Pichai: Google Sacrifices Cloud Revenue to Fund Frontier AI, Argon Outperforms Rivals

Google is currently forgoing short-term cloud revenues to allocate compute resources to frontier AI development, a strategy Pichai describes as non-negotiable since 2015.
Gemini 4 Argon outperforms OpenAI and Anthropic's latest models in coding, math, science, and knowledge-work benchmarks, processing 25 trillion tokens in the first 72 hours—three times the usage of Gemini 3.0 in the same period.
Internal metrics show Argon has increased code change lists by 30% and decreased rollbacks by one-third, indicating higher productivity and quality in Google's engineering workflows.
Pichai views compute constraints as beneficial, arguing they inspire creativity and more efficient algorithmic breakthroughs, such as reducing memory usage.

Compute Strategy & Availability

Current compute holdings are the result of decisions made in 2024, with plans based on demand 2-3 years out.
Meeting 2027 compute demand is described as very difficult, comparable to Andy Jassy's "sold out by 2028" statement.
Overbuild is possible due to bottlenecks in memory, power, electricians, and land, but Pichai is confident long-term demand will exist.

A unified Gemini agent is being brought to consumers, running 24/7 in the cloud with a micro VM spun up for every query.
The goal is to unify experiences across Workspace, Search, and Gemini, with a growing percentage of inputs now being agentic tasks rather than simple chats.
Pichai rejects the framing that Google is bolting AI onto existing products, citing the Gemini app as a world-class, AI-native experience built from the ground up.

Google has invested tens of billions of dollars in Anthropic and is licensing Gemini for Apple's Siri.
Space data centers are 2-3 years into the journey, with TPsUs already sent to space and a "hello world" moment expected imminently.
Research bets include 1 million output tokens for Argon, 1 million context window for Gemini models, and world models for embodied reasoning.