NVIDIA Inception, and saying the word out loud

SoftBacon Software is now a member of the NVIDIA Inception program. We applied on June 4 and were approved on June 10 — six days. Inception is NVIDIA's program for startups building on accelerated computing: no equity, no fee, and a door to the things an independent lab cannot easily buy its way into — DGX-class hardware access, engineering contact, and a channel to be evaluated on the work.

Why we applied

This lab's standing claim is that the architecture around models matters more than the size of any one of them — and that claim is only as good as its portability. Everything published here so far runs on Apple Silicon unified memory. The honest next test is the same harness on different metal: NVIDIA's unified-memory machines are the natural second substrate, and one NVIDIA seat is already in production — a Jetson running a quantized model at roughly 0.8 joules per token on twenty watts, the smallest member of the mesh.

The application itself was written under the same rule as these notes: claim only what runs. What runs on NVIDIA today: the Jetson seat, a consumer GPU serving generation jobs, cloud H100s for training runs. What doesn't yet: DGX-class demonstrations — that's what the membership is for.

The other thing that happened today

Filling out a member profile forces a question every small operation dodges: what are you? We have been calling this site "an independent research publication," which was true and slightly evasive — a publication is a thing a lab produces, not a thing a lab is. As of today the description reads what it should have read for a while: an independent AI research lab. One human, a persistent local AI squad, owned hardware, instruments we calibrate in public, and corrections committed in git next to the claims they correct.

External recognition doesn't make that true — the practice does, and the practice predates the email. But there is a real difference between describing yourself and being described, and this week had both: a platform vendor shipping the architecture thesis on a keynote stage, and now a hardware vendor reading the receipts and opening the door.

The plan does not change: do good work, make cool stuff, publish the receipts, see what happens. This note is a receipt.