Perplexity's Portable Computer runs AI agents entirely offline
Perplexity and Nvidia are making the most aggressive push yet to pull AI agents off the cloud: a desktop app where the model, your files, and the work itself never have to leave the machine — and the token meter never moves.
Perplexity launched Portable Computer on Monday, a version of its agentic Computer platform that runs entirely on local hardware — starting with Nvidia's DGX Spark desktop machine and Linux PCs with RTX GPUs carrying at least 24GB of VRAM. Built with Nvidia, it bundles the model, agent harness, tools, connectors, and a security sandbox into one install, replacing what has traditionally been an evening of duct tape: downloading weights, standing up an inference server, wiring tools together. Work done locally burns no billing credits, and every task starts on-device by default — the system asks before sending any single step to a frontier cloud model. "We've basically brought the exact same UI to a fully local app," Perplexity's VP of engineering for infrastructure, Nate, said at a press briefing.
The pitch is really about economics. Chat was bursty — ask, answer, done. Agents are always-on and insatiable, and in the cloud that hunger shows up as a bill. On hardware you already own, the marginal cost of tokens approaches zero; Nvidia's director of developer technology argued that's exactly what makes agents practical on a desk instead of in a data center. In one demo, an agent running a 27-billion-parameter Qwen model combed through a folder of tax documents with the credits counter "just parked at zero" — the kind of sensitive financial work many users hesitate to hand to any cloud service.
The benchmarks are Perplexity's own, so treat them accordingly: on the company's internal 53-task knowledge-work suite, the local setup scored 82.6% versus roughly 74–78% for two open-source harnesses running the identical Qwen 3.8 27B model. The more commercially interesting number is hybrid escalation: on a hard coding benchmark, fully-local execution hit 59.6% for free, routing hard steps to a Claude Opus 5 advisor reached 73.0% at about $0.42 per task, and pure frontier ran 82.4% at $0.65 — recovering around three-fifths of the frontier gap at two-thirds off the cost. Before any escalation, a privacy classifier shows exactly what would leave the device, and the remote model returns advice only, never touching local files or tools.
It lands amid Apple's M5 Ultra Mac Studio push — we covered this morning's launch in Apple's new Mac Studio and Mac mini are built for local AI — and together they mark local AI graduating from hobbyist project to shipping product category. The caveats are real: Linux only for now (Windows follows in September), no Apple silicon, and compact models still trail the frontier on hard reasoning. But if agent workloads keep scaling, "runs on the machine you own" stops being a niche perk and becomes the default privacy-and-cost story.
What to watch: whether Portable Computer ships preinstalled on future DGX Sparks — Nvidia wouldn't rule it out — and whether enterprise buyers bite on zero-marginal-cost agents.
Local agents with no per-token bill: would you trust your sensitive files to a box on your desk, or is the cloud still worth it? Tell us in the comments.
AI safety startup Alice raised $140 million led by Apax Digital, valuing the company at roughly $700–800 million as it scales from content moderation into guarding the models themselves. Formerly ActiveFence, the Tel Aviv and New York company spent nearly a decade tracking fraud, manipulation campaigns, and abuse for platforms like Google, Meta, TikTok, and Amazon — and now sells that playbook to AI labs, stress-testing models against jailbreaks and prompt-injection attacks before release, then supplying continuous red-teaming and runtime guardrails after deployment. CEO Noam Schwartz says Alice protects eight of the world's ten leading model developers and is approaching $100 million in annual recurring revenue. The take: guardrails are quietly becoming infrastructure, and a near-decade-old abuse-intelligence dataset turns out to be a moat.
Waymo will bring robotaxis to Munich, its first EU market, targeting commercial rides by the end of 2027. The Alphabet unit starts mapping and testing with small fleets of Jaguar I-Paces and human drivers within weeks, working with Germany's Federal Motor Transport Authority under the country's Level-4 framework. London was already lined up; Brussels' approval process is the real gatekeeper here. Waymo currently runs over 3,500 vehicles across 11 US cities and more than 500,000 paid trips weekly.
Which matters more for agents' next chapter — cheaper local hardware or stronger guardrails? Tell us in the comments.
