Pichai says cloud yields compute to frontier AI, 2027 capacity hard to lock

In October 2026, Google released its flagship model Gemini 4 Argon, which consumed 25 trillion tokens within 72 hours — three times the volume Gemini 3.0 used over the same period in its launch window, and a record for the company. Sundar Pichai then gave his first in-depth public interview since the launch, speaking with tech journalist Alex Kantrowitz on the Big Technology Podcast at a studio in Mountain View. It was his first extended public appearance after the model's release.

During the interview, Pichai pushed back on claims that Google has stepped back from frontier AI, and acknowledged that the company's cloud business is deliberately ceding compute capacity to frontier model development. He said compute is severely short across the industry: every company now wants twice the capacity it originally planned for, and customers trying to lock in compute for 2027 today face an extremely difficult task.

On recursive self-improvement (RSI), Pichai drew a clear line. He described the current phase as a transitional period in which AI agents and human researchers work in tandem, and said the point of an autonomous, runaway endgame remains quite far off. The report also noted a view that has circulated in Silicon Valley over the past year — that the frontier model race has entered a phase of diminishing marginal returns, and that lightweight, fast models such as Meta's open-source Llama series and Google's own Gemini Flash are already sufficient for most enterprise use cases. Pichai, for his part, called frontier models Google's first priority. The story currently rests on the positions he laid out in this interview.