Perplexity CEO: SMBs Could Drive the Next Economic Boom
Summary
Srinivas’s central economic bet is that AI-native small and mid-market businesses could become a larger source of US GDP growth. These firms can be more capital-efficient, headcount-efficient, and productive than incumbents, while agents may help them cope with the difficulty of early hiring. His rule is simple: “You don’t think about how to be worth a trillion. You think about how to be useful.”
Perplexity argues that AI economics depend on recursive usage rather than social-media-style network effects. Revenue is up 5X year over year, and one one-person company reportedly spends more than $1 million annually on Perplexity while making $3 million by automating newsletters for clients. Srinivas’s point is that “the depth of usage and the quality of users matters a lot more than the quantity.”
The claimed moat is neutral orchestration across search, browsers, models, files, tools, chips, and cloud sandboxes. Every form of agentic action needs web access and context; better frontier models add capabilities while model competition lowers serving costs. Srinivas’s formulation is deliberately provocative: “We don’t have competitors. We have contributors.”
Perplexity is betting that long-context agents will make local compute economically and strategically important alongside the cloud. Chat can cost 30X more locally, but around 100K context the data-center advantage may narrow to 3X or nearly 1X; meanwhile, 100 million Apple-silicon devices at 50 watts represent gigawatts of potential power, albeit by Srinivas’s own “back-of-the-envelope math.” “The data center is like a bus,” while local hardware is the user’s own car.
Mass adoption hinges on hiding AI’s model and prompting complexity from users. Inspired by Larry Page’s “the user is never wrong” philosophy, Srinivas says failure should be treated as a product issue—not evidence that customers lack prompting, skills files, or sub-agent expertise. His acid test: “Imagine if the iPhone needed an FDE to onboard everybody.”
Privacy and sovereignty make hybrid deployment a trust feature, not just a cost optimization. Perplexity is pursuing fully local, cloud-to-local, and sovereign on-prem configurations; Srinivas also says a completely air-gapped version is possible. The company accepts weaker feedback loops and occasionally “janky” experiences in exchange for user control. For multi-agent security, Bumblebee and Numbat are intended to audit trails, detect sensitive access, and alert; separate remediation agents could act after threats are detected.
Srinivas says writing code is nearly solved while problem selection remains scarce. Once constraints, specifications, security risks, and go-to-market logic are clear, writing code is “almost already a solved problem”; the harder frontier is persistent research where truth cannot be checked instantly. His end state is an agent that tests hypotheses in the world so “anybody can feel like a PhD.”
Deep dive
1. Curiosity is Perplexity’s brand and its theory of human advantage
Srinivas defines Perplexity’s brand as celebrating curiosity, which he calls “the only quality we human beings can be proud of” as AI becomes better at supplying answers. Humanity’s role shifts from being rewarded for knowing things to asking the next, sometimes embarrassingly basic, question.
Borrowing David Deutsch’s framing, he says humans uniquely revisit what is already familiar from new perspectives. The best corporate missions similarly transcend a technology cycle: Apple celebrates beautiful products, while NVIDIA’s mission can encompass AI today and quantum computing or another frontier tomorrow.
Meeting rooms named for scientists reinforce that questions precede invention. Srinivas’s specimen is the junction transistor: amid an industry built around hot vacuum tubes, three people asked whether something fundamentally better was possible.
Ramanujan is his favorite because he represents “the spirit of human intuition”—a faculty that remains useful even as more context and compute encourage people to rationalize everything.
2. AI-native SMBs could become the next American growth engine
Srinivas calls the US a “recursive country”: satisfied customers keep spending, lifetime value is high, and entrepreneurial ambition compounds. He cites Berkshire Hathaway at $1 trillion and JPMorgan near $940 billion as examples of enormous value supported by largely US customer bases.
His American-dream framing centers on outsiders being permitted—and financed—to improve the system. He says this may be especially associated with Silicon Valley, where other founders’ ambition is “infectious,” expanding the dream beyond the job or stability an immigrant initially sought.
Early-stage hiring explains why agents matter. Srinivas made Perplexity’s first 10–20 people work trials because he genuinely did not know how to recruit; good fits were usually apparent within two days, while uncertainty remained after seven days for people who did not work out. “It was a survival mechanism,” not recruitment genius.
He expects future companies to be smaller yet able to make a lot of money, using agents to become more capital- and headcount-efficient. Rather than optimize for Fortune 100 logos because growth investors consider that revenue durable, Perplexity wants to provide tools to smaller and mid-market firms that may create value faster with less bureaucracy.
3. Great AI products must make the user’s mistakes disappear
Larry Page’s “the user is never wrong” philosophy most shaped Srinivas. When engineers explain the correct way to use a failing product, they are solving the user instead of redesigning the product so the user never encounters that problem.
Users should not need to know prompting, skills files, sub-agents, or whether Astra, Fable, Opus, or Sol best fits a task. The harness should anticipate intent, route work, and preserve the familiar prompt-box interface—even when a task is dictated from a phone.
