Google celebrates cloud and AI chip wins but dodges the hard question about generative AI revenue

For example, WPP is creating campaigns with up to 70% efficiency gains, and Swarovski has increased email open rates by 17% and accelerated campaign localization by 10 times.

These numbers are vague, hard to compare, and tell us almost nothing about actual business impact.

Token consumption is standing in for hard revenue numbers

Instead of reporting concrete revenue figures, Google leans on token consumption by its 150 (presumably) largest customers as a performance indicator. Each of these customers consumed roughly one trillion tokens last year.

That metric is flawed for several reasons. Token consumption alone doesn't say much about how deeply or profitably a company is using AI. It's also unclear which models were involved, whether the tokens were input or output, and what the actual pricing looked like.

A back-of-the-envelope calculation exposes the weakness. Even under an extremely Google-friendly assumption, using only output tokens at the highest list price of $15 per million tokens (Gemini 2.5 Pro), one trillion tokens per customer would work out to about $15 million per customer per year.

More realistic assumptions change the picture fast. Factor in a mix of input and output tokens, mostly short prompts (≤200k), context caching, and cheaper models, and the number drops well into the low single-digit millions. Depending on the mix, revenue typically lands around $2.5 to $4.5 million per customer per year, with a plausible average of about $3 to $4 million.

Across 150 customers, that adds up to roughly $450 to $600 million. With Google Cloud on track for up to $60 billion in annual revenue in 2025, that share amounts to only about 0.75 to 1.0 percent. Even the theoretical ceiling of $15 million per customer would total around $2.25 billion, or about 3.8 percent. Yet this is the revenue that ultimately has to justify Google's massive AI buildout.

Why aren't Google's customers buying more tokens or paying higher prices?

The low monetization suggests that Google's customers haven't scaled their generative AI applications in any meaningful way yet. One likely reason is that integrating generative AI into existing workflows is complicated. Concerns about data privacy, regulation, quality control, and costs add friction. Reliability and cybersecurity are probably the biggest sticking points. Meanwhile, growing competition in the AI market is pushing token prices down, which further limits revenue potential.

Google's cloud and AI chip business is also riding the generative AI hype more than real demand. Companies are buying infrastructure and specialized processors because they expect generative AI to become economically important soon. But if they can't build scalable, profitable use cases, that enthusiasm will fade, and so will their willingness to invest in the underlying infrastructure.

In that scenario, Google's current cloud and chip revenue would look less like proof of lasting AI value and more like a temporary side effect of inflated expectations.