Perplexity CEO Aravind Srinivas believes this new superpower will decide who wins AI race

Think the AI race will be won by the company with the most powerful model? Perplexity CEO Aravind Srinivas disagrees. He says the real winner will be the firm that creates the most economic value from every watt of computing power it uses.

In Short

• Perplexity CEO says AI success won't be decided by the biggest models
• The winning company will have to deliver economic value from the computing power
• Perplexity is betting on AI agents and orchestration to gain a long-term edge

Who will win the AI race? The company with the most powerful AI model, the biggest data centres, or the deepest pockets? Not necessarily, or at least that's what Perplexity CEO Aravind Srinivas believes.

Speaking to CNBC, Srinivas argued that the future winners of the AI industry will not be determined by who builds the largest or most expensive models. Instead, he suggest the companies that generate the most economic value from the computing power they consume will ultimately come out on top.

“The companies that can provide the most economic value from the power their AI uses will ultimately command the highest valuations,” Srinivas told CNBC. He added that whichever company can deliver the “most token value per watt per user” will be the real winner in the years ahead.

A token is the basic unit of data processed by an AI model. You can read more about tokens and other AI terms in our AI Dictionary we created for you readers, which explains the technology in simple language.

Coming back to Srinivas' argument, he believes the strongest AI companies will be those that can maximise the value generated from each unit of energy consumed, rather than simply throwing more hardware at the problem.

According to Srinivas, long-term success in AI depends on balancing several factors simultaneously, including accuracy, latency, cost, privacy and intelligence. While some AI providers may appear to be generating substantial revenue because their models are expensive to use, he suggests that such gains may represent short-term growth rather than a durable competitive advantage.

The comments from the CEO come as Perplexity ramps up its focus on agentic AI, systems capable of carrying out more complex tasks with limited human input. Earlier this year, the company introduced Perplexity Computer, an AI agent designed to execute complex tasks over extended periods.

Although Perplexity develops some of its own AI models, many of its products also rely on models from other companies, including Anthropic. Srinivas said the company's focus is on improving efficiency and delivering the best results while minimising energy consumption. To support that vision, Perplexity recently unveiled Personal Computer, which it describes as an AI "orchestrator".

Orchestration refers to selecting the most suitable AI model for a particular task, determining how multiple AI agents should work together and deciding where the processing should take place. Rather than relying entirely on large data centres, Perplexity sees a future where more AI workloads run directly on devices such as laptops and smartphones.

“The data centre is coming to your laptop,” Srinivas said, highlighting a growing industry trend towards on-device AI. He highlights how running AI locally could make systems faster, improve privacy and reduce the energy required to process user requests because data would not always need to travel to remote servers.

Perplexity's push into orchestration comes at a time when competition in AI is intensifying. OpenAI, Google, Anthropic, Microsoft and Apple are all investing heavily in AI agents and next-generation models. Yet Srinivas believes Perplexity's platform-agnostic approach gives it an advantage.

To bring that vision to life, Perplexity has already launched Personal Computer on Apple's Mac platform and this week announced support for Microsoft Windows, allowing the AI assistant to connect with applications such as Word and Outlook, as well as files stored on users' devices.

- Ends