Own your AI.

A $200 box that runs a frontier AI model at home. Pay once, keep 20 tokens a second for life, and sell back what you do not use. Nothing you ask ever leaves the house. A solar panel, but for AI.

Reserve Cronus One Watch it run

$200

Pay once. No subscription, ever.

20 tok/s

Yours, for life.

~120B

A real frontier model.

Chip errors: DRC, LVS, antenna.

It runs on the box, not the cloud.

Local inference / live~20 tok/s · 0 bytes sent

Answering from your own files, on the device. Nothing leaves the room. 20.0 tokens a second.

Local only0 bytes leave

Your prompts

Stay on the box. Nobody uploads them. Nobody logs them but you.

Your files

Read and answered on the device. The model sees them. The internet does not.

Your usage

No account. No telemetry. No profile of how you think.

Your model

Yours. Open weights you run, not an API that can change or vanish.

Your prompts never leave the house.

Rent AI and your questions become their training data. When a lab set out to crack Navier-Stokes, it learned from what people had typed in. Cronus One runs the model on hardware you own, so what you ask stays on your desk. No account. No logs. No profile of how you think.

A cell, not a subscription.

Universal constructor / live

A few boxes hold one model and serve it together, and keep going as nodes come and go. generation 1 · nodes live —

A frontier model is too big for one cheap box. So a few Cronus units team up into a cell and run it together, like RAID, but for intelligence. You own a node, and your node gets you 20 tokens a second, for life.

The honest part: the boxes in a cell sit on your own network, not scattered across strangers' homes, so your data stays inside your cell. The compute you do not use does not idle. The cell sells it back, and you get paid. Nvidia calls compute the new asset class. This is how you own a piece of it.

The chip that makes $200 possible.

Pooling ordinary machines only gets you so far. To hit $200, we design our own inference silicon, the compute and the memory it feeds on, with an AI-assisted loop, not a room of veterans. It runs in simulation today, and checks its own answers.

Subsystem	Serial → SRAM + DMA → arithmetic
Memory	6 × 1 KiB SRAM macros, SKY130
Verified	321 transactions, 2,781 checks, all passing
Negative control	A wrong answer is caught, not missed
Global route	0 congestion overflow

See the test and the code

Actual placementsix SRAM macros

The Cronus compute and memory test chip: six 1 KiB SRAM macros placed in the global-route database

Six 1 KiB SRAM macros from the global-route database. Cell footprints, not metal.

How Cronus compares.

Plenty of boxes can run AI at home. None of them do it for $200, guarantee your speed for life, or pay you back for the compute you do not use.

	Cronus One	Mac mini (M4)	Nvidia DGX Spark
Price	$200, one time	From $599	~$3,000 to $4,000
Runs a ~120B frontier model	Yes, as a cell	No, not on 16GB	Yes, 128GB
20 tokens a second, for life	Guaranteed	You tune it	You tune it
Your data stays with you	Yes	Yes	Yes
Sell spare compute back	Yes	No	No
Setup	Plug it in	Install and serve it yourself	Developer box
Built for inference	Custom chip	General purpose	General GPU

Prices are approximate, from public listings. A Mac mini or a DGX keeps your data local too. What they do not do is start at $200, guarantee your throughput for life, or sell your spare compute back.

$200

Buy AI once. Keep it for life.

One payment. Twenty tokens a second, for good, with the compute you do not use paying you back. No subscription can say that.

Reserve Cronus One Read the manifesto