You are pricing a machine to run a 2.8 trillion parameter model inside your own office.
The new tinybox Red Kimi comes back at $30,000.
The parts list reads well.
A 32-core AMD Epyc Genoa processor with 32 GB of DDR5 behind it.
Four AMD Radeon AI Pro R9700 cards, 32 GB of GDDR6 each, 128 GB combined and 2,560 GB/s of memory bandwidth.
Then sixteen 1 TB NVMe SSDs in RAID, 238.4 GB/s of storage bandwidth, wired up by a custom driver that lets the AMD GPUs read those drives directly.
That driver is the clever part.
Moonshot AI's Kimi K3 does not fit entirely in 128 GB of GPU memory, so the design pushes weights out to flash memory and pulls them back fast enough to keep the model answering at the 40 tokens per second Tiny Corp claims.
Good engineering.
Fixed price, though.
$30,000 for the box, plus power, plus floor space, plus a machine shaped around one model that will be replaced.
The other way to buy compute is by the hour.
An RTX 3090 runs $0.15/hr at BHK Cloud.
Book a weekend to test your own model, add cards for the two weeks the job needs them, and leave the hardware lifecycle to someone else.
Buy Now Pay Later, zero upfront, pay after one month.
What did your last GPU box cost, and how many hours did it actually run last month?
https://ai.bhkcloud.com/?utm_source=linkedin&utm_medium=social&utm_campaign=bhk_social_2026w44&utm_content=evening
Originally posted on linkedin_personal