There is a moment in the history of technology when a product’s name stops being a brand and starts being a statement. When Perplexity CEO Aravind Srinivas published a blog post on March 3, 2026, titled “The AI is the Computer,” he wasn’t being metaphorical. He was describing, with unusual candor, exactly how his company built something it believes is the next evolution of a machine that has defined human civilization for half a century.​

The story starts, as so many Silicon Valley experiments do, with a Slack message.

Inside Perplexity’s offices, engineers quietly built what they called ASI — a name that technically stood for nothing, though the team occasionally joked it meant “Artificial Super Intelligence.” The more honest joke, Srinivas later admitted, was that it could just as easily have stood for “Another Slack Integration,” because that’s precisely where the project lived in its earliest days. You’d send ASI a task in Slack, and it would get to work — quietly, without drama. It would spin up sub-agents, make files, search for information, write code, and check in only when absolutely necessary. You could go to sleep and wake up to find weeks’ worth of work had been completed. The vision was a digital worker who never clocked out.​

While this was happening inside Perplexity, the broader AI world was moving fast. Something called Clawdbot — now known as OpenClaw — had surfaced publicly, drawing a lot of attention. Srinivas wasn’t particularly impressed. He thought ASI was already safer and considerably more capable: it had its own file system, could connect to hundreds of tools, browse the internet, write and read files, execute code through a command line, and orchestrate tasks across multiple specialized sub-agents — all within a secure cloud sandbox. The realization that emerged from those experiments was quietly profound. This wasn’t just a digital worker. With a file system, a secure shell, and a browser, ASI was essentially a computer. And so, officially, it became one — Perplexity Computer.​

To understand why Perplexity landed here, it helps to zoom out to 2011, when Google launched the Chromebook with a striking thesis: that a computer only really needed the internet, some files, and a terminal. The Chromebook was right about something important — the web is where work happens, where we communicate, where knowledge lives. But Google underestimated a glaring flaw in that vision. While the web has always been brilliant at storing and publishing knowledge (the first blog appeared in the early 1990s, and the web has been a write machine ever since), the READ side of that equation was broken. For 28 years, we’ve been relying on algorithmic search to access knowledge — a system that was never designed to synthesize, reason, or act on what it finds. Search engines, by their nature, can’t do deep research or behave like an agent. They surface links. That’s it.

Then came AI. Perplexity built what it claims is the first answer engine, and the rest followed. But Srinivas’s argument in his blog post is that solving search was only solving the READ problem — giving people accurate access to the web’s knowledge base. Once you solve READ, and pair it with AI that can take action, the web transforms into a fully functional read/write storage system. And once that’s in place, the only thing a computer still needs is what makes it personal: persistent memory, your files, your tools, your preferences — all of it private and tailored to how you work.​

That’s one half of the Computer equation. The other half is something called orchestration, and it’s where Perplexity is making its most provocative bet.

Here is the central argument Srinivas is advancing: no single AI model, no matter how good, can do its best work alone. The frontier models — Claude, Gemini, GPT, Grok — are increasingly specializing, not commoditizing. One is better at reasoning. Another shines in coding. A third handles long-context recall. A fourth is optimized for speed. When you try to make one model do everything, you get mediocrity at the margins. When you orchestrate them — when you route each task to the model best suited for it — the results are, in Srinivas’s words, more impressive than anything he’s seen in AI for a while.

Perplexity Computer currently pulls from 19 models in its backend, routing dynamically across providers including Anthropic, Google, xAI, and OpenAI. A coding-heavy task might get dispatched to Claude, while a lighter query uses a smaller, faster model. Deep research might go to Gemini. Speed-sensitive tasks might hit Grok. The orchestration layer makes these decisions automatically, weighing latency, cost, and domain alignment before dispatching each call. The system currently uses Opus 4.6 as its core reasoning engine, while models like Gemini handle research, Grok covers speed, and ChatGPT 5.2 supports long-context recall.

To explain why this matters, Srinivas reaches for an orchestra analogy — specifically Steve Jobs‘s famous line about playing the orchestra rather than an instrument. A Stradivarius violin is a masterpiece, but it can’t shake a room the way a bass drum does. And a bass drum can’t move you the way Mahler’s Ninth does with every instrument playing in concert. The model makers, Srinivas writes, are the Antonio Stradivaris of this era — and he means that with genuine admiration. But the models alone are not the product. The harness is.​

In technical terms, the architecture works like this: a planner breaks a complex objective into subtasks, an orchestrator assigns those tasks to the right agents and manages execution, specialized sub-agents perform focused actions in isolated cloud sandboxes with their own file systems, and shared memory stores context and learnings for continuity across sessions. The orchestrator even manages dependencies between sub-agents — if one agent needs data that another hasn’t returned yet, the system queues the task rather than hallucinating an answer from assumed information. It’s this dependency management, more than any individual capability, that separates Perplexity Computer from a glorified chatbot.

The launch itself is significant in ways that go beyond the product. Perplexity’s valuation has climbed from $500 million to $20 billion in 18 months, with backers including NVIDIA, Jeff Bezos, and SoftBank Vision Fund 2. To justify that number, the company needs to show that AI is no longer just a better way to find information, but a genuine replacement for the labor of organizing and executing it. Perplexity Computer is that argument in product form — a bid to evolve from a knowledge engine into what Hindustan Times aptly called an “action engine.”​

Access, for now, is exclusive. Computer is available only to Perplexity Max subscribers, a tier that costs $200 per month and was launched last year for power users and enterprises. Max subscribers receive 10,000 credits per month, with Computer operating on a usage-based credit system rather than the all-you-can-eat model most AI subscriptions use. That shift in pricing philosophy reflects a hard reality: running nearly 20 models in parallel to complete a single project costs real money, and Perplexity isn’t pretending otherwise. Enterprise Max access is coming soon.

TechCrunch, covering the launch, framed it as proof that orchestration beats monolithic design — that the future of AI infrastructure isn’t one model or one agent, but an intelligent system coordinating many. That framing is exactly what Srinivas intends. In his historical sweep, he traces the computer from its origins as a human job title — 400 years ago, a “computer” was an astronomer’s apprentice who calculated orbital paths — through mechanical, electromechanical, and digital phases, all the way to the GUI era that Jobs helped define. But here’s his sharpest observation: the GUI, which was supposed to give us control over computers, has arguably inverted that relationship. We scroll. We tap. We stare. The interface evolved from how we control the computer to how the computer controls us.

Perplexity Computer is, in Srinivas’s framing, a correction — a return to the computer’s original meaning: something that works autonomously, accurately, and asynchronously, freeing humans from the tyranny of watching it work. In 2025 alone, a new frontier model entered the market on average every 17 days. With each new capability, the orchestra gains another musician. The conductor — in this case, Perplexity’s orchestration layer — taps the baton, and the computer goes to work.​

Whether or not you believe that framing is visionary or convenient, the engineering underneath it is real. And for a company that started by challenging Google’s 28-year-old grip on how humans access knowledge, building a product that challenges what a computer fundamentally is might be the most Perplexity thing Perplexity has ever done.