Perplexity CEO Aravind Srinivas on AI Agents, Compute and Competition

Aravind Srinivas says the prospect of winning motivates him more than fear of failure. He recalls growing up in a lower-middle-class family in India and initially seeing a job as an engineer at Google as a major success. His approach to building Perplexity is shaped by that personal experience.

Srinivas says he was more outspoken about competition with Google in Perplexity’s early days, but now sees the product as broader than search. He describes an answer engine as the entry point for later research and agent products. This is his account of the company’s strategy.

He claims Perplexity influenced Google’s search interface and offers his own comparison of the products. That is the CEO’s competitive view; the interview does not establish the causes of Google’s design decisions or independently compare product quality.

Srinivas says more sophisticated research and agent work are important to Perplexity’s paid products. He argues that a company cannot assume an early advantage will persist. His assertions about competitors’ positions and future prospects are opinions rather than established outcomes.

Stebbings and Srinivas discuss OpenAI’s potential readiness for an IPO and distinguish financial readiness from perceived product leadership. Their judgments are speculative; no offering, date or financial assessment is established by the conversation.

Srinivas argues that subscription value may move toward systems that perform work, while conventional answers can be a way to attract users. Stebbings asks whether advertising in chat could become a major business. Both are discussing uncertain future markets.

Srinivas is skeptical that a conversational interface will capture all the advertising value of search or visual discovery. He gives travel and shopping as examples where people may want to explore alternatives. The advertiser spending figures he mentions are estimates and are not independently verified.

He also argues that advertising can create trust concerns if users cannot distinguish a recommendation from a paid placement. That is his product opinion, not proof that any particular service’s answers are inaccurate or that chat advertising cannot succeed.

Srinivas clarifies that by “frontier” he means the most valuable outcome a system can currently produce, not necessarily only the largest model. He describes an agent harness as the rules, tools, connectors and context around a model that help turn output into useful work.

In his view, simply reselling model output is hard to differentiate. Perplexity’s proposed advantage is orchestrating multiple models as well as tools and files. This is a company strategy and competitive claim, not an independent assessment of other providers.

He proposes “value per watt per user” as a way to think about useful output relative to energy and cost. The concept is his framing, not a standardized metric established in the interview. It does not by itself show that one product is more efficient than another.

Srinivas says a small group of power users can account for significant agent consumption and offers examples of high spending. These anecdotes are not independently verified, and they do not show that such spending is typical or produces a positive return.

He distinguishes one-off tasks from recurring agent workflows that monitor events, triage information or investigate problems. He expects continuous work to become more important, though reliability, permissions and cost remain practical questions.

Srinivas predicts agent products could eventually generate revenue comparable to or greater than major advertising businesses. That is a long-term business forecast, not a reported result. Stebbings asks how AI spending might relate to developer salaries and company valuations.

The CEO says non-developer work could also be a large market for agents, naming finance, sales and research. He does not give a reliable estimate of future token spending as a percentage of salary; his market-size comparisons are speculative.

He expects the price of a capability to fall as models improve, even if customers pay for new frontier tasks. A hypothetical cheaper open model could perform some current work, while a more advanced system might enable different work. Neither future performance nor spending is guaranteed.

Srinivas imagines smaller software teams using many agents and points to possible scientific and engineering applications. These are visions of future capability, not proof that present systems can autonomously make discoveries or solve medical problems.

Asked whether frontier problems will eventually run out, he argues that human curiosity and new questions can continue to create opportunities. His comments about abundant intelligence, energy and future economics are philosophical speculation rather than established economic outcomes.

For an always-on agent, Srinivas describes a tradeoff among capability, accuracy, privacy and cost. He proposes using local models where appropriate and server-side models for harder work, with a system that selects tools and preserves relevant context.

He says a continuously running server-side model could be expensive and imagines local computation reducing some costs. Actual costs depend on hardware, energy, usage and implementation; local processing is not free or automatically private.