Steve Jobs’s goal of making computers usable by “mere mortals” is the precedent. Srinivas concedes Perplexity is “not there yet,” but argues that widespread FDE-led onboarding often compensates for poor design: an AI failure should be a product issue, not a “skill issue.”
4. Perplexity’s economics rely on recursive depth, not network breadth
The product moved from a 2022 answer engine to multi-step search in late 2023, deep research in 2024, and browser agents last year. Agents naturally extended that stack because real computer work requires internet access, browsing, synthesis, files, and personal context.
Frontier labs supplied stronger coding models; Perplexity combined them with search and browser tools used as code, cloud sandboxes, context compaction, and multi-model routing. That helped drive 5X year-over-year revenue and faster enterprise than prosumer growth through self-serve onboarding.
The sharpest concrete monetization example is a solopreneur spending upward of $1 million annually on Perplexity while making $3 million by automating newsletters. The people using those newsletters are getting clients for their businesses, creating the recursive chain Srinivas distinguishes from a network effect.
The episode’s introduction notes NVIDIA was reportedly considering an investment at a valuation above $30 billion. Srinivas grounds the strategic case elsewhere: search compounds because every agentic action needs web access, while each new model contributes capabilities and price competition reduces serving costs.
5. Hybrid compute turns existing devices into agent infrastructure
Cloud inference wins for repeatable workloads with small context windows because batching and caching raise throughput. As context expands—especially around 100K context—the context window can matter more than model weights in memory. Srinivas says chat can cost 30X more on local hardware, while the data-center advantage for long-context work may narrow to 3X or nearly 1X.
His counterweight to slow data-center construction is distributed hardware: millions, potentially hundreds of millions, of devices already available in homes and offices that can be plugged into power on the grid. Multiplying 50 watts by 100 million Apple-silicon machines yields gigawatts, though he explicitly calls this “back-of-the-envelope math” and notes the devices are not necessarily always plugged in for AI.
The analogy is a bus versus a car. Shared cloud inference remains cheaper for many jobs, but users may prefer locally owned hardware for persistent personal agents, private context, customization, and workloads where the harness matters more than the underlying model.
Portable Computer runs fully locally—with its harness, model, and sub-agents on-device—and can escalate to the frontier. The Apple implementation starts in Perplexity’s cloud but can move sensitive files down to Apple silicon. Srinivas says Perplexity will work with NVIDIA, Apple, Intel, and others.
Molly’s direct question about building Perplexity data centers gets a clear answer: “We don’t have plans to.” Srinivas expects Perplexity to increase cloud demand, but prefers orchestrating existing capacity to dealing with permits, grid queues, substations, power equipment, and gas turbines.
6. Trust requires neutrality, sovereignty, and agent-level defenses
Molly challenges the industry’s rotating “cash grabs,” from coding agents to neoclouds and assistants. Srinivas accepts neocloud growth as evidence that demand for tokens and agents is there, but sees little differentiation when everyone pursues the same market.
His contrarian call is that even the best engineers are not principally bottlenecked by typing code. They need clarity on the problem, constraints, system design, security, and go-to-market; a useful agent should question whether a proposed clone deserves to exist before generating it.
Multi-model operation is therefore both a feature and “a trust signal.” Businesses can build on Perplexity, he argues, because it intends to remain model-neutral, protect privacy, and permit orchestration between local and cloud rather than locking users into one provider.
Sovereign funds are already interested in deployments running entirely on hardware they own, where even Perplexity cannot see anything. The cost is losing product-feedback loops and visibility into bugs; the benefit is a self-contained harness, runtime, agent loop, and possibly fine-tuned open-weight model.
The Hugging Face incident prompts a distinction between one parent agent with sub-agents and multiple collaborating parent agents, each with its own sub-agents; Srinivas considers the latter riskier. He expects slower model rollouts, secure sandboxes, background agents that guard them, and classifiers trained on audit logs.
Perplexity’s proposed tools are Bumblebee for user devices and Numbat for servers. They are meant to inspect trails, flag sensitive access, and alert users; Srinivas separately says remediation agents will be needed once a threat is detected. Fully local or air-gapped runtimes offer maximum control for users willing to sacrifice some frontier intelligence.
7. Long-horizon agents must learn where truth arrives slowly
Srinivas wants Perplexity to move beyond discovering answers on the web toward building prediction models, running experiments, testing hypotheses in real time, and revising them. If that works, “everybody can turn into a scientist.”
The constraint is capability: models are not yet sufficiently persistent, parallel, or autonomous. Math and code permit rapid hill-climbing because correctness can be verified; many questions people are curious about require real-world deployment and slow feedback, where truth cannot be readily verified.
His starting thought experiment is to give a team 10 excellent research interns working 24/7. The aspiration is to make that leverage intuitive enough that it is not reserved for PhDs: the product should let “anybody feel like a PhD.”