Physivitis Joins NVIDIA Inception: From Simulation to Silicon

We are delighted to share that Physivitis has been approved to join NVIDIA Inception, NVIDIA's programme for deep-tech and AI startups.

For a company like ours, the value here is concrete rather than symbolic. Inception is not an endorsement or an investment; it is a working relationship that gives us access to the resources that matter most at our stage. Chief among those is compute. Through the programme we gain access to NVIDIA's infrastructure to build our AI innovations and, just as importantly, to validate them on the latest-generation hardware they were designed for.

That second point is the one we care about most.

The gap between simulation and silicon

In deep-tech, the distance between a promising result and a proven product is measured in validation. It is one thing to show that an idea holds up in simulation. It is another to prove it on the silicon it was built for, under real conditions, at real scale. That gap is where many good ideas quietly stall, because the hardware needed to cross it is expensive and hard to access.

This is exactly the gap Inception helps us close.

First in line: IGTR

The first innovation we are taking through this process is IGTR, our optimiser for stabilising neural-network training at very low numerical precision, specifically fp8 and fp4.

Low-precision training is one of the most important frontiers in AI infrastructure right now. Training and running large models at reduced precision can dramatically cut the cost, energy, and memory footprint of AI, but it introduces instability that can derail training. IGTR is designed to hold that training stable where it would otherwise break down.

IGTR is validated in simulation today, with strong results. The next step is to prove it on the hardware it was designed for, at scale. Inception puts that hardware within reach, and that is the work now in front of us.

Next: SLATE

The same logic applies to SLATE, our stabiliser for compressed-attention models. Compressed attention is one of the key techniques making large language models faster and more efficient, and SLATE is built to make that compression provably stable rather than fragile. It is another innovation where moving from a validated concept to a hardware-proven result is the decisive step, and another that benefits directly from the compute access Inception provides.

Why this matters for how we work

Physivitis exists to turn rigorous physics and mathematics into practical, commercial technology. Our model depends on moving quickly from idea to working, validated product, because a validated innovation is one industry can actually license and deploy, while an unvalidated one is just a claim.

Access to frontier hardware shortens that path. It lets us replace "this should work" with "here is the result on the silicon," which is the only kind of evidence that earns the trust of serious partners.

What comes next

We will share more as the results land. Our focus over the coming period is straightforward: take IGTR and SLATE from simulation to hardware-validated performance, and report what we find honestly, whatever it shows.

If you are working on low-precision training, efficient AI infrastructure, or the broader challenge of making large-scale AI cheaper and more stable to run, we would be glad to compare notes.

Build. Validate. Accelerate.

Working on low-precision training or efficient AI infrastructure? Let's talk.

Get in touch