The proposed system would combine models, a harness, devices and tools. Srinivas believes users may want more control over sensitive context, but local deployment still requires security and governance. The interview does not demonstrate that the envisioned system eliminates surveillance or data breaches.

He uses an orchestra analogy: models and tools are instruments, sub-agents perform parts, and the orchestration system coordinates the work. Perplexity positions its own product in that role. His claim that an orchestrator will capture the most value is a prediction with a clear company interest.

Srinivas says Perplexity’s revenue has more than tripled since the start of the year and that model competition helped reduce its burn. These are company-reported statements; the interview does not supply audited financials, a full baseline or evidence for each causal attribution.

The discussion shifts to infrastructure. Srinivas identifies power, land, cooling, equipment and permitting as possible constraints on data-center development. He expects new hardware to improve model capabilities, though future performance and deployment timelines remain uncertain.

Stebbings asks about the economics of Meta’s capital spending. Srinivas speculates about possible subscription and cloud services, then suggests Micron might become more valuable than Meta. This is a securities prediction based on his view of memory bottlenecks, not verified pricing or investment advice.

He argues that memory, processors and storage can gain value when they constrain AI workloads. His descriptions of individual suppliers and demand are market opinions; a bottleneck does not guarantee a particular stock return or valuation.

Asked whether neocloud providers can become durable large businesses, Srinivas says some could if they secure power, capacity and reliable service. He does not identify a certain winner and says operational execution matters. The company examples and valuations are illustrative.

He believes simply renting hardware may be less differentiated than combining capacity with software services. The comparison with cloud providers is his business thesis, not an independent measurement of any provider’s margins or prospects.

Srinivas says a very large inference business is possible under assumptions about revenue, margins and access to capable open models. He also identifies model concentration as a risk. These are conditional scenarios, not evidence that any company will reach the hypothetical valuation.

On routing providers, he distinguishes selecting among different models from obtaining reliable access to endpoints for the same model. He suggests reliability and capacity can have value. His description of another company’s contracts and margins is an external estimate, not verified financial disclosure.

Srinivas expects power to remain a major bottleneck and discusses public opposition to data centers. He dismisses some water and power concerns, but the interview offers no evidence for a blanket dismissal: electricity demand and water use depend on facility design, location, grid and measurement.

Stebbings suggests that employment and wealth inequality can also shape opposition. Srinivas agrees concerns can take several forms, including environmental impacts and electricity prices. The percentage of projects the host cites is not sourced in the interview and should not be treated as an established rate.

Srinivas imagines data centers moving to locations with available resources and potentially more favorable permits. He says infrastructure requires capital, supply chains and physical execution. The location, environmental effects and feasibility of any specific project depend on its circumstances.

He acknowledges a risk of overbuilding if future models become substantially more efficient. His probability estimate is a personal guess, not a measured likelihood. Lower compute needs for a task could alter infrastructure economics, but the overall effect would also depend on demand.

Srinivas discusses US export controls and suggests they can both constrain access in the short term and encourage alternative approaches over time. His claims about specific Chinese systems, supply chains and the causes of capability gaps are not independently established in the interview.

He argues China has strengths in physical infrastructure and manufacturing and suggests the US should invest in its own capacity. His broad statements about power, permits and relative national capabilities are opinions; actual constraints and policy effects vary by project and region.

The CEO argues for public discussion grounded in evidence rather than fear. He dismisses some environmental concerns too broadly; local water, emissions and electricity impacts require independent assessment. His recommendation to use Perplexity for fact-checking is a product claim, not a guarantee of accuracy.

Srinivas challenges messaging that emphasizes job losses without describing new opportunities. He believes AI may make it easier for small groups to build companies, while acknowledging a debate about work and displacement. His projections do not establish that benefits will be evenly distributed.

He describes cloud credits received by Perplexity and a proposed credit program for new companies. The figures and program details are his account, and access to credits alone does not guarantee a viable business or an investment outcome.

Asked about headcount, he says Perplexity has around 400 people and speculates about future growth. He imagines more companies producing valuable products with small teams. The staffing numbers, valuations and comparisons are his statements and scenarios, not verified economic results.

