Perplexity launched Portable Computer today, a locally-run version of its AI agent platform that operates entirely on user-owned hardware with no cloud dependency or token costs. The product rolls out initially on Nvidia's DGX Spark desktop supercomputer and Linux machines equipped with Nvidia RTX GPUs, developed through a direct partnership with the chip maker.
Portable Computer represents a strategic pivot for Perplexity. The startup's core Computer platform, which gained traction as an enterprise agent capable of autonomous web browsing and task execution, historically relied on cloud infrastructure and token-based pricing. By shifting computation onto local hardware, Perplexity eliminates per-query costs while offering users complete data sovereignty. No API calls leave the device. No usage metrics feed back to Perplexity's servers unless the user explicitly opts in.
The timing matters. As large language models grow more capable and compute requirements stabilize, the economics of local inference are improving. Nvidia's RTX GPUs and the DGX Spark (a $35,000 desktop workstation designed for AI researchers and engineers) provide sufficient computational muscle to run capable language models on-premises. Perplexity's decision to launch here positions the startup directly in the "AI on your machine" movement that rivals like Anthropic, OpenAI, and startups building offline-first tools are exploring.
The partnership with Nvidia goes deeper than platform support. Nvidia likely optimized its CUDA libraries and TensorRT inference engine for Portable Computer's specific workloads. This kind of hardware-software co-optimization has historically given Nvidia leverage in new computing paradigms. For Perplexity, the partnership signals Nvidia's confidence in the startup's agent approach and locks in an important distribution channel through Nvidia's developer community and enterprise sales motion.
Enterprise adoption drivers are clear. Companies handling sensitive data, financial institutions managing proprietary information, and government agencies operating in restricted networks all face constraints that cloud-based AI agents cannot satisfy. Portable Computer removes that friction. Users retain full control over model weights, prompts, and outputs. Compliance becomes easier. Latency drops to network-local levels.
The zero-token-cost model also reshapes Perplexity's competitive positioning against OpenAI's ChatGPT, Claude, and other API-first models. Instead of paying per query, users pay once for hardware and model weights. For high-volume use cases, this arbitrage can be dramatic. A financial services firm running thousands of daily agent queries sees costs collapse from cents-per-query to electricity-only math.
Perplexity's execution on this front matters because local AI agents remain immature. Most agentic systems require cloud orchestration, external tool integration, and constant connectivity. Building a fully self-contained agent that can autonomously browse, reason, and act while running on consumer or prosumer hardware is harder than it sounds. If Perplexity pulls this off cleanly, it gains a defensible advantage in the emerging local-AI-agent category.
The DGX Spark is a niche entry point. At $35,000, it targets researchers and enterprises with real budgets. Nvidia RTX GPUs in existing machines offer a much larger addressable market. Perplexity's roadmap likely includes expanding to consumer-grade Nvidia hardware like RTX 4090 cards and eventually to other GPU vendors if local inference becomes table stakes. This launch is the beachhead.