Ghost Core: a $3,499 local AI agent computer, priced vs. APIs

A startup now sells a computer that exists only to run your AI agents. Ghost came out of stealth this week with an $11 million seed round led by Andreessen Horowitz and opened preorders for Core, a $3,499 local AI agent computer with no screen, TechCrunch reports. The pitch: your data and your models stay in one box you own. The real question for a small business is simpler. Does $3,499 of hardware beat a year of API invoices?

What actually happened

The facts, from TechCrunch:

• Funding: $11M seed led by a16z, with Abstract, Audacious Ventures, SV Angel and Nova. CEO Zain Javaid is 19.
• Hardware: an Nvidia RTX Pro 4000 SFF Blackwell GPU, included in the $3,499 price.
• Form factor: a screenless "brain in a box." You talk to it through a phone app.
• Models: ships with Qwen-3.8-Next, Qwen-3.8-27B and Gemma-4-31B. You can load other models from Hugging Face or your own fine-tunes.
• Privacy claim: Javaid says personal data is encrypted so no cloud provider or third party, Ghost included, can meaningfully read it. The company says it is building a firewall that watches outbound network requests.
• Timing: preorders opened Monday; shipping is set for the last week of October.

What we did not see: published benchmarks, memory and storage specs in the report, or any subscription price. Treat the privacy claims as claims until someone independent tests the box.

Why it matters for your business

Do the math before you buy the box. A $3,499 device that replaces $300 a month of API spend pays back in about a year. One that replaces $40 a month is an expensive paperweight. Pull your last three months of model invoices first. If your spend is small and spiky, stay on the API.

Local models have a quality ceiling. A 27B or 31B model does good work on drafting, summaries, tagging and data extraction. It is not a frontier model. The smart setup is routing: cheap, private, repetitive work goes to the local box, and the hard reasoning goes to a cloud API. Your code should not care which one answers.

Version 1 hardware from a seed-stage startup is a risk. If Ghost pivots or folds, you still own a GPU box, which is fine. But if your workflows depend on its app and its agent layer, you inherit its roadmap. Keep the agent logic in code you control, and use the hardware as a model server.

Privacy is the strongest reason, not cost. Client files, HR records and anything under NDA are where local inference earns its price.

Sources: TechCrunch.