Perplexity's New AI Orchestrator Decides What Runs On-Device vs. Cloud in Real Time

At Computex 2026, Perplexity CEO Aravind Srinivas and Intel CEO Lip-Bu Tan took the stage to demonstrate a new approach to running AI: a hybrid local/server agentic inference orchestrator. Instead of forcing users to choose between powerless on-device AI or privacy-invasive cloud processing, this software layer automatically dissects a complex task, keeps sensitive parts locked down on your machine, and sends the computationally expensive parts to the cloud. The result is a seamless, single output stitched together from both worlds—and, according to early ecosystem analysis, a potential 30–50% reduction in inference costs .

An 'Air-Traffic Controller' for AI Workloads

Perplexity describes the orchestrator as an "air-traffic controller" for AI tasks, and the analogy fits . When you ask the system to, say, analyze a confidential investment deal, it doesn't just blindly send all your data to a frontier model in a faraway data center.

The process follows four distinct steps:

• Local Assessment: Small models running directly on your PC's NPU or GPU first scan the task to identify any sensitive, confidential, or compliance-relevant data .
• Local Execution: Any subtask touching that sensitive data—like reading a confidential PDF or analyzing a private financial model—is executed entirely on-device. That information never leaves your machine .
• Cloud Delegation: The non-sensitive but computationally intense parts of the job, such as broad web research or complex reasoning that requires a more powerful model, are routed to cloud-based AI agents .
• Recomposition: Once both local and cloud models finish their work, the orchestrator reassembles the results into a single, coherent response.

This isn't a simple decision tree based on file size or task type. As Srinivas explained, it's the orchestrator "reasoning about what work should run on your device and what work should go to agents in the cloud" in real time .

The Key Features Driving Enterprise Adoption

The Computex announcement highlighted several core capabilities that position this orchestrator for enterprise use:

Chip-Agnostic by Design
Even though the stage demo exclusively used Intel hardware, Perplexity was careful to stress that the orchestrator is not locked to Intel's ecosystem. In an interview following the keynote, CEO Aravind Srinivas explicitly called the platform "chip agnostic," designed to work across silicon from different vendors . The initial launch on Intel-powered Windows PCs is a partnership milestone, not a permanent lock-in .

Massive Multi-Model Orchestration
The orchestrator operates within Perplexity's broader "Computer" AI operating system, which can coordinate up to 20 different AI models at once. For any given subtask, the system selects the optimal model based on its capability, speed, and cost—ensuring you're not using a sledgehammer to crack a nut .

A True Privacy Firewall
For regulated industries like finance and healthcare, keeping sensitive data off third-party cloud servers is non-negotiable. The orchestrator acts as a privacy gatekeeper, ensuring that compliance-relevant documents are analyzed locally before any non-sensitive context is ever sent to the cloud for enrichment .

Significant Cost Reduction
Running every enterprise AI query through a cloud-hosted frontier model is becoming prohibitively expensive. By intelligently offloading simpler or private tasks to an on-device NPU, the hybrid approach is estimated to cut inference costs by 30 to 50 percent .

The Intel Partnership: Local Silicon Meets Cloud Brains

The live demo at Computex showcased a laptop powered by Intel's new Core Ultra Series 3 processor, built on the company's advanced 18A process node . The demonstration revealed how a private equity analyst could load a project codenamed "Project Falcon"—complete with NDAs, financial models, and confidential transcripts—and have the local Intel silicon handle the sensitive document analysis while cloud agents simultaneously pulled external market research, all without exposing private data .

Intel framed the collaboration as validation that on-device inference is no longer a novelty. The Core Ultra Series 3's integrated GPU and NPU can handle meaningful AI workloads that were previously data-center-bound, enabling a genuine division of labor that "maximizes token value per watt" . At launch, the hybrid inference capability is exclusive to the Perplexity app on Intel-powered Windows PCs, but the orchestrator's underlying architecture is designed to support a much wider range of hardware going forward .

Why Computex 2026 Marked a Turning Point for AI Architecture

The Perplexity-Intel demo wasn't an isolated experiment—it was the flagship announcement of a dominant theme at Computex 2026: the distribution of AI workloads across edge and cloud.

• The Economics Are Shifting: Running every single agentic AI query in a centralized data center is unsustainable at scale, both in terms of raw dollar cost and physical power consumption. Hybrid execution eases the load on data centers by shifting simple, low-latency tasks to the client device .
• Privacy Demands It: As enterprise adoption of AI accelerates, corporate compliance teams are increasingly refusing to let sensitive data touch external APIs. Hybrid inference makes AI tools usable in these tightly regulated environments .
• Edge Hardware Is Finally Ready: With Intel's 18A chips shipping across hundreds of PC designs, on-device NPUs can now do more than just blur video backgrounds—they can run substantive, intelligent AI inference alongside cloud agents .

Intel used its entire Computex keynote to reinforce this distributed future, announcing complementary partnerships with SambaNova, Vista Equity, and Cambium for cost-efficient inference, and framing hybrid compute as the architectural foundation for the coming "agentic AI" era .

The message is clear: the data center isn't disappearing, but it's no longer the only place intelligence happens. In the hybrid model that Perplexity and Intel are betting on, the future of AI is a conversation between your device and the cloud—orchestrated automatically, intelligently, and privately.