Ur AI Joins NVIDIA Inception to Accelerate Sovereign AI Infrastructure

I’m happy to share that Ur AI has been accepted into NVIDIA Inception, NVIDIA’s program for startups building in AI and accelerated computing. Our application was approved on October 26, 2026.

We applied for one reason: to accelerate our research and development on sovereign AI infrastructure, and to bring that infrastructure, along with the tools and resources around it, to our enterprise customers.

I don’t think enterprises should have to choose between capable AI and control over their own data. Inception brings us closer to the hardware and software that make sovereign AI practical, and that lets us build it into Nebula from the ground up.

Why sovereign infrastructure

Our thesis is that enterprises will run most of their AI workloads on locally deployed, fine-tuned models that they own, and use frontier models only where they truly need them. For companies that handle sensitive documents, such as financial institutions, M&A teams and regulated industries, this is not a preference. Data residency, security reviews and auditability decide what they can deploy.

Sovereign AI means the whole stack, from the GPUs to the models to the applications, runs where the customer decides, under the customer’s control. Building that well takes a deep understanding of the compute layer underneath.

Starting with Nebula

Nebula, our document intelligence engine, turns charts, tables and complex business documents into structured Markdown and JSON. It is where our sovereign work is furthest along.

Nebula Sovereign runs on open-weight models that we post-train ourselves. Post-training lifted its score on RCRR, our benchmark for Japanese business documents, from about 67–70 to 92.4, within two points of Nebula Frontier (94.4). Our next goal is to close that gap on all three things enterprises care about: accuracy, cost and speed. Inception’s technical resources, training and tooling give us a faster path to getting there.

Closer to the source

Today, we run on NVIDIA GPUs through private cloud environments with AWS and Microsoft. That works well for our customers, but it puts a layer between us and the hardware.

Working with NVIDIA’s resources more directly helps us learn what GPU optimization our models actually need, and plan ahead for the infrastructure our customers will run tomorrow. It also gives us access to preferred pricing on NVIDIA hardware and software, partner cloud offers, and NVIDIA’s network of venture capital firms.

Our R&D roadmap

Our roadmap has two simple ideas: rent compute first and own it as we scale, and move our R&D from post-training to pre-training. Inception supports both.

• Now, through mid-2027: Post-train vision-language models for document conversion, then LLMs for embeddings, retrieval and reasoning. Optimize how we allocate GPUs and how our models perform on them. Run a domestic sovereign cloud pilot, and bring in our first GPUs of our own for R&D in 2027.
• To early 2028: Pre-train our first small, specialized models, such as handwriting and digit recognition, and prepare for LLM pre-training with data annotation and Japanese-native tokens and embeddings.
• 2029 onward: Pre-train our own LLMs and multimodal models, designed and built in Japan end to end, with our services running mostly on our own GPUs.

Along the way, we will keep publishing our benchmarks and sharing what we learn. You can read more about us on our blog at ur-ai.net/blog.