Google Parent Company CEO Reveals Google Data Center Allocation Priority: DeepMind First, TPU First Direct Sale in a Decade

According to Dongcha Beating monitoring, during Alphabet's first-quarter earnings call, analysts questioned how computational power scarcity would be allocated. CEO Sundar Pichai outlined the priorities: first, ensuring computational resources for training cutting-edge models at Google DeepMind, with the rest allocated among Search, YouTube, and Cloud based on the Return on Invested Capital (ROIC) framework. He admitted that the company is currently facing constraints on computational power and that "if demand could be met, cloud revenue could have been higher".

To alleviate the supply pressure, Google will, for the first time, directly sell TPU hardware to selected customers for deployment in their own data centers. Previously, TPUs were only used internally and leased in the cloud, never sold directly. The target customers include quantitative trading firm Hudson River Trading, AI lab Thinking Machines Lab, and robotics company Boston Dynamics. CFO Anat Ashkenazi provided revenue guidance: TPU hardware sales revenue will only account for a small portion this year, with the vast majority realized in 2027. These hardware agreements are already included in the cloud backlog (which nearly doubled this quarter to $462 billion), but the majority still comes from regular Google Cloud Platform (GCP) contracts, with slightly over 50% expected to be recognized within 24 months.

Another signal of increasing demand: the token processing volume of Google's in-house models APIs such as Gemini has increased from 100 billion per minute in the previous quarter to 160 billion per minute, representing a 60% quarterly growth. In the past 12 months, 330 cloud customers have individually processed over 1 trillion tokens each, with 35 customers surpassing 10 trillion.