Artificial Intelligence / Architectures / On-premises
Artificial Intelligence on Your Premises: Data, Models, and Computations Remain on Your Premises
When the online service is unavailable, nothing is processed—neither documents, nor requests, nor the template.A contract that prohibits it, a trade secret, a regulated dataset, source code, or data volumes that would make the online service too expensive: certain use cases require that the model run on your datacenter. We then design the infrastructure to meet those needs—the server with graphics processing units, identity management, integration—and we maintain it. We’re not selling you a server—we’re operating a system for you.
Three situations where hosting on your own premises isn't an option
As a general rule, we do not recommend the venue. But in these three cases, the question no longer arises: the requirement is specified somewhere, and it takes precedence over everything else.
A clause prohibits the transfer of data
A contract with a client, a duty of confidentiality, an internal policy on research data, an industry regulation. No matter how good the online service is, the transmission itself is the problem. Regulatory constraints are the only ones that are non-negotiable.
You don't argue with a clause.
What the model analyzes is the company's assets
A publisher’s source code, a manufacturer’s processes, a firm’s files, a laboratory’s results. Even when processed by a third party that does not own it, such data should not be entrusted to a service that one does not control from start to finish.
The risk isn't in the use; it's in the exposure.
The volume makes the online service too expensive
Reviewing millions of pages every month, transcribing thousands of hours of calls, and handling hundreds of requests per minute: with pay-as-you-go billing, this ends up costing more than a server and its operating expenses. The break-even point is calculated over three years, and it comes sooner than you might think.
Once you exceed the threshold, renting a space is the most economical option.
Use cases that support it
Three detailed pages today, and three more that are taking shape as the projects unfold. Each one captures what remains human, and what the local context demands in addition.
Private RAG
Search a confidential corpus using everyday language, citing the source, using a model that runs on your premises because the online service is not an option.
PossiblePrivately Housed Business Assistant
The company's chatbot, which uses Teams as its interface and a model that never sees a document leave the system.
PossibleIntelligent Document Processing
When documents are regulated, Microsoft's reading services run in containers on your premises, on Azure Local.
Private Development Assistant
A code sample that runs on your graphics processors, for source code that must not be released.
Local Voiceovers and Transcriptions
Record, transcribe, and summarize calls without a single minute of voice traffic leaving your network.
Large-Scale Document Analysis
Millions of pages are processed each month: once that threshold is exceeded, local processing is the most cost-effective option.
What “everything on your premises” Really Means
A server with graphics processors—yes. But the server is the easy part. There are six other factors involved, and local projects stand or fall on them.
The calculation, based on the actual load
The number of concurrent users, the model size, the document length, and the acceptable response time determine the number and type of graphics processors. Not the other way around.
Power supply, cooling, chassis
An inference server doesn't fit into a chassis designed for virtualization. It's added alongside it, with its own requirements—that's the sizing process for our Dell servers: PowerEdge.
Storage of Corpora and Indexes
A document corpus and its vector index consist of millions of small files. The storage solution is chosen based on this profile and on the speed at which the model needs to read from it.
Identity and Rights
A local assistant should not see any more than a user: Entra ID remains the source of permissions, and each response adheres to what the user is authorized to view. Without this, the local assistant is nothing more than a well-organized internal leak.
The models and their updates
An open-weight model needs to be replaced every few months. Someone has to test the new one based on your questions, compare it, make a decision, and roll it out without interrupting service.
Supervision and Feedback
Who notices when response times slow down, latency increases, or the disk fills up? Who reverts to the previous version? These two issues must be resolved before the system goes live, not after.
What you measure determines what you buy
The sizing guide—cores, memory, disk access, accelerators, and growth—is described on the Dell servers page at PowerEdge. It applies word for word to an artificial intelligence platform.
What Powers a Private Artificial Intelligence System
Technology comes last, and we choose it on a project-by-project basis. We implement what’s appropriate for each situation, and that’s what each of us does.
Microsoft Foundry Local
Run templates from the Microsoft catalog locally, without a subscription Azure in the query path: on a workstation, in Windows Server 2025, and on Azure Local—this last option is currently in preview at Microsoft, with scheduling across multiple nodes and inference that continues if the connection is lost.
This approach preserves Microsoft's tools—the same interfaces, the same agents—while keeping computing on your premises.
Azure Local
Hardware that you own, located on your premises, managed and updated from Azure — including in offline operations, available as of February 2026. This is the foundation on which Foundry Local, Edge RAG, and document reading services run locally.
The hardware must be included on an approved list: Dell nodes for Azure Local are included in this list.
Open-weight models
Models published by their authors, which you can download and run yourself. They allow you to choose the model’s size based on your hardware, lock it into a specific version, and test it on your own questions before putting it into use.
Choosing the model is part of the project; it is redone every few months.
Dell PowerEdge with NVIDIA graphics processors
From general-purpose servers that support one or two accelerators for inference, to servers in the XE series designed for multiple graphics processors and specialized cooling. The right model is determined by the measured workload, not by the product catalog.
