Pichai says Google gives up cloud revenue to keep frontier AI research first

Google is forgoing short-term cloud revenue to reserve computing resources for its most advanced AI research, CEO Sundar Pichai said. In an October 20, 2026 interview with Alex Kantrowitz on Big Technology, he described that priority as non-negotiable—even as Google struggles to accommodate demand for computing capacity stretching into next year.

The decision puts Google DeepMind’s research and frontier model development ahead of planning for its other businesses. Pichai called that work “ground zero” for Google. Cloud demand still informs long-term investment, he said, through customer commitments, contracts and business forecasts—but research gets the first claim on resources.

Capacity follows decisions made years earlier

Pichai traced today’s constraints to the long lead times involved in building Google’s own data centers and computing infrastructure. Its current capacity reflects decisions he made in 2024. Google plans two to three years ahead, he said, leaving it to anticipate workloads before the technology and customer demand fully take shape.

That gap has practical consequences for customers. Supplying capacity to someone asking now for compute in 2027 would be “incredibly difficult,” Pichai said. Google leased short-term computing capacity last year to accommodate some customers, illustrating that its own infrastructure has not always been enough to meet demand.

Asked whether the industry risks overbuilding, Pichai defended the scale of investment over a long enough period. He said Google is using its infrastructure and faces capacity constraints daily. His stated approach is to weigh contracted revenue opportunities and demand signals against technical progress, internal usage and the trajectory of its models.

Why “good enough” is not his research strategy

Pichai’s argument for pursuing the most capable models goes beyond selling access to them. Working at the frontier gives Google an earlier view of emerging capabilities, he said. He also expects those models to help develop their successors, making research capability an input into the next round of progress.

He nevertheless accepted that less advanced models can power very good products. Frontier models tend to be slower and more expensive, he noted, and Google still sees room for smaller offerings such as Flash and Flash-Lite. His position was not that every task needs the strongest model, but that Google must keep developing one.

Gemini 4 Argon provided his immediate example. Pichai said deploying it internally had sharply improved what Google’s developers could tackle, including problems earlier models could not solve. He also acknowledged pockets of coding performance that need work, citing clear feedback rather than presenting the model as uniformly strong.

Pichai put Argon’s early usage at about three times Gemini 3.0’s over the same period. He said Google was capacity constrained while serving it.

Safety requires more than training the model

On AI helping build AI, Pichai drew a boundary around current capabilities. He described agents assisting human researchers in parts of model development, not a fully self-improving system operating on its own. He expects that work to advance into next year, while saying today’s systems remain far from that end state.

He acknowledged clear gaps in alignment—the work of getting models to behave as intended. Google also relies on real-time monitoring to spot unexpected behavior and intervene, he said. Isolated operating environments and customer controls over deployment and data access form other parts of that approach. Safety testing itself requires dedicated computing resources.

Google’s October 13 safety overview supplies detail behind that description. The company says Argon’s protections screen incoming requests, inspect internal model signals for dangerous activity, and monitor responses as they unfold. A separate misalignment monitor examines agents’ actions and internal reasoning, blocking severe boundary violations and flagging serious internal cases for human responders.

Pichai said it is reasonable to worry about whether safeguards are enough. He called shared industry standards and independent evaluations natural next steps, comparing them with accounting standards and outside auditors. Regulation, he said, will also have a role. The full 47-minute interview is available on Big Technology Podcast.