Perplexity's hybrid Comet AI keeps sensitive data off the cloud

What happened

Perplexity has launched hybrid compute for Comet, its agentic AI platform, allowing a single AI agent to split a task between frontier models running in the cloud and smaller open-weight models running locally on Apple silicon Macs. Sensitive portions of a task are routed to the local machine and never leave the device, while the rest of the work continues in the cloud.

According to Perplexity, this marks the first time an AI agent can start a task in the cloud and dynamically hand off the confidential elements of that same task to a model running on the user's own hardware — without restarting the job or losing context.

The hybrid feature is available now through Perplexity's desktop app for opted-in enterprise customers, as well as Pro and Max subscribers, on any Apple silicon Mac running macOS 15 or later.

Why it matters

This is fundamentally a technology story about how agentic AI handles the trade-off between capability and confidentiality. Frontier cloud models typically offer stronger reasoning, but sending sensitive data off-device has been a persistent blocker to enterprise adoption. By letting one agent seamlessly split work between cloud and local compute mid-task, Perplexity is attempting to remove that trade-off rather than force organisations to choose between accuracy and data control.

For organisations weighing AI adoption in regulated or data-sensitive functions — legal, HR, finance, healthcare — this points to a broader shift: agentic platforms that can dynamically route work based on sensitivity, not just model choice, may become a baseline expectation rather than a premium feature. The approach also signals how local device compute (Apple silicon, open-weight models) is increasingly being positioned as a complement to cloud AI rather than a lesser alternative.

By the numbers

• macOS 15 or later is the minimum requirement for hybrid compute to run on Apple silicon Macs.
• Two subscriber tiers — Pro and Max — gain access alongside opted-in enterprise customers.

The Renascence take

The headline feature here isn't really "privacy" — it's continuity. Most confidentiality tools force a hard switch: either the whole task stays local (and gets dumber) or it goes to the cloud (and gets riskier). Perplexity's bet is that users won't tolerate friction, so the system needs to make the security decision invisible and automatic.

The real design insight is behavioral, not technical: people don't trust systems that ask them to choose between "good" and "safe" mid-task, so they either avoid the tool or route around its safeguards. By making the cloud-to-local handoff automatic and context-preserving, Perplexity is removing a decision point that most users would get wrong under pressure anyway. Any operator building or buying agentic AI should ask not "is this compliant?" but "does this force my employees to make a trust trade-off in the moment?" — because that's the question that actually predicts adoption and shadow-IT workarounds.