Perplexity CEO Aravind Srinivas has a new favorite metric for the AI industry: “token value per watt per user.”
In a CNBC interview on June 3, Srinivas laid out a thesis that the winners in AI won’t be the ones who spend the most on infrastructure. They’ll be the ones who extract the most economic value from the energy they consume.
The power user problem
For Perplexity, this dynamic is playing out at meaningful scale. The company reported approximately 45 million monthly active users as of early 2026, processing around 780 million queries each month. That’s roughly 17 queries per user per month on average, but averages are deceiving. The real compute load is concentrated among the heaviest users.
Srinivas framed the competitive landscape around five variables that need to be balanced simultaneously: accuracy, latency, cost, privacy, and intelligence. Companies that can optimize across all five dimensions per unit of energy consumed will, in his view, dominate the next phase of AI competition.
