ONPREMBENCH / THE STORY BEHIND THE LAB
We put a machine on the office floor.
Then we started writing down what happened.
OnPremBench grew out of Understand Tech’s own work bringing AI onto local hardware for businesses. This is why we built it, what we learned and what we hope it becomes.
- 01
IT STARTED WITH A BOX
We put our software on a real machine.
Understand Tech develops AI in a Box: a software platform that brings AI models, applications and operational support together on local hardware.
When we began deploying that stack on NVIDIA GB10 and GB300 workstations, we discovered two things. First, these machines can support useful business deployments, well beyond the individual experimentation people usually associate with desktop AI hardware. Second, making that happen reliably takes substantial work.
- 02
THE PART NOBODY SHOWS
Loading a model was only the beginning.
We spent months evaluating serving stacks, dealing with compatibility issues, investigating startup behaviour, managing resources and learning from customers.
A model fitting in memory is not a machine supporting a workload. “How many users?” needs assumptions about activity, request rate, simultaneous requests and acceptable response time. Operating an AI system for a business is a different job from running one on a desk.
Read the operator’s notebook - 03
THE MOMENT THAT CHANGED THE CONVERSATION
People needed to see the box.
In customer meetings and on LinkedIn we noticed something powerful: showing the actual machine changed the conversation. People could see a physical box beside a team and understand that useful AI could run there. The opportunity became tangible.
That observation is where the Lab started. Not with a benchmark or a product page, but with a machine on the office floor and the questions it raised.
“Two months ago it was a slide. Today it’s on our floor.”Naama Bak · Co-founder, Understand Tech
- 04
A PRACTICAL MIDDLE GROUND
Not cloud or nothing.
Many organizations use cloud AI successfully. Others have workloads where confidentiality, data restrictions, operating costs or disconnected operation make local execution attractive. They may want their own AI capability without becoming datacenter operators.
A complete workstation is another option for suitable workloads: hardware, models, serving software and business applications operating together inside the organization. We are not claiming that local machines replace frontier cloud models. We are helping people understand when a local system is a useful alternative and what it takes to operate it.
Privacy and cost are central, and both need evidence. Owning a box does not automatically secure the whole application. Local tokens have a real cost: hardware, power, software, administration, maintenance and unused capacity.
Privacy, security and the cost of tokens - 05
WHY THE MACHINE COMES FIRST
Start from something you can point at.
Visitors usually already know what they want AI to do. Their question is whether a real system can support it. So the Lab starts with an identifiable machine and its photograph, then opens into its specifications, memory architecture and installation needs; the models, quantizations and serving stacks with documented support; what is known about speed, latency and simultaneous requests; the applications and deployments that make it concrete; what owners learned, including failures; and its pricing, availability and support options.
Every record distinguishes evidence from the exact machine, evidence from a related platform, supplier claims and planning estimates.
Explore the machines - 06
WHY A COMMUNITY
The catalogue brings people in. Operating knowledge brings them back.
A useful record connects hardware, model, runtime, application, workload and outcome. It might document a successful deployment, a configuration improvement, a slow startup, a memory limit or an experiment that failed.
Understand Tech contributes its own experience. Customers can contribute approved deployment stories. Independent owners and supplier engineers add what they learn. Existing tutorials and recipes are indexed with attribution and context. We take inspiration from how Hugging Face made models discoverable; our organizing object is the complete AI machine and its deployment record.
Join the conversation - 07
THE LAB AND UNDERSTAND TECH
Useful without a purchase. Honest about who we are.
The Lab should remain useful to someone who never buys anything from us. People can explore, learn, compare and contribute. Organizations that want help can engage Understand Tech for the software platform, applications, deployment, maintenance and support.
Sharing operating knowledge does not require publishing our proprietary software. We explain findings, diagnostic methods and configuration evidence while protecting customer assets. The relationship with Understand Tech stays visible, and commercial relationships never determine technical rankings.
How Understand Tech can help - 08
WHERE THIS GOES
The platform still has to earn its community.
Today the Lab holds sourced machine records, model configurations, a technical library, member workspaces, moderated contributions. Each contribution can make the next deployment easier to understand and reproduce.
The ambition is simple to state: OnPremBench should become the place a professional visits when considering an AI machine for a real workplace. To see what exists, understand what is possible, learn from others and make a better deployment decision.
Help build the shared record
YOUR TURN
Someone else is figuring this out, too.
The records on this site come from a real office, real customers and named sources. Add yours: a question, a working setup, a failure you solved or a deployment you can talk about.