Google is turning down some cloud revenue so its own AI researchers get first call on computing power, and its CEO now says so openly. “Our cloud business is at any given moment in time forgoing short-term revenues because we are making compute trade-offs to support frontier AI development,” Sundar Pichai told Alex Kantrowitz in a Big Technology Podcast interview recorded at Google’s Mountain View headquarters.
Chips and data center capacity are the scarcest goods in the industry, and Google could sell them to outside customers at a premium. Pichai called funding the frontier “ground zero” and said the commitment is “non-negotiable.” Google DeepMind, he said, gets cutting-edge compute access first, and every other business plans around that.
The interview followed the launch of Gemini 4 Argon, which Big Technology says tops the newest public models from OpenAI and Anthropic across a set of tests covering programming, mathematics, science, and office-style knowledge tasks. Those are Google’s own comparisons as relayed in the piece, not an independent test. Pichai said Argon handled almost 25 trillion tokens (the chunks of text a model reads and writes) in its first 72 hours, roughly three times what Gemini 3.0 managed in the same window. He added that Google is “capacity constrained” while serving it.
That detail explains the trade. A model that people use this heavily needs the same chips a cloud customer would rent. Pichai said Google has already felt the squeeze in the other direction: it leased short-term compute last year to help some customers, and a request for capacity in 2027 today is “incredibly difficult” to meet.
He tied the shortage to planning horizons. Google builds its own data centers, lead times are long, and the capacity it has now reflects choices he made in 2024. Last year, he noted, the agentic coding boom had not happened yet. Everyone, in his words, wishes they had twice what they planned for.
Pichai also gave Argon’s first product numbers. In the four weeks since launch, he said, Google’s engineers saw a 30 percent increase in code changes and a one-third drop in rollbacks. Those figures come from inside Google and cover Google’s own developers.
On whether the industry is overbuilding, he rejected the fiber comparison: “It’s not dark. We are using it.” He still allowed that a shortage in any of roughly 2,000 components, from memory chips to electricians, could create bottlenecks, and that a shock to the wider economy could leave someone overbuilt.
The sharper tension is inside Google. The cloud unit has contracts and demand signals, and Pichai said it gets a long-term compute plan. But when a customer wants capacity and DeepMind wants the same chips, DeepMind wins, and enterprise buyers on Google Cloud are paying for that choice with longer waits.
Kantrowitz also pressed him on safety. Pichai said Google is “pretty far” from full recursive self-improvement, where models build better models, and described humans and AI agents working together on research today. He called for independent evaluators and cross-industry standards, comparing them to accounting audits.
Anyone planning a large Google Cloud commitment for 2027 should ask for capacity guarantees in writing, because the CEO has just said where frontier research sits in the queue.
Reported by Big Technology (Alex Kantrowitz) on 20 October 2026.
