Ghost raises $11M for a $3,499 AI computer that isn’t fully offline
Ghost came out of stealth on Monday with an $11 million seed round led by Andreessen Horowitz. The product is a $3,499 computer called Core, pitched as a machine that runs your AI agents at home instead of in someone else’s cloud. Its own privacy policy, effective 5 October 2026, is narrower than that pitch. The policy describes a Ghost-run gateway that forwards web search and phone call requests, including personal details from the task, out to third-party providers.
That second data path didn’t make the coverage. SiliconANGLE reported that Core’s agents reason on private data and perform tasks “without ever sending so much as a byte of information to the cloud,” in its account of the device. TechCrunch reported 19-year-old founder Zain Javaid saying all personal information is encrypted so no cloud provider or third party, including Ghost, can meaningfully access it. Both are fair readings of what Ghost says about inference, which is the part that holds up.
The privacy policy adds the part that doesn’t. It says Ghost-managed services let you turn on features such as phone calls and web search without obtaining a separate API key. The request then travels from Core to Ghost’s managed-service gateway, and on to the provider. Ghost says that gateway saves none of that content in databases, logs or backups.
It also says what it can’t promise.
Ghost’s decision not to store forwarded content is not a guarantee that the provider does not store it.
Ghost and Core Privacy Policy, effective 5 October 2026
That matters because what leaves the house is the user’s own task text, not telemetry. The policy does offer two exits. You can plug in your own provider keys so Core calls the provider directly, and you can swap Ghost’s remote-access relay for your own network such as Tailscale. On local inference the policy backs the pitch, stating that Core does not use a remote AI model as a fallback for its conversational AI or memory.
The Core FAQ is blunter about everything else, saying you’ll still need an internet connection for web search, email and over-the-air updates.
The spec sheet lists a fourth model
Ghost’s product page also carries hardware detail neither write-up printed, plus a model both of them missed. They each listed three preinstalled models. The page lists four.
| Component | What ghost.ai lists for Core |
|---|---|
| GPU | Nvidia RTX PRO 4000 Blackwell SFF Edition, 24 GB GDDR7 ECC, 432 GB/s |
| CPU | AMD Ryzen 5 7600, 6 cores and 12 threads |
| Memory | 64 GB DDR5 |
| Storage | 1 TB NVMe SSD |
| Models | Qwen 3.8-Next, Qwen 3.8-27B, Gemma 4-31B, Muse-Glimmer-30B |
| Price and batch 1 | $3,499, ships 31 October, 30-day return, 1-year warranty |
The fourth one is Muse-Glimmer-30B, which Hugging Face lists under the verified Meta Inc. organisation on an Apache 2.0 licence. So Core ships with a Meta model too, not only the Alibaba and Google ones SiliconANGLE named. SiliconANGLE also printed the second Qwen build as “Qwen-3.8-2.7B”, where Ghost’s page says Qwen 3.8-27B.
Memory is where the sheet gets interesting. Nvidia’s page for that card confirms 24 GB of GDDR7 with ECC, 432 GB/sec of bandwidth and a 70 W power ceiling. It also footnotes the 770 TOPS figure Ghost quotes as theoretical FP4 using sparsity.
Muse-Glimmer’s model card reports 29.78 billion parameters in BF16, which is roughly 60 GB of weights by our arithmetic. That’s more than twice the card’s memory, and Ghost’s page doesn’t say what quantisation it ships.
The FAQ sells it to developers, quietly
One FAQ answer reframes who Core is for. It says Core exposes an endpoint compatible with the OpenAI Responses API, so its local models can drive any app that accepts a custom endpoint. The answer promises “no per-token inference fees” for that use. That’s the same arithmetic behind Perplexity’s Portable Computer, which charges nothing for work the device does itself.
The investor list differs from the coverage as well. Ghost’s about page credits Andreessen Horowitz, Abstract Ventures, Audacious, SV Angel and Z Fellows, while both write-ups named Nova and neither named Z Fellows. The firm’s interest tracks its $1.1B Machine Age Fund, whose announcement named home AI appliances as a target. The policy also names the legal entity behind the brand, ZMJ, Inc., at 325 9th Street in San Francisco.
Two things are worth watching from here. SiliconANGLE reported that batch one sold out in hours. Yet ghost.ai still showed “First batch open now” and free shipping for batch 1 when we read it on 6 October, with 31 October in the FAQ as the ship date. Ghost has separately invited independent security firms to audit the gateway and the relay, and that audit is the one test that would settle whether the second data path matters.
Get the daily rundown
One email each weekday with the AI news that matters, every claim linked to its primary source.
Free, one email each weekday, unsubscribe in one click. We never sell or share your address.
