Pichai: Google needs 'much larger base models' as cloud revenue surges 82%

Google’s AI bets are starting to show real financial muscle. Alphabet’s Q2 2026 revenue hit $119.8 billion, a 24% jump that handily beat Wall Street’s $116.9 billion expectation. Google Cloud was the standout, exploding 82% to $24.8 billion. The company is so bullish on the trajectory that it raised its 2026 investment forecast to between $195 billion and $205 billion, insisting demand still outpaces what it can build.

Usage numbers are equally staggering. The Gemini app now claims 950 million monthly active users, up from 750 million in February. AI Mode in Search crossed the one billion user mark since its October launch, which Google says is driving more overall search queries. On the advertising front, the AI Max tool just left beta and already has 500,000 advertisers on board. Alphabet is framing this as unlocking billions of new, previously hard-to-monetize queries.

But scale doesn’t equal dominance at the frontier. CEO Sundar Pichai was candid about the gap: “There are areas where we have acknowledged we need to improve. Coding and agentic coding is an example of that.” His fix is Gemini 4, now in what he calls Google’s “most ambitious pre-training run yet.” The core insight? The next leap requires building “much larger base models” to catch rivals. Gemini 3.5 Pro is already testing but apparently won’t ship as part of the new monthly model cadence.

Here’s where I see a fascinating tension. The cost per AI response keeps dropping even as models get more powerful—a trend the Flash 3.6 release reinforced. That efficiency could give Google a genuine margin advantage in serving AI at planet scale. For the AGI race, a frontier model like Gemini 4 matters enormously. For the tens of millions of daily queries hitting Flash models, maybe less so. Pichai himself noted the “sweet spot of performance and cost” is what developers actually reach for. The real question isn’t whether Google can build a model that tops the benchmarks. It’s whether the economics of serving AI will make that capability a winner-take-all advantage or just one card in a much bigger hand.