AI-stack
Our soeverein AI stack
We do not believe in the "one size fits all" of Big Tech
Open-weights versus closed-source
We resolutely choose leading European open-weights models. Why? Because absolute data sovereignty requires that we maintain full control over the model weights and the infrastructure within which the model operates. By hosting these models entirely within our own secure stack, no data leaks to external servers or central parent organizations in the US.
Custom fine-tuning
We train and fine-tune these models specifically for the complex legal, administrative, and operational context of the European Union.
When you use closed-source AI models via public APIs, you pay per token (the number of processed words). In large-scale data integrations and continuous use, this leads to unpredictable, exponentially rising costs.
With Thor, we completely move away from this token payment model:
Unlimited processing: Because the open-weights models run dedicated on private hardware (within your perimeter or at LCL), you do not pay for consumption. Whether Thor analyzes 10 or 10,000,000 documents per day: the costs remain the same.
Fixed, predictable ROI: Your investment shifts from variable operational costs (OpEx) to a fixed, predictable infrastructure. You can scale up without fear of an unexpectedly high bill at the end of the month.
No 'data tax' on integrations: Because the API connections remain internal, you can feed Thor without financial barriers with the rich, extensive context needed for enterprise tasks.
We ensure compliance and ethical standards not retroactively, but through a proactive, layered architecture (compliance-by-design).
Real-time Guardrails: Every interaction and action is assessed in real-time against predefined risk assessments. Our systems utilize advanced guardrail layers that filter outputs based on the AI Act and specific company policies.
Automated Intervention: As soon as a response or action threatens to enter a legal gray area, the system proactively intervenes by blocking, rephrasing, or providing a disclaimer for the output.
Human-in-the-Loop & Escalation: In complex or boundary-pushing cases, our escalation protocol is activated. The interaction is paused and handed over to our engineers and compliance experts, who take control to ensure quality and legal robustness.
By hosting our stack with LCL, we firmly commit to Belgium's highest-certified data center infrastructure (Tier III).
Immunity from the US CLOUD Act
Because LCL is a 100% Belgian entity and we only run European software, US authorities have no legal grip on your environment. Your data remains under strict Belgian and European jurisdiction at all times.
Built-in certification
You directly benefit from the strict ISO 27001 and ISAE 3402 standards that are already fully embedded in the foundations of our hosting.
The arguments in a row
Open weights-models (Thor) vs closed source models (zoals ChatGPT & Claude)
- Infrastructure & Hardware
- With a closed-source model, everything runs entirely on the server farms of OpenAI or Microsoft. The underlying technology remains completely invisible to the user.
- With an open-weights model, you have to host the 'engine' yourself. Because AI models require powerful graphics cards (GPUs), this demands specialized hardware.
- With our Hardware-as-a-Service model, companies can flexibly rent the necessary computing power and hardware from DataNerds.
- Implementation & Interface
- Open-weights models are often delivered 'bare-bones.' They lack the commercial, user-friendly wrapper that Big Tech companies typically build around them.
- This is precisely why DataNerds exists. We don't just deliver a standalone model; we deliver Thor. Thor is a pre-trained, user-friendly digital assistant that comes fully connected to data.
- Knowledge & Management
- A closed-source model updates itself silently in the background. An open-weights model works differently. This requires in-depth technical knowledge. An average SME IT professional cannot just take this on.
- DataNerds acts as both the architect and the operator. Our engineers maintain the integration engine and ensure that Thor remains up to date.
- Speed & Relevance
- Gigantic closed-source models (such as GPT-4o or Claude 3.5 Sonnet) feature hundreds of billions of parameters. They excel at doing everything simultaneously. For broad, general world knowledge, such a commercial giant is often superior out-of-the-box.
- An employee in an SME or industrial environment doesn't need an AI that writes poetry; they need an AI that understands stock levels, CRM data, and accounting. By explicitly coupling a more compact, open-weights model (such as Mistral) to the ERP system, it performs faster, more purposefully, and far more securely on the shop floor than a generic American cloud AI.