Lutflow joins NVIDIA Inception

The news

On October 14, 2026, NVIDIA accepted Lutflow into NVIDIA Inception, its program for startups. Lutflow measures the cost of language-model inference from the counters of the server that runs it, and much of that work happens on NVIDIA GPUs: we measured the most recent use case on this blog on an NVIDIA L4.

What NVIDIA Inception is, according to NVIDIACITED

NVIDIA presents Inception as a free program that guides AI startups through its platform and ecosystem. It charges no application or membership fees, takes no equity, has no deadlines or cohorts, and accepts startups at any funding stage. According to its official page, the benefits fall into these groups:

• Training and guidance: free self-paced technical courses and discounted instructor-led workshops, recommendations of tools and resources, full access to the NVIDIA Developer Forums and a monthly members’ newsletter.
• Tools and offers: SDKs, model libraries and developer platforms; preferred pricing on selected NVIDIA hardware and software (the page notes it works as a rebate on the supplier’s price and only for a subset of enterprise GPUs); offers from other companies; and cloud credits from NVIDIA and other providers.
• Investor access: exposure to NVIDIA’s VC network through Inception Capital Connect, based on eligibility, and networking events.
• Brand and reach: official badges, co-branded social content and event materials.

The same page says what it does not include: membership gives no special access to GPUs that are not available. It also publishes no amounts for the program in general, so we do not quote any.

How we plan to use itPROPOSED

What follows is a plan. None of it is done yet.

1. Add the physical GPU next to the server counters

Today the Lutflow SDK measures from the counters the inference server publishes: tokens, requests, KV cache usage. It reads no GPU sensor, and every report says so. In our AWS Community Day Paraguay talk we put it in one line: the physical GPU and answer quality need other instruments.

For the physical GPU, the natural instrument is NVIDIA DCGM (Data Center GPU Manager), which exposes the activity, memory and power of each GPU. The plan is to study how to place that telemetry next to the server counters, with separate labels: chip activity is not throughput, and neither one, on its own, is efficiency. To do it well we will lean on the technical resources and developer forums the program gives access to.

2. Training

The NVIDIA Deep Learning Institute courses included in the program are a structured way to go deeper into inference on GPUs, which is where our problem lives: how much useful work comes out of a GPU paid for by the hour.

3. Access to NVIDIA technology

The program’s SDKs, libraries and developer platforms help us validate our measurements on the same stack used by people serving models on NVIDIA GPUs. If the cloud credit offers apply to our case, we will use them to repeat the measurements on more configurations, always with their labels.

What does not change

• NVIDIA Inception membership describes program participation. Lutflow remains responsible for its product and its published numbers.
• Every number we publish keeps its label (MEASURED, MODELED, CITED) and its limits.
• We use the official member badge as NVIDIA provides it and following its brand guidelines: unmodified and never larger than the Lutflow logo.

NVIDIA, NVIDIA Inception and their logos are trademarks of NVIDIA Corporation.