Press & Recognition

RakuAI joins NVIDIA Inception. Here is what we are doing with it.

·Kevin Griffin

This week NVIDIA accepted RakuAI into the NVIDIA Inception program. It is a free program for AI startups - we are not paying for it, and NVIDIA does not own equity in RakuAI - but it unlocks four things that meaningfully change the timeline for our spatial pipeline: NGC catalog access, partner cloud credits, the Deep Learning Institute, and the global network of more than 19,000 AI startups already inside.

Why this matters: we are bringing NVIDIA-accelerated 3D Gaussian splatting to Raku Capture, the part of RakuAI that turns a phone scan of a room into an explorable splat. Today that pipeline runs on CPU - COLMAP on Azure Container Apps, producing sparse point clouds in about twenty minutes. With NVIDIA-accelerated splat training, the same scan lands an explorable splat in two to three minutes. That is the difference between an interesting tech demo and a consumer-grade capture experience, and Inception is how we accelerate that jump.

From a phone scan to an explorable splat CPU SfM today (~20 min) - GPU splat training (~2 min) via NVIDIA Inception cloud credits Scan CPU COLMAP Azure Container Apps Sparse PLY ~20 min today NVIDIA-accelerated splat GPU trainer - NGC base image 3D Gaussian splat output ~2 min target via Inception partner cloud credits

What Inception actually gives a startup

The four things we care about:

  • NGC catalog. Pre-built containers for CUDA, RAPIDS (we will use cuDF for spatial queries), Triton, and the building blocks of a GPU splat trainer. Free, automatic for members. We base the reconstruction worker on an NGC image and skip several days of dependency wrangling.
  • Partner cloud credits. Cloud credits across Microsoft Azure, AWS, and Nebius. The Azure piece runs on the same Azure stack the rest of RakuAI uses, and the broader credits give us room to scale GPU capacity as the capture pipeline proves out.
  • Deep Learning Institute. Self-paced courses on fundamentals of accelerated computing with CUDA, RAPIDS for accelerated data science, and Triton for model serving. We will burn through the relevant three or four.
  • Networking and brand assets. The Member badge that now lives in the footer of every page on rakuai.com. Curated VC introductions. The kind of relationships that compound over a year.

The exact piece this unblocks

If you read our earlier post about the team-of-agents architecture, you know we have been treating reconstruction as a back-end detail that AI assistants drive through MCP. The capture pipeline that runs today does its first stage on CPU: COLMAP produces a sparse point cloud, and the viewer renders it the same way the GPU-trained splat will. The orchestration is in place; the next step is richer geometry.

The second stage is dense 3D Gaussian splat training on GPU, which we base on the open-source Brush trainer. The Inception cloud credits let us bring that stage online and validate it end-to-end with real capture traffic before scaling GPU capacity further.

What changes in the product this month

Three things, in order:

  1. The NGC base image lands in the reconstruction worker. We move the worker onto an NVIDIA CUDA base image. Side effect: even the CPU path becomes faster because we get NVIDIA's optimized math routines for free.
  2. GPU splat training comes online on Azure. We bring up GPU capacity, install the splat trainer, and route opted-in users to the GPU backend. The CPU path stays for everyone else while the GPU path rolls out in early access.
  3. The iPhone capture experience gets faster. As GPU splat training rolls out, reconstruction completes while the user is still standing in the room they scanned. Whole different feel.

The NVIDIA Inception disclaimer

NVIDIA Inception is a free program for AI startups. RakuAI's membership does not imply that NVIDIA endorses RakuAI's products, nor that we are affiliated with NVIDIA beyond the program itself. The brand badge that now appears on rakuai.com is the official Member badge we are entitled to use under the brand portal's guidelines; we have not altered it, recolored it, or paired it with the NVIDIA corporate logo.

Try the current pipeline today

Scan a room yourself at rakuai.com/capture-app/ - iPhone Safari or Android Chrome, no install. The reconstruction you get back today is the CPU version. As GPU splat training rolls out in early access, you will be among the first to compare them.

If you are building something spatial and want to compare notes on GPU splat reconstruction, mail partners@rakuai.com. We will be at Microsoft Build next week - NVIDIA is a partner there - if you would rather catch us in person.

← All posts