With Dell storage for the corpora, the network to connect the nodes, and data protection to ensure we don't lose three months' worth of indexes.
Microsoft Entra ID, and your systems
Your identity remains the same. The permissions that the assistant respects are the permissions for your shared content. And integration— Teams for the interface, SharePoint for documents, Dynamics 365 orERP for actions—takes place without copying your data elsewhere.
The capabilities of Foundry Local,Azure Local and the PowerEdge servers are documented by Microsoft and at Dell Technologies, and are re-verified before each proposal: they evolve from quarter to quarter, and this architecture changes faster than others.
The message isn't "buy a server"
Many providers sell a server with a pre-installed operating system, set up in two days. We offer something different: the choice of architecture, followed by the complete system, and then its operation.
What a hardware reseller does
- It provides a server with a preinstalled template and a chat interface.
- It connects your documents to it, just as they are.
- He leaves the operation, updates, and rights issues up to you.
- He'll never tell you that Copilot was enough, or that a hybrid deployment was cheaper: all he knows how to do is the server.
What We Do
- We prioritize architecture first, and we know how to do all four—that’s what makes consulting possible.
- We connect the assistant to your permissions and systems: Entra ID, Teams, SharePoint, Dynamics 365, andERP.
- We size and supply the equipment—which is prepared prior to delivery—both in Switzerland and at your overseas locations.
- We manage the platform —templates, oversight, costs, and feedback—or we train your team to do so.
What You Need to Know Before Making a Decision
Four points that aren't listed on a product sheet, and that we bring up at the first meeting.
The scale does not make the model more accurate
A model running on-premises makes just as many mistakes as an online model—and it does so with the same confidence. Privacy changes where the processing takes place, not the quality of the responses. Human validation remains a key component of every use case.
The model you'll be able to run is smaller
The largest models on the market require computational power that few companies have in their server rooms. A smaller model—carefully selected and trained on your data—often answers your specific questions better than a large, general-purpose model—but performs less well on everything else. We test it on your questions before reaching a conclusion.
Backup and rollback are included in the scope
It takes days to rebuild an index containing several million fragments. It is backed up as a database, and the process of reverting to a previous version of a model is prepared before the update, not during the incident.
The cost is compared over a three-year period, including operating expenses
Equipment, electricity, cooling, licenses, and the time spent by the people who maintain the platform, compared to the cost of using the online service for the same volume. Sometimes the comparison does not favor the local option in certain areas. We note this when it is the case.
How We Do It
Don't buy anything until the prototype has been tested with your actual data. This approach prevents the server from sitting idle, waiting to be used.
The Framing
The use case, the corpus and its rights, written requirements, expected volumes, and acceptable response time. And the calculation of the cost threshold for online and hybrid services.
The prototype, based on your data, using loaned or existing equipment
A model, a representative sample from the corpus, about twenty users, three weeks. We measure the accuracy of the responses, latency, and how users actually use the system.
This is where the project can end, and that's fine.
Architecture and Hardware
Servers, graphics processors, storage, networking, and data protection—all sized to match the prototype’s specifications. Pre-configured prior to delivery, installed, and connected to Entra ID and your systems.
Industrialization and Day-to-Day Operations
Gradual rollout, supervision, model updates, index backups, and review of responses. Performed by us or by your team, whom we will train.
The scope is defined before we begin. This includes what we’ll handle, what you’ll retain, the planned rollback, and the clearly defined limits. The same engineers will support you from the initial “ assessment ” phase through day-to-day operations.
Related pages
Other architectures, and the infrastructure that supports them.
When only part of the group needs to stay at your facility
The interface at Microsoft, sensitive data on your premises: often the most cost-effective solution.
View the page PlatformWhen multiple departments want their share of the local infrastructure
A shared, managed, and monitored infrastructure, rather than one server per team.
View the page EdgeWhen the decision is made by the machine
The facility extends all the way to the production line, the construction site, or a remote location.
View the page Dell Servers PowerEdgeA server scaled to meet the needs of your applications
Four product lines, the sizing chain, and the section on artificial intelligence.
View the page Azure On-Premises on DellThe Microsoft ecosystem in your datacenter
Dell infrastructure validated for Azure Local, and the three other ecosystems compared.
View the page Microsoft AzureAzure Office: Staying in the Office Without Feeling Isolated
The four situations in which we recommend it, and what we do not promise.
View the pageLet's talk about your use case
Tell us what shouldn't be released
The corpus, the people who need access to it, the text that prohibits the online service, and a rough estimate of the volume. We’ll come back to the architecture we’d be exploring, the cost threshold, and what a three-week prototype would allow us to measure.
What we offer is the opportunity to meet the engineers who will do the work. Assessment — prototyping and sizing are part of the project. The initial consultation, however, is free of charge.
Renens, Sion, Châtel-Saint-Denis
Microsoft Solutions Partner and Dell Technologies Gold Partner. The template, the identity, the integration, and the server that hosts them—all handled by the same point of contact.
Renens VD +41 21 806 37 15
Sion VS +41 27 552 00 22
Châtel-Saint-Denis FR +41 26 322 59 05