Key Takeaways:
A single efficiency metric may determine which AI company commands the highest valuation — and it has nothing to do with revenue.
Perplexity Chief Executive Officer Aravind Srinivas said Wednesday that the companies able to maximize "value per watt per user token" will emerge as the long-term winners in artificial intelligence, proposing a framework that shifts investor focus from top-line growth to operational efficiency. The metric combines accuracy, latency, cost, privacy and intelligence into a single ratio of economic output per unit of energy consumed, he said in a CNBC interview.
"Whoever can maximize that — balancing accuracy, latency, cost, privacy and intelligence — will be the long-term winner," Srinivas said. He contrasted his framework with what he described as short-term revenue growth driven by high pricing, saying some model providers "look like they're making a lot of money because their models are very expensive," but that such revenue may not translate into durable competitive advantage.
The comments come as AI token costs have collapsed from roughly $20 per million tokens in late 2022 to around $0.40 per million tokens by mid-2025, according to Epoch AI data. Yet total enterprise AI spending has moved in the opposite direction: average monthly corporate AI expenditure rose to $85,500 in 2025 from $63,000 in 2024, and nearly half of companies now spend more than $100,000 per month on AI infrastructure or services. The divergence — cheaper tokens driving exponentially more consumption — mirrors the Jevons paradox effect that Mercor CEO Brendan Foody recently described when he said his startup now spends more on AI tokens than employee salaries.
The Efficiency Metric That Could Reshape AI Valuations
Srinivas's framework arrives as the AI industry confronts a growing disconnect between spending and returns. Research shows that 85 percent of companies struggle to demonstrate clear return on investment from AI deployments, even as model capabilities improve at a rapid clip. Uber Chief Operating Officer Andrew Macdonald recently said he has yet to see a proportional link between rising AI token spend and productivity gains, underscoring the challenge of translating technical progress into measurable business outcomes.
Perplexity itself illustrates the valuation gap Srinivas's metric is meant to address. The AI search startup is valued at roughly $20 billion, according to reports — a fraction of Anthropic's estimated $1 trillion valuation and OpenAI's $850 billion-plus price tag. By proposing a framework that weights efficiency over scale, Srinivas is effectively making the case that Perplexity's leaner approach to inference cost management could deliver superior per-dollar returns even without matching the raw compute budgets of its larger rivals.
The company is simultaneously expanding its product footprint. Perplexity Computer, an AI agent capable of executing complex multi-step tasks over extended periods, is now available on Microsoft Windows after launching on Apple's Mac platform in February. The Windows version can access Word, Outlook and local device files, extending Perplexity's reach from cloud-based search into on-device productivity — a move that puts it in direct competition with Microsoft's Copilot and OpenAI's Codex agent, which Microsoft recently integrated with its new Execution Containers (MXC) sandbox for secure agent deployment.
The expansion into Windows gives Perplexity access to the operating system that powers the majority of enterprise desktops, but it also places the startup in competition with Microsoft's own AI ambitions. Microsoft's Project Solara, unveiled at Build this week, envisions a future where devices run AI agents instead of traditional apps, while its MXC platform provides kernel-level security containment for autonomous agents — infrastructure that Perplexity's agent will need to operate within on Windows machines.
For investors, Srinivas's efficiency metric offers a new lens for evaluating AI companies at a time when the industry is spending hundreds of billions on infrastructure with uncertain payback. Companies that can demonstrate lower inference costs per successful outcome — whether through model architecture, hardware optimization or smarter workflow design — may command premium valuations even if their absolute revenue trails larger peers. The question is whether the market will adopt a metric that penalizes the very scale that has driven the AI sector's most spectacular gains.
This article is for informational purposes only and does not constitute investment advice.
