From Simulation to Production: What It Actually Takes to Make an AI Factory Operational
September 24, 2026
AI Factory
Introduction
That’s the subject of “From Simulation to Production: How to Make AI Factories Operational,” the latest episode of our series The Human Voice in the Age of AI. Nicolas Kremer, CTO and Co-Founder of Polarise, David Iles, Senior Director of AI Networking Solutions and Ecosystems at NVIDIA, and Lukas Gentele, Co-Founder and CEO of vCluster, joined our own Vishal Shukla, Co-Founder and CEO, and Ilona Gabinsky, VP of Marketing, for a conversation about what separates a factory that works in a lab from one that runs reliably in production.
What Actually Makes an AI Factory Production-Ready?
David laid out a clear definition from NVIDIA’s side. An AI factory isn’t production-ready until compute, network, storage, security, and operations have all been tested to work together as one system. He framed it as three tests: can you deploy it predictably, can you operate it reliably at scale, and can you make changes without putting production at risk. He also pointed out that an AI factory’s value isn’t measured the way a traditional data center’s is, it’s measured in productive GPU time, and the metric for that is token generation.
What’s the Hardest Part of Scaling a NeoCloud?
Nicolas described the moment this becomes real for his own customers. “There’s a lot of people coming to me and telling me, yeah, I have this great solution,” he said. “But then I ask them, okay, how do I scale that across 10,000 different nodes? And then it gets interesting, right? Because that’s the scale we’re talking about. It’s not just having something run in a lab.” He also made the case that no single company can solve this alone: strong partners are what let a NeoCloud deliver at real scale and quality, rather than trying to build the full stack itself.
Why Is Automation at the Software Layer Non-Negotiable at That Scale?
Lukas explained why vCluster’s own focus stays deliberately narrow. vCluster isn’t building data centers, it’s solving the software layer once the factory, networking, and storage already exist. “Automation at the software layer is really the solution for a lot of these things,” he said. Even in a worst-case scenario, like a grid outage taking a data center offline, things need to come back up in an orderly, structured, predictable fashion, not through someone manually SSHing in and running scripts by hand.
How Does the Network Stay Reliable as an AI Factory Scales?
Vishal spoke to this from Aviz’s side of the stack. “That’s where predictable network automation comes into the picture, designed to get the best ways to do multi-tenancy at scale,” he said. As networks evolve from scale up to scale out to scale across, reference architectures define how multi-tenancy should work for a given factory, and predictable automation is what carries that design into a running system without it becoming unmanageable.
What’s the One Piece of Advice Each Guest Would Give Teams Building an AI Factory Today?
The panel closed with each guest offering their single best piece of advice. David’s was to stick with proven reference architectures rather than treating your factory as a science experiment, and to try NVIDIA DSX Air directly, since pre-built AI factories can be up and running there in minutes. Nicolas encouraged teams not to be afraid to break things, since AI infrastructure is new enough that everyone is still learning how to operate it at scale. Lukas summed up the through-line of the whole conversation in three words: automate, test, assume failure. And Vishal offered a three-piece framework for moving fast without losing control: process, meaning your reference architecture; people, the right experts who know where the real risk starts; and a predictive platform, the automation built from what the first two pieces teach you.
Listen to the Full Conversation
There’s more in the full episode than fits here, including a deeper look at how multi-tenancy gets defined differently across storage vendors, NeoClouds, and enterprise customers, and why that inconsistency is itself one of the harder problems in building an AI factory.
Frequently Asked Questions
What does it mean for an AI factory to be production-ready?
According to NVIDIA’s David Iles, an AI factory is production-ready when compute, network, storage, security, and operations have all been tested to work together as one system. The three key tests are whether you can deploy it predictably, operate it reliably at scale, and make changes without putting production at risk.
How is an AI factory’s value measured?
Unlike a traditional data center, an AI factory’s value is measured in productive GPU time, with token generation as the core metric.
What is the biggest challenge when scaling a NeoCloud?
The hardest part is scaling a solution across thousands of nodes, not just getting something to run in a lab. No single company can solve this alone, which is why strong partners are essential for delivering at real scale and quality.
Why is software layer automation non-negotiable at scale?
Manual processes cannot reliably manage recovery and operations across a large AI factory. Even in scenarios like a grid outage, systems need to come back up in an orderly, predictable fashion rather than through manual intervention.
How do you keep a network reliable as an AI factory scales?
Reference architectures define how multi-tenancy should work across scale-up, scale-out, and scale-across environments, and predictable network automation is what carries that design into a running system without it becoming unmanageable.
What is the single best piece of advice for teams building an AI factory today?
The panel offered several: stick with proven reference architectures, be willing to break things since the field is still maturing, and follow a three-part framework of process, people, and a predictive platform to move fast without losing control.
Listen to the full episode, “From Simulation to Production: How to Make AI Factories Operational,” to hear the entire conversation.
