Not every AI startup wants to be a chatbot in your browser. Ghost, a San Francisco startup, emerged from stealth on Monday with an \$11 million seed round led by Andreessen Horowitz, joined by Abstract Ventures, Audacious Ventures, SV Angel, and Nova, along with individual investors including Instacart co-founder Max Mullen and Codecademy co-founder Zach Sims. Its bet: the next wave of personal AI won't run in someone else's data center — it will run in a box under your desk.
That box is Core, a \$3,499 computer built specifically to host always-on AI agents. It has no screen; users connect through a phone or desktop app, with voice mode handled through the app. Under the hood: an Nvidia RTX Pro 4000 SFF Blackwell GPU with 24GB of GDDR7, 64GB of DDR5 RAM, a 1TB NVMe SSD, a 6-core AMD Ryzen 5 7600, and a company-claimed 770 AI TOPS. The GPU is bundled into the price — no separate purchase. Pre-orders opened Monday, with first-batch shipping scheduled for the last week of October. Units are assembled at Ghost's San Francisco facility.
Core ships with three models preinstalled — Qwen-3.8-Next, Qwen-3.8-27B, and Gemma-4-31B — and users can pull others from Hugging Face or load models they've trained themselves. The premise, according to founder and CEO Zain Javaid, is that useful personal AI needs continuous access to the full context of your digital life: desktop files, apps, even smart-home devices. "You can run models on existing hardware, but it's a terrible experience on many levels," said Javaid, a 19-year-old former quantitative trader who founded the company after ChatGPT's 2022 release pulled him out of finance. The founding team — Nicholas Chua, 19, Yifei Chen, 24, and Gautam Sharda, 23 — averages barely legal drinking age.
The privacy architecture is the real sales pitch. Models and personal memory run locally, and personal information is encrypted so that no cloud provider — including Ghost itself — can meaningfully access it. Javaid says the device is designed to keep working even if the company disappears: "All of the source code, logic, and model weights live on the device, not the services." A built-in firewall monitors every outgoing network request, rejecting and flagging any agent action that looks like unauthorized use of credentials or data exfiltration — a feature that reads as a direct answer to the rogue-agent incident stream that has dominated this year's AI security headlines.
The comparison set is telling. Against a Mac Mini, Javaid's argument is specialization: Apple's desktop is general-purpose and needs configuring, while Core arrives with models ready to run. Against the cloud, the argument is ownership: no subscription, no usage meter, no terms-of-service change rewriting your assistant overnight. Updates arrive over the air, Tesla-style, with users choosing when to move to a new model generation.
Whether a consumer will pay \$3,499 up front to escape \$20-a-month subscriptions is the entire business model on trial — and incumbents with far deeper pockets could ship a similar "local agent hub" as a feature. But the timing is not accidental: agent autonomy is rising exactly as trust in hosted AI wobbles. Ghost is betting that the next platform shift is not a smarter model, but a place to run it that you actually own.
Comments (0)
Log in to join the discussion
Log InNo comments yet