Srinivas encourages people unfamiliar with AI to start with a question that interests them. He also clarifies a past comment about people who dislike their jobs, saying he intended to describe an opportunity for some individuals to pursue ideas. His account does not establish that entrepreneurship is accessible or suitable for everyone.

Stebbings asks whether adoption and productivity will rise smoothly. Srinivas expects setbacks and says controlling agent costs will matter. He favors an orchestrator that can choose models and manage budgets, a role aligned with Perplexity’s product strategy.

Asked whether Google could be the lowest-cost token producer, he cites its infrastructure advantages but argues about its position in coding models. Those competitive rankings and future cost outcomes are his opinion, not an independent benchmark.

The speakers discuss a reported change in web traffic and whether agents will displace websites or advertising. Srinivas predicts that subjective choices such as fashion may retain a role for browsing and advertising, while more objective transactions may be automated. This is his forecast, not a settled market division.

Srinivas encourages leaders to ask what they would do if they could delegate far more work. Stebbings describes an imagined investment research and outreach system. Their examples depend on hypothetical capability, budget and oversight rather than current guaranteed performance.

Asked about Perplexity’s goal, Srinivas describes making people more curious and says financial milestones can be motivational rather than the whole mission. His discussion of trillion-dollar valuations is aspirational and does not establish a likely outcome.

On inequality, Srinivas argues that access to useful tools and examples of entrepreneurship could spread benefits. He recounts a driver who told him about building an application. The anecdote does not demonstrate typical income, eliminate barriers or establish that inequality will decline.

The speakers differ over how much personal agency people can exercise in difficult circumstances. Srinivas says not everyone will become wealthy but hopes more people can try building things. Their discussion is a values-based perspective, not evidence of equal opportunity.

Stebbings raises possible future IPOs for large private companies. Srinivas speculates that investors might reallocate capital from existing holdings. No offering, timing, investor decision or valuation is established; the hypothetical allocation is not investment advice.

Srinivas says Perplexity is growing and hopes to go public, but does not disclose a complete set of current financial results. He argues that growth and a path to efficiency both matter. His revenue claims and any IPO timing are company statements and remain uncertain.

He says Perplexity is training its own models on open foundations to reduce reliance on expensive tokens for existing uses, while expecting to use frontier models for new capabilities. Cost savings and full migration are goals, not independently verified results.

Srinivas says rapid model progress leaves all providers under competitive pressure. He describes how a negative external assessment of Perplexity contrasted with the company’s subsequent growth according to his own figures. The interview does not independently verify that growth or establish the views of the people he criticizes.

In a quick-fire round, he says the company could adopt its own and competitors’ tools more deeply. He imagines semi-autonomous systems assisting business functions with human scaffolding. That is a vision for future operations, not a claim that such a system already exists or can safely run a company.

With unlimited capital, Srinivas says he would invest in physical data-center infrastructure, while continuing to focus on useful outputs. This is a hypothetical allocation and an argument for the importance of execution, not a project announcement.

Asked to choose a ten-year investment among three companies, he chooses SpaceX because he values its infrastructure and connectivity businesses. The selection, assumed IPOs and possible future applications are speculative; no transaction or investment recommendation follows.

He advises a recent graduate to remain curious and avoid treating others’ wealth as the only measure of opportunity. He also argues that new AI labs need genuine differentiation rather than simply replicating an existing approach. Both are his personal views.

Srinivas imagines Perplexity growing by improving accuracy and orchestration, but says the process and impact motivate him more than a particular valuation. The very large valuation milestones mentioned are aspirations, not forecasts supported by disclosed financial projections.

He reflects on people he has met in technology, praising the ability to concentrate on a limiting problem and to remain demanding of one’s work. His descriptions of individual leaders and their motives are personal impressions, not independently established facts.

Srinivas says he admires long-term ambition and continued engagement in work. He and Stebbings close by acknowledging the difficulty of building companies. His views on retirement and family are personal preferences rather than universal advice.