Skilak joins NVIDIA Inception

The news

Skilak has been accepted into NVIDIA Inception. Thank you to the NVIDIA Inception team for having us.

We started Skilak because using AI should not mean handing over your data. People should be able to run useful models on hardware they control, and know exactly what goes in and out of those systems.

Inception does not change that. It gives us access to technical resources, training, and other startups working on similar problems, and we plan to use all three.

More testing on NVIDIA hardware

The first job is testing. We will run more local-inference tests on NVIDIA hardware with representative, non-sensitive workloads. Loading a model is the easy part. We want to see how it holds up over hours of use, with long contexts, with several people asking questions at once.

We will measure latency, throughput, and memory use, and record where model quality has to give way to the hardware available. We want to be honest about what runs well locally and what does not.

For Skilak Keep, our on-prem AI appliance, this should mean clearer hardware sizing and deployments that go the same way every time. A customer should get the performance they were told to expect when the system is switched on.

Learning from other teams

We intend to explore the program's technical resources, training, and startup community, mostly to learn about inference performance and deployment patterns.

Other teams have hit the same problems we have: limited GPU memory, models that change every few months, setup steps that break, and the distance between a demo and something a firm can rely on every day. We would rather learn from them than repeat their mistakes.

What we learn should show up as documented configurations, better setup guides, and tests anyone can rerun. When there is a chance to ask questions, we will ask them.

None of this involves customer data. The data people trust us to keep private stays out of our testing.

What does not change

Skilak Keep and Skilak Mesh come at the same problem from two sides. Keep brings AI onto hardware you control. Mesh keeps your data out of other people's AI.

Joining NVIDIA Inception does not change what we owe the people who use them. Your AI. Your hardware. Your data. Anything we improve has to keep those commitments intact and easy to understand.