A $3,499 screenless computer has one job: running AI agents for you, 24/7

Ghost's first product, Core, is a small box with no screen, no keyboard, and no display output. Its only job is to run AI agents on a person's behalf, around the clock, using a discrete Nvidia RTX Pro 4000 SFF Blackwell GPU and a small bundle of pretrained models that ship in the box. Preorders opened Monday for $3,499, and the first units are scheduled to ship in the last week of October.

The device is reached through a phone or desktop app, never a monitor. Voice mode works only when spoken into the Core app itself. Founder and CEO Zain Javaid, 19, calls the box a "brain in a box": his shorthand for a personal computer built specifically for the agent workload, rather than a general-purpose PC reconfigured for it. The RTX Pro 4000 SFF Blackwell is a workstation card, not a gaming or data-center part, which is the price point that lets Ghost put it in a $3,499 box without subsidizing the rest of the bill.

The software stack is what the bet is actually on. Core ships with three preinstalled open-weight models, Qwen-3.8-Next, Qwen-3.8-27B, and Gemma-4-31B, and lets users load other models from Hugging Face or their own trained weights. A custom browser, filing system, and agent layer sit on top. The pitch is that the user's data, apps, and smart-home context stay on the device, and the agents that touch them run there too, instead of round-tripping to a cloud API. That's the privacy and latency pitch, and it's also the only reason a buyer would pay a $1,000-plus premium over a comparable small-form-factor PC.

Ghost announced an $11 million seed round led by Andreessen Horowitz to fund the launch, with Abstract, Audacious Ventures, SV Angel, and Nova participating. The bet is that dedicated local AI compute is a hardware category of its own, not a Mac Mini with extras. Andreessen Horowitz has backed plenty of category-creation bets that didn't work out, and the firm has also backed plenty that did. The seed is small enough that a16z is treating this as a category test, not a platform bet.

That pitch is fair, and the most obvious objection to it is also fair. The $3,499 price is real consumer money, especially for a first-generation product from a company that has never shipped hardware. The Mac Mini comparison, which Javaid raised himself, is the harder one: a Mac Mini configured for AI costs less than Core at the entry level, runs macOS, and can do everything else a person needs a computer for. Javaid's counter is that Core is built specifically for the agent workload, with the GPU and the software stack tuned for it from the start. That's a real architectural difference, but it's also a difference the buyer has to take on faith until the benchmarks and the third-party reviews land.

The agent reliability question is the one that matters most and the one Ghost controls least. AI agents that book, buy, and message on a person's behalf are still hit-or-miss in 2026, and running them continuously on a dedicated box does not change the underlying model. A brain in a box is still running someone else's weights, on someone else's GPU, with the same failure modes the cloud versions have. Dedicated hardware buys latency, privacy, and uptime, not better agents. If Core's defaults flake on routine tasks, the buyer is paying $3,499 for an unreliable assistant that lives in the closet.

The founding team is unusually young. Javaid is 19, as is co-founder Nicholas Chua; Yifei Chen is 24 and Gautam Sharda is 23. Javaid worked as a quant before ChatGPT's release in 2022 pushed him into personal AI. Youth isn't, on its own, a reason to discount a hardware bet, and the team has hired veteran engineers around the founders, but it's one more unknown on a category-test product from a first-time shipper.

The open-weight model choice is a small tell. Ghost did not train its own foundation model and is not asking the buyer to trust a closed one. Qwen and Gemma are public, can be audited, and can be swapped. That's consistent with the local-compute pitch, and it lowers the lock-in risk for an early buyer who decides Core's defaults aren't for them. It's also an admission that the agent layer is the product, not the model.

Core is the first real test of whether dedicated personal-AI hardware has a market, or whether the category collapses into "a small PC with a good GPU." Preorders opened this week. The first units ship in late October. That's the deadline for finding out.