East Wenatchee, Washington · Established 2018 · Tuesday, 08 September 2026
Compute, milled on the Columbia
A mining farm went up in East Wenatchee in 2018. It cut Ethereum until the chain stopped paying for the work, then stood cold for two years. Through 2025 we rebuilt it, rack by rack, into a plant that reads documents, draws pictures and makes film.
One hundred and sixty machines carry nine hundred and sixty AMD RX 580 cards, six to a machine, on one flat segment with a single switch above the row. The hosts are Celerons with eight gigabytes of memory, which is all a card‑feeder needs. Power comes off the river at seven and a half cents the kilowatt hour.
A nine hundred and sixty‑six page ophthalmology textbook goes in as one file. It is cut into pages, the pages are dealt across the fleet, and the text comes back assembled in seven minutes and thirty seconds. The same shape of machine draws images and renders video; only the model on the card changes.
Nothing here came off a shelf. The inference server is built from source against Mesa and the Vulkan SDK, because the vendor compute stack drives one card per machine on this silicon and Vulkan drives all six. Where the tooling was wrong about the hardware, we fixed the tooling and sent the patch upstream.
| Nodes in service | 160 | six cards to a machine |
|---|---|---|
| Graphics cards | 960 | AMD RX 580, Polaris, 4 GB and 8 GB |
| Power | $0.075 / kWh | Columbia River hydroelectric |
| Cost of VRAM | $80 / 8 GB | against $365 per GB on an H100 |
Plate I
Document extraction
A thousand-page book goes in as one file and comes back as text, because every page is read by a different machine at the same time.
In service
Plate II
Image generation
Batch image work on the same cards, with the styles kept as house recipes rather than as a text box the customer has to fight.
In service
Plate III
Video generation
Clips rendered on racks of cards that were never meant to see a diffusion model, at a price per second that makes a long cut affordable.
In service
Plate IV
Engineering
The cards were fine. The software did not know what it was holding. Most of the speed came from fixing that.
Continuing
A medical textbook, read end to end, on cards bought for mining. The trick is not the model. It is that the book is nine hundred and sixty-six separate jobs.
One attention kernel went from three thousand and twenty-two milliseconds to one hundred and fifty-two. Nothing was optimised. A test was corrected.
The vendor compute stack reports six cards and uses one. The failure is silent, and it decided the architecture of everything we have built since.
We take work that suits the plant: long documents, large batches of images, and video by the minute rather than by the second. Tell us the size of the job and the shape of the output.