Perplexity AI CEO says AI race comes down to value per watt

Why it matters

The fight to define the AI race is also a fight to define valuation. Operators should watch whether AI companies can tie usage metrics to retained, paid work rather than demos or raw engagement.

Perplexity AI CEO Aravind Srinivas told CNBC on Wednesday that the AI winner will be the company that delivers the "most taken value per watt per user," reframing the race around energy efficiency, user utility and economics rather than model benchmarks alone.

"Whoever is able to maximize this particular objective really will, by balancing accuracy, latency, cost, privacy and intelligence all together, they're going to win, that's what's going to win long term," Srinivas told CNBC's Elaine Yu.

CNBC framed the point through tokens, the basic units of data processed by AI models. Each task is broken into tokens, and each token requires energy to process. Srinivas's argument is that the most valuable AI companies will be the ones that generate the best ratio of economic output to power consumed, not simply the ones charging the most for expensive models.

"And so it might feel like some model providers are making a lot of money because their models are very expensive ... but that's short-term revenue growth," Srinivas said.

Orchestration as the moat

Srinivas is using the metric to push Perplexity AI's case for orchestration: the software layer that decides which model should handle a task, how agents should work together and whether processing should happen in a data center or on a user's device.

Perplexity AI develops some of its own models, but its core products also integrate models from outside AI companies, including Anthropic. The company's bet is that a neutral layer across models, chips, operating systems and devices can compound as underlying models improve.

The Perplexity Computer agent, announced in February, is designed to execute complex tasks over long periods of time. The Perplexity AI Personal Computer tool, announced Tuesday, is positioned as an orchestrator that can route work to the best processing environment. On Wednesday, the company said the Personal Computer product will be available on Microsoft's Windows operating system, where it can connect to apps including Word and Outlook as well as files on a user's device. Perplexity AI has already launched the product on Apple's Mac.

"The data center is coming to your laptop," Srinivas told CNBC, adding that an AI operating system that unifies cloud and device-based processing will be crucial.

That matters because running more AI locally could reduce power needs, cut latency and improve privacy when data does not need to be sent to a server. For Srinivas, the strategic claim is that this routing problem is not a feature but the center of the company.

"We believe that by solving that, we'll be building a pretty valuable company that has endurable, long-term advantage," he said.

The competitive pressure

Perplexity AI's orchestration pitch lands as OpenAI, Anthropic and Google all push further into AI agents, while Microsoft and Apple are building AI systems inside the operating systems and apps Perplexity AI wants to span.

CNBC reported that Perplexity AI was last reportedly valued at $20 billion, behind Anthropic and OpenAI, whose valuations have climbed to nearly $1 trillion and just over $850 billion, respectively. Anthropic this week confidentially filed for a U.S. initial public offering, according to the report.

Srinivas argued that Perplexity AI's advantage is that it is not tied to one platform or one model provider.

"I think they absolutely will try to build their own AI systems, but we believe we're building the most versatile operating system by making it work across different models, across different chips, across different traditional operating systems, different hardware providers, different laptops," he told CNBC.

He added that Anthropic's model improvements flow through to Perplexity AI because those models are integrated into its products. Srinivas told CNBC that Perplexity AI has tripled annualized revenue since the beginning of the year, "thanks to model advances that have been made by Anthropic."

For founders, the useful read is that Srinivas is trying to move the AI scorecard from demo quality to operating leverage: how much value a product can deliver per user, per watt and per workflow. If that becomes the investor lens, the companies that win will be those that convert model progress into repeatable, efficient work rather than simply more compute consumption.