How a packaging company built AI infrastructure without the cloud.
Trivium Packaging makes cans. Aerosol cans, food cans, metal packaging in all shapes and sizes. Not the first industry you’d associate with AI.
Yet today, Trivium runs its own AI platform. Hundreds of employees use it daily. The company has a dedicated AI Center of Excellence. And every piece of data, every model, every integration runs inside its own walls.
It started with one server. No GPU. Ten-minute response times. Here’s how they got from there to here.
Where it started
Somewhere in mid-2022, Trivium’s IT team was in the middle of a global network refresh. New Cisco equipment is replacing old infrastructure across multiple sites.
Then GPT launched. The team started experimenting. They used it to generate switch configurations, prepare templates, validate setups. Work that used to take hours now takes minutes.
That sparked a question. The team was juggling multiple tools every day: security dashboards, firewall consoles, inventory systems. Tabs upon tabs. What if they could build something like ChatGPT, but internal? One interface that connects to all their tools and retrieves information on demand?
They called it Kenny. A nod to the cans they produce.
The first version was rough. A single server. No GPU. Every response took about ten minutes to generate. But hey, it was working.
That was enough to prove the concept. The rest would follow.
Why did they keep it inside
When the team considered where to host Kenny, one question stopped them early.
Their tools had open APIs. Useful for internal automation, but not something they wanted exposed to the public internet. Hosting externally would mean opening doors they preferred to keep closed.
The instinctive reaction: “Why would we even consider hosting this outside of our own organization?”
So they didn’t. They started with a single on-prem server, isolated to only the systems it needed to access. No external dependencies. No third-party SLAs.
What they didn’t anticipate was how much easier this made other conversations. When legal asked questions, when compliance needed explanations, when security wanted to verify controls, the answer was simple.
“It’s sitting here. This is what it does. This is how it communicates to the outside. Here’s the log. If you have any questions, let us know.”
No complicated data processing agreements. No debates about where information might end up. The infrastructure was on-site, observable, and fully under their control.
Growing without losing grip
Kenny didn’t stay on one server for long.
The single server was upgraded to a GPU-equipped server. That became three servers. Then four. Within ten months, they had nearly outgrown their initial setup and needed to rethink the architecture.
The infrastructure supported the growth. It was modular from the start. Trivium could add capacity without rebuilding from scratch. When they moved to Kubernetes for orchestration, the underlying hardware handled the transition.
The harder question was access control. Canny connected to sensitive systems. Not everyone in the organization should see everything. How do you prevent an AI assistant from exposing data that users wouldn’t normally access?
Their solution was a marketplace model. Users request access to integrations through an internal app store, similar to how you’d download an app on your phone. Each plugin requires approval from the application owner before activation. No blanket access, no shortcuts.
This meant the team never had to worry about Canny having more access than it should. The permissions followed the user, not the system.aving more access than it should. The permissions followed the user, not the system.
When the math starts working
Every AI infrastructure conversation comes down to the same question: what’s the return? GPUs are expensive. If they only run when someone asks a question during office hours, justifying the cost is difficult.
That equation changes when you introduce agents.
Trivium started automating tasks that run continuously. Ticket resolution. System monitoring. Processes that don’t wait for a human to type a prompt. When agents use your GPUs around the clock, you generate enough tokens that owning hardware starts costing less than renting cloud capacity.
Cloud pricing charges per token. The more you use, the more you pay. On-prem hardware is a fixed cost over five years. Once your usage crosses a certain threshold, ownership wins.
When agents run 24 hours a day, generating tokens continuously, cloud bills climb faster than hardware depreciation.
The math works once you know usage will keep growing. And at Trivium, it keeps growing.
What made adoption real
Building the platform was one thing. Getting people to use it was another.
One moment made the difference clear. When the procurement team first saw Kenny in action, they were stunned. One person asked: “Is it a he? Is it a she? How does this work?”
It was a reminder that most people have never interacted with AI this directly. What IT teams see as normal feels completely new to everyone else.
That same person became one of Kenny’s most active users. Daily use, all day. The shift happened because they saw it work with their own eyes.
PowerPoint presentations explain what AI could do. A working demo shows what it actually does. After seeing Kenny answer questions in real time, stakeholders who had been skeptical started asking when their team could get access.
What this means for your starting point
Trivium didn’t wait for flawless infrastructure. They started with a server that made users wait ten minutes for a response. That was enough to prove the concept and secure approval for the next project.
The team was motivated and flexible. When they needed to move from one server to a Kubernetes cluster, people were willing to learn, experiment, and rebuild when things didn’t work. Sometimes at 3 a.m. when a new idea couldn’t wait. That willingness to iterate matters more than having the newest hardware.
You don’t need the latest GPUs to begin. You don’t need a perfect architecture. Sixty or seventy percent ready is enough to launch. Ship something, learn from it, improve as you go.
And find a partner who can help you think through the first steps. Not just the infrastructure, but the use case itself. What should AI do for your organization? That question deserves as much attention as which server to buy.
Just getting started
Trivium began with one server, no GPU, and a ten-minute wait for every response. Today, they run a company-wide AI platform with a dedicated team and a list of departments awaiting onboarding.
They’re already planning a company-wide coding challenge. Inviting people from across the organization to build their own small applications using AI. Finance, procurement, operations. Anyone can participate.
Twenty years ago, nobody built their own office suite. You bought Microsoft licenses and moved on. Now, anyone can build tools that used to require a development team. And if you get stuck, the AI helps you through it.
This is still the beginning. For Trivium, and probably for you too.
Wondering what your first step could look like? Let’s talk.
