. are launching a joint venture that will provide a “compute-as-a-service” platform for artificial intelligence workloads. 

The companies announced the collaboration on Monday. The yet-unnamed venture is launching with $5 billion in equity funding from Blackstone, which will reportedly receive a majority stake. The investment is said to be worth $25 billion including leverage, a term that usually refers to debt financing.

The venture will be led by Benjamin Treynor Sloss, a long-time Google executive who currently serves as its Chief Programs Officer. In that capacity, Treynor Sloss oversees many aspects of the search giant’s cloud business including its data center design efforts and network.

The new venture will reportedly use the search giant’s custom Tensor Processing Unit, or TPU, AI chips to power its infrastructure. The processors are currently only available through Google Cloud.

Last month, Alphabet Inc. Chief Executive Officer Sundar Pichai revealed that the company will let a limited number of customers deploy TPUs in their data centers. He cited demand from “capital markets firms” as one factor behind the decision. Blackstone is one of the world’s largest institutional investors with about $1.3 trillion in assets under management. According to the firm, that sum includes more than $150 billion worth of data centers.

Google debuted its two newest AI chips, the TPU 8t and TPU i8, last month. The processors are optimized for training and inference workloads, respectively.

When a chip receives a prompt, it splits the message into chunks and assigns each one to a different core. After all the cores process their respective chunks, the chip collects the results and distills them into a single prompt response. Google’s inference-optimize TPU 8i processor includes a module called the CAE that is specifically optimized to speed up that phase of the inference workflow. According to the company, the engine can reduce the latency involved in the task by a factor of five.

The TPU 8t also includes multiple task-specific optimizations. One of them is is a module called the SparseCore that enables the processor to more quickly retrieve the data needed for a training run from memory. For added measure, Google ships both of its new chips with a suite of networking technology called TPUDirect. It speeds up the flow of data center traffic by enabling packets to bypass components such as servers’ central processing units. 

“This joint venture with Blackstone helps meet growing demand for TPUs, which are optimized specifically for efficiency and performance in the AI era,” said Google Cloud CEO Thomas Kurian

The TPU clusters in Google data centers include many auxiliary components that the search giant developed in-house. The TPU 8i, for example, is deployed in so-called pods that use a custom architecture called Boardfly. It’s possible that Google’s joint venture with Blackstone will adopt not only GPUs but also other custom components of the search giant’s AI infrastructure. 

The newly formed company will initially build 500 megawatts of computing capacity. Google and Blackstone plan to significantly expand its data center footprint over time. 

The deal raises the possibility that the investment firm may also bring GPUs to its other portfolio companies. In 2024, Blackstone spent $16.1 billion to buy Australian data center operator Airtrunk. Three years earlier, it acquired another AI infrastructure builder called QTS Realty Trust for $10 billion.

Photo: Google