Why we build our own hardware
Software that must run fully offline, on site, on a power budget, cannot treat hardware as someone else’s problem. The device is where the promises get kept.
A promise like "fully offline" or "the data never leaves the site" is not a software feature. It is a property of the whole machine: the compute, the storage, the power draw, the failure behaviour when the grid browns out at two in the morning. If you do not control the hardware, you are asking the customer to assemble your promises themselves.
So Vatar ships as a device. Small enough to sit in a rack or on a shelf, frugal enough to live on the backup power a site already has, and complete enough that plugging in the data sources is the whole installation. The models are tuned to the silicon, the storage is sized to the retention the customer actually needs, and when something fails it fails visibly and locally, not silently in someone else’s region.
Building hardware is unfashionable in software companies, and for good reason: it is slow, it is capital-hungry, and it punishes mistakes. We do it anyway, because the environments we build for leave no choice. The places that most need intelligence infrastructure are precisely the places where you cannot assume a data centre, a stable grid or a fast link. The device is where the promises get kept.