Ghost Wants AI Agents to Have Their Own $3,499 Computer
Ghost came out of stealth on October 5, 2026 after a year of building, announcing an $11 million seed round led by Andreessen Horowitz to ship Core, a $3,499 personal computer built for AI agents. TechCrunch
The company was founded by Zain Javaid, 19, who serves as founder and CEO. The founding team includes Nicholas Chua, 19, Yifei Chen, 24, and Gautam Sharda, 23. Other investors in the round include Abstract, Audacious Ventures, SV Angel, and Nova.
Core is priced at $3,499. Pre-orders for the first batch opened October 5, 2026, with shipping scheduled for the last week of October. The machine bundles an Nvidia RTX Pro 4000 SFF Blackwell GPU, the specialized processor that handles AI calculations, into that price. It has no screen and is accessed through a phone or app. Ghost builds the units at its San Francisco facility.
Core ships with three AI models preinstalled: Qwen-3.8-Next, Qwen-3.8-27B, and Gemma-4-31B. Those are the ready-made AI systems the computer can run out of the box. Users can install other models from Hugging Face, a public library for sharing models, or load their own trained models.
The broader context here is where this kind of work happens. Cloud services rent out shared computing that can expand and shrink on demand. A dedicated local machine keeps data, files and credentials in one place and gives the owner direct control over which models and tools run. For long-running agents that need access to private information, that difference is practical, not theoretical.
In my view, Ghost is testing whether developers and power users will pay workstation prices for independence. The $3,499 is an upfront purchase instead of paying a little at a time for cloud use. We have seen this tradeoff before with early PCs and workstations, when buyers accepted higher upfront cost for predictable access and the freedom to change the system without asking permission or paying per use.
Looking at what this means for builders, the open model policy is the detail to watch. Preinstalled models reduce setup work. Support for Hugging Face downloads and custom models leaves room to experiment. That flexibility counts for more than any one model, because the tools around agents change fast and teams need to swap models without buying new hardware.
Worth flagging for enterprise readers, the screenless, phone-managed design solves one set of problems and raises another. It removes monitors and peripherals and fits work that happens in the background over time. It also places more responsibility on remote management, access control and regular updates, the normal upkeep for any machine that stays on all the time, now sitting in a home or office.
There is a longer arc worth keeping in mind. Computing has often moved from shared systems to personal ones once costs allowed, from timeshares to PCs, and from hosted servers to phones that carry real computing power. Local hardware for agents follows that pattern if models keep getting more efficient and if people prefer to keep their working context close. Ghost will not answer that question with a first batch. It gives engineers something concrete to test, and useful categories often start that way.
