Alphabet puts AGI work ahead of TPU sales as AI spending rises

Alphabet is reserving its tensor processing unit capacity first for internal artificial general intelligence work even as it has begun selling TPUs to some customers, CEO Sundar Pichai told analysts on the company’s second-quarter earnings call. The allocation choice is tied to a larger infrastructure bill: CFO Anat Ashkenazi said Alphabet now expects to spend $195 billion to $205 billion this financial year, above its prior $180 billion to $190 billion range, while supply constraints force it to buy outside compute capacity.

Pichai’s comments came after Goldman Sachs analyst Eric Sheridan asked how Alphabet weighs external demand for TPUs against its own need for accelerators. Pichai said the starting point is keeping enough capacity to compete at the frontier of AGI development, referring to artificial general intelligence, the term generally used for systems with human-like capability.

Bernstein analyst Mark Shmulik pressed the same issue from another angle, asking how Alphabet splits compute across search, cloud and model training. Pichai said the baseline remains AGI work, followed by core products including Search and YouTube, then Google Cloud workloads. Inside cloud, he said Alphabet is prioritizing capacity for its own models in Vertex AI and Gemini Enterprise, along with data analytics and cybersecurity offerings.

The company did not disclose how many TPUs it has sold, which customers bought them, pricing, margins or how much capacity it is keeping for itself. That leaves investors with the broad priority order, but not the economics of selling chips versus using them to train and serve Alphabet’s own models.

Alphabet said demand for AI infrastructure and AI products is feeding Google Cloud growth. The company reported Google Cloud revenue of $24.75 billion for the quarter, up 82 percent year over year, and profit of $8.8 billion, up 214 percent. It also said Google Cloud backlog has reached $514 billion, meaning contracted services that customers have not yet consumed.

Core advertising businesses also grew. Alphabet reported search revenue growth of 17 percent and YouTube ads growth of 13 percent. Pichai said Google’s AI Mode in search is increasing overall query volume, a relevant data point after generative AI products raised questions about whether conversational interfaces would pressure Google’s search advertising business.

Serving AI search is costly because each answer consumes specialized compute. Pichai said Alphabet reduced the cost of AI Mode responses to the lowest level since launch during the quarter, attributing that to engineering and hardware improvements while the company added more advanced AI features.

Supply remains the limiting factor. Ashkenazi said Alphabet cannot obtain all the equipment it wants in the current constrained market and will expand its use of third-party capacity in the third quarter as a temporary measure while it builds more internal infrastructure.

Pichai framed that outside capacity as a way to keep large cloud customers engaged during a period of intense AI demand. He said Alphabet may accept high short-term costs over several months when a customer opportunity spans multiple years and the expected returns over the life of the deal are attractive. The company did not name those customers or quantify the margin trade-off.

Alphabet reported quarterly revenue of $119.8 billion, up 24 percent year over year, and operating income of $40.8 billion, up 34 percent. Free cash flow was negative $5.9 billion, the first such shortfall since 2004, and Alphabet shares fell 4 percent in after-hours trading.

This story draws on original reporting from The Register.