.png){kind=logo}
Companies across industries are building generative AI, machine learning, and automation proofs of concept at record speed. Many of them work well in a controlled setting. Far fewer survive the real demands of enterprise production, such as complex infrastructure, security rules, legacy systems, and cost pressure.
In today’s Analytics Insight Podcast, host Priya Dialani spoke with Rhys Oxenham, Vice President and General Manager of AI at SUSE, about why so many enterprise AI projects stall before they reach production.
The conversation covered the production chasm, open infrastructure, AI sovereignty as a spectrum, the AI factory model, the India AI Mission, and the one barrier he would remove. Here are the key excerpts from the interview:
SUSE is a global leader in enterprise open source. For well over 30 years, we have focused on stable, resilient platforms for mission-critical environments. Many people know us for enterprise Linux and our partnerships with companies like SAP. Today we also offer a cloud-native open infrastructure platform that uses Kubernetes across data centers, public clouds, and the edge. I joined about three and a half years ago, worked in edge computing and telecommunications, and took over the AI business earlier this year. My job is to turn our hardened open source platform into a secure, sovereign base for what we call private enterprise AI.
A pilot on a laptop is easy to demo. Production brings scale, security, tooling, steady operations, and many different teams. We call this gap the production chasm. The rate of change is one cause. A new model or framework appears almost every week, so companies worry that what they build today will be outdated in months. Isolated pilots and scattered tools make things harder. Pay-as-you-go frontier models can lead to uncontrolled costs, and recent export control changes showed that access can be lost. Shadow AI adds data risk when teams feed corporate data to outside tools. Rules also differ by region, and some workloads must stay within a certain jurisdiction.
Boards want real business results from AI, but the people on the ground often lack approved tools. That gap pushes them toward third-party services. Open infrastructure gives customers freedom, flexibility, and the ability to pivot. I call this operational resilience. It means a transparent, inspectable stack built on open standards at every layer. That covers where the workload runs, which hardware and silicon are used, and which application stacks are consumed. It is not about being locked into one stack. It is about deploying intelligence wherever it suits the business.
Sovereignty is not a compliance checkbox, and it is not only data governance. I see it as a spectrum that runs from silicon and operating systems to runtimes and models. Total isolation is not possible today. The goal is to match technology choices to the organization’s risk profile and always keep a plan B. A company should be able to move between platforms and protect its intelligence from export control changes or foreign access limits.
An AI factory is not about buildings or gigafactories. It is about running AI workloads at industry scale. I describe it as an assembly line for applications and the infrastructure under them. It includes GitOps workflows, zero-trust security, and a developer-to-production pipeline. Standard tooling removes technical debt and shortens time to value. It supports generative AI use cases such as chatbots and retrieval augmented generation, along with MLOps and agentic workloads that need policy control and human oversight. Private enterprise AI is not private cloud. It means running intelligence within the customer’s own control, where it can never be taken away.
I visited India a few weeks ago, and I see the mission as a major catalyst. Indian enterprises have faced GPU procurement problems and high costs. The mission aims to deploy thousands of GPUs through public and private partnerships, which opens compute access to enterprises and startups. It also supports local AI gigafactories and regional cloud capacity, so companies need not depend only on hyperscalers or hosting outside India. The effort also covers data quality, local talent, and open datasets tuned for the Indian market. GPU capacity is only half the battle, though. Enterprises need an open infrastructure platform to deploy, scale, and own their workloads from day one.
I would remove the false trade-off between speed and control. Teams feel pressure to adopt whatever technology is available fast, and that risks security and intellectual property loss. Control protects resilience and ownership, but it can feel slow. This choice creates shadow AI, uncontrolled costs, and governance failures. When teams standardize on runtimes built for private enterprise AI, they can move fast inside their own perimeter and keep full ownership of their intelligence and data.
Listen to the full discussion on the Analytics Insight Podcast.