Aravind Srinivas sells an SMB boom that local silicon cannot ship
By Nikhil Raghavan · Reporting from San Francisco ·
Perplexity founder Aravind Srinivas pitches autonomous agents and local chips as an economic savior, but offloading data-center compute onto small businesses creates operational fragilities no solo operator can survive.
A one-person company is currently paying Perplexity upwards of $1 million a year in compute fees to generate $3 million selling automated newsletters. On the podcast Sourcery, host Molly sat down with Perplexity founder and chief executive Aravind Srinivas. Srinivas pointed to that solitary operator as proof of a broader macroeconomic thesis. The next boom in American gross domestic product will not come from enterprise software contracts. It will come from hyper-efficient, AI-native small businesses delegating their operations to autonomous agents. Perplexity, Srinivas argued, scales not by chasing passive consumer logins, but by becoming a neutral orchestrator. It combines search, cloud sandboxes, and multi-model harnesses for users whose depth of utilization compounds their own revenue.
Local Silicon Cannot Solve Distributed Latency
Srinivas’s pitch is seductive because it promises to sidestep the physical wall currently stalling generative infrastructure. With data-center buildouts mired in multi-year utility interconnection queues, substation backorders, and gas-turbine shortages, Srinivas proposes offloading inference from shared cloud clusters to client machines. By his back-of-the-envelope math, Apple silicon chips represent untapped gigawatts of distributed compute already wired to residential grids. Perplexity is building for this by orchestrating hybrid workloads. It escalates sensitive files downward to local hardware while routing complex frontier reasoning up to the cloud.
As an engineering architecture, this is sellable, but it is not shippable at scale. Having managed cloud capacity and platform integrity, I recognize the mechanical trade-offs Srinivas glosses over. Batch inference in a central data center is economical because high throughput amortizes expensive memory bandwidth across thousands of uniform queries. Srinivas correctly notes that as context windows swell, batching advantages shrink. But pushing long-horizon agent trajectories to local laptops introduces brutal failure modes. Client devices sleep, throttle thermally, drop Wi-Fi connections, and lack redundant power. When a local autonomous process hangs midway through reconciling a client database, there is no site reliability engineering team on call. The sole proprietor is the person who gets paged at three in the morning to debug an orphaned execution loop.
Liability Follows the Capital, Not the Model
The strongest case for Srinivas’s vision lies in his critique of corporate software bloat. He rightly observes that enterprise software historically required dedicated full-time employees just to onboard users because the underlying product was too brittle to navigate unaided. If multi-model orchestrators can automate research, synthesis, and execution without forcing non-technical founders to master prompt engineering or context compaction, small firms can operate with unprecedented capital efficiency.
Yet the historical record reveals why this decentralized paradise fails to materialize. Technological shifts from the junction transistor to spreadsheet automation lowered entry barriers, but they did not eliminate the structural advantages of scale: liability absorption, regulatory compliance, and distribution monopolies. When Srinivas points to sovereign deployments and open-source audit classifiers like Bumblebee and Numbat to contain autonomous multi-agent risks, he acknowledges the underlying threat. In reality, an unmonitored agent hallucinating a legal contract or leaking private data exposes a small business to catastrophic statutory liability. Regulators and council members evaluating data-center zoning ordinances cannot rely on the fantasy that residential MacBooks will absorb national compute demand. Power demand does not vanish when fragmented across neighborhoods; it merely hides from the grid operator.
AI will certainly spawn profitable one-person agencies, but autonomous agents running on local laptops will not decentralize economic power. When systems scale, fragility scales with them, and capital flows toward whoever can afford to insure against the wreckage.