Every few weeks, another AI company launches another chat interface. Different logo, same concept: you type, it responds, you close the tab. The thing is gone the moment you're gone.
Perplexity just did something different. They launched two products in close succession — Perplexity Computer (February 25) and Perplexity Personal Computer — and together they're making an argument that the entire premise of "AI as a chat window" is the wrong model. The right model is an AI that's running whether you're there or not.
I've been watching this space for a while, and I think this is worth slowing down to understand. Not because Perplexity is guaranteed to win — they're not. But because the product direction they're pointing at is where this is all going, whether they build it or someone else does.
What Perplexity Computer Actually Is
The cloud product — Perplexity Computer — launched behind their $200/month Max subscription. At its core, it's an orchestration layer. You describe an outcome: "build me a dashboard tracking my competitors' pricing" or "plan a marketing campaign for this product launch." Computer breaks that into subtasks and routes each one to whichever of 19 AI models handles it best.
Claude Opus 4.6 handles the core reasoning. Gemini does deep research. GPT-5.2 manages long-context recall. Grok handles fast lightweight tasks. Specialized models handle image generation, video, code. The orchestration layer decides who does what, in parallel, and keeps going until the job is done — which can take hours or days.
This isn't a gimmick. There's actually data behind it. In January 2025, 90% of Perplexity's enterprise queries routed to just two models. By December 2025, no single model had more than 25% of usage. The market was diversifying on its own — enterprises were already mixing models based on what worked. Perplexity built the router for that reality.
Their CEO Aravind Srinivas put it plainly: "I don't think Computer could have been as powerful as it is if we had launched it even three months ago. You have to have the right harness at the right time."
That's an honest framing. The timing matters as much as the product.
The Personal Computer Is the More Interesting Bet
The Personal Computer product is what caught my attention more. It's separate software — you run it on a Mac Mini — and it gives the AI persistent access to your local files, apps, and sessions. The agent runs continuously. It keeps context between sessions. It works on tasks while you're not there.
That might sound like a minor quality-of-life improvement. It's not. It's a completely different relationship with the software.
Right now, every AI interaction starts from zero. You open the tab, re-explain your context, get a response, close the tab, and the thing forgets you exist. The "memory" features most apps have bolted on are half-measures — keyword recall dressed up as continuity. They work well enough until they don't, and when they fail, they fail weirdly.
What Perplexity Personal Computer is selling is a model where the AI has actual continuity. It knows your files because it can read your files. It knows what you were working on because it was running when you were working on it. It can take a task you gave it at noon and finish it at 3pm, with no check-in required.
"An always-on AI agent for a Mac that keeps running locally rather than only responding inside a live session." — How outside coverage framed it, and honestly, they nailed it.
The Mac Mini framing is smart product positioning too. It's not your primary machine. It's a dedicated device that exists to run the agent. You don't give it your whole desktop environment — you give it a box whose job is to be available and persistent. That keeps the security concerns more contained and makes the pitch cleaner.
Why This Direction Is Right Even If Perplexity Isn't
I'm not saying Perplexity wins this market. They're competing against OpenAI, Anthropic, Google, and every enterprise software company that's retrofitting AI into existing tools. The distribution advantages stacked against them are enormous.
But the product direction they're betting on is correct. Here's why:
The chat interface was always a temporary default, not the destination. We got chat because it was the fastest way to demo large language models to a general audience. Type a thing, get a thing back. It works. But it's not how most actual work gets done. Most actual work is ongoing, context-dependent, and distributed across time and tools. A tool that forgets you every session is fundamentally misaligned with how that work flows.
Orchestration is the unlock nobody's talking about enough. The obsession with which single model is "best" misses the point. For real workflows, no single model is best at everything. The value is in routing — knowing which model to use for which subtask and stitching the results together. Perplexity's data shows the market already figured this out organically. They're just building the infrastructure for what's already happening.
Persistence changes the economics of AI use. Right now, you have to babysit AI to get value out of it. You prompt, review, prompt again, correct, iterate. The time cost is real. A system that can run a multi-hour research and build task while you do something else is a qualitatively different productivity proposition. It's the difference between having a fast assistant who needs constant direction and having a colleague who you can hand a project to and trust.
The Obvious Concerns
None of this comes without tradeoffs worth naming.
An always-on agent with local file access is an attack surface. The security model for "AI that reads everything on my machine" needs to be airtight — and historically, the AI industry's track record on security has been optimistic at best. Perplexity's answer seems to be the dedicated Mac Mini approach: physical isolation. Whether that's sufficient depends on what you're putting on that machine.
There's also the cost. $200/month for the Max subscription, $2,000/year. That's a real line for individuals. Enterprise pricing is $325/seat/month. For a company with 50 people who need this capability, that's $195,000 a year just in subscription costs — before you account for the Mac Mini hardware, setup, and management overhead. The value case needs to be very clear at that price point.
And there's the concentration question. Running your entire workflow through a single orchestration platform means a single point of failure, a single vendor relationship, and a single point of leverage over your data and operations. The more valuable these systems become, the more uncomfortable that dependency gets.
The Signal in the Noise
I keep watching AI product launches, and most of them are easy to summarize: faster, cheaper, better at coding, better at reasoning. The benchmarks shuffle around and the vibe shifts, but the fundamental product model stays the same — chat interface, zero memory, starts fresh every session.
Perplexity's Computer products are saying something different. They're saying the chat paradigm isn't the right paradigm, and the right one looks more like a colleague or a system than a search box. That's a harder sell, a harder build, and a harder support model. It's also more honest about what people actually need AI to do.
I don't know if Perplexity executes this well or survives the competition long enough to find out. But I know that the problem they're solving — AI that actually fits into how work happens instead of requiring you to fit your work around it — is the right problem. The next few years are going to be a race to see who builds the best answer to it.
Watch that race carefully. It's more important than the benchmark shuffling.