AI is being judged on revenue now

For several years the AI conversation ran on capability. Benchmark scores, parameter counts, what the newest model could do that the last one could not. Research output was the currency.

The questions changed. What I hear now in procurement conversations and budget reviews:

• What does this cost per unit of work
• Who owns the data it touches
• What happens when it is wrong
• Does the revenue survive the vendor raising prices

Those are the questions asked of mature technology. Palantir reporting 93% revenue growth in Q2 2026 belongs to the same shift, as do enterprises building internal AI cost controls and governments moving on oversight. Different signals, same direction.

The counterweight is worth stating alongside it. A 2026 WRITER survey of 2,400 enterprise respondents found 54% of C-suite leaders saying AI adoption is tearing their company apart. Revenue growth and organisational strain are both real at the same time, and the second is what the commercial questions are reacting to.

What it changes for builders

Unit economics become a design constraint. Which model runs which step stops being an implementation detail and becomes a margin decision. I have written about where this hides in Snowflake specifically, and the general version applies everywhere: cost per request, cached against live, frontier against open-source per task.

Governance moves from compliance to product. Data ownership, retention, and deletion turn into things enterprise buyers ask about during evaluation. A product that cannot answer loses to one that can, regardless of model quality.

Reliability outweighs peak capability. A product that works consistently at 85% beats one that reaches 95% and fails unpredictably, because the buyer builds a process around your behaviour and unpredictable behaviour breaks it.

Distribution reasserts itself. The best model does not win a market on its own. It never has in any previous technology cycle.

The open question

Which of capability, distribution, trust, or execution ends up deciding the next five years is genuinely unsettled, and anyone claiming certainty is selling something. My own bet leans toward execution and trust, on the reasoning that capability keeps commoditising while the operational work of running AI in production stays hard.

That is a bet, not a forecast. What I hold with more confidence: the era where a capability demonstration was sufficient has closed, and products now have to show they create value that lasts past the pilot.