Unpacking TPU-As-A-Service

Research evidence excerpt

Unpacking TPU-As-A-Service

Barclays | Alphabet Inc.

Background On TPU-As-A-Service

Two major announcements have happened in the past quarter that could be material to GOOGL

financials over the coming years. The first was on May 18th where Blackstone and Google Cloud

set up a JV run by a former Google exec to provide TPU-as-a-service to customers, with the first

0.5 GWs coming online in '27 and more capacity into the future1 . This was followed up by

another mega-deal announcement forming an AI XPV Platform between Broadcom and

Blackstone/Apollo, enabling as much as 20 GWs of AI capacity through 2028, with the first >1 GW

coming online starting in "mid-2026"2 . The latter will involve AI labs bringing both TPUs and

other custom ASICs (OpenAI's Jalepeno, etc.) online.

For purposes of this report we only really care about the portion that flows through Google's

P+L. These two entities could bring online 15-20 GWs of TPU-as-a-service AI cloud capacity

outside of what Google is doing inside GCP.

First Off, What Is TPU-As-A-Service (TPU-aaS)?

Historically, the only way an AI lab or enterprise customer could access AI specific compute

running on Google's TPUs was via Google Cloud Platform (GCP). Customers would enter into

agreements with Google to rent AI capacity on various terms. However, this is starting to change

in '26 with the new "off platform" structures noted above. In the very near future, AI labs who

are the largest buyers of AI compute (Anthropic, OpenAI, Meta, etc.) will be able to buy compute

through a new vehicle that allows for more direct control. Broadcom explained it best in their

press release: "It also establishes a scalable framework for future deployments of XPU-based