Beyond Policy-to-Code: Asenion Awarded 750,000 GPU-Hours to Advance AI Compliance Research

Following successful technical validation under two earlier European supercomputing allocations, Asenion secures new EuroHPC AI Factory resources to advance research into context-aware AI compliance.

By David Van Bruwaene, Co-Founder and CEO, Asenion

What does it take to build AI systems that can determine not just what is technically possible, but what is permitted? This is one of the most important questions facing enterprises deploying AI agents.

A general-purpose large language model can interpret a policy. But interpreting a policy is not the same as reliably determining whether a specific AI action complies with an organization’s requirements, context and intended use.

At Asenion, we believe solving this problem requires more than applying off-the-shelf LLMs to Policy-to-Code. It requires fundamental research into how machines understand and operationalize compliance.

Today, I’m pleased to share an important milestone in that journey.

Asenion has been awarded 750,000 GPU-hours through the EuroHPC AI Factory Large Scale call, following technical peer review and evaluation by the Industrial Innovation Group. The allocation provides access to Leonardo BOOSTER, a European supercomputer hosted by CINECA in Italy.

This builds on two earlier 50,000 GPU-hour allocations, which supported successful technical validation on LUMI-G, the GPU-accelerated partition of Finland’s LUMI supercomputer. Together, these milestones reflect our continued investment in advancing the technical foundations of AI compliance.

Why AI compliance needs more than bigger language models

As AI evolves from generating content to autonomously taking actions, compliance becomes a fundamentally more complex problem.

Consider an AI agent handling personal information. For one organization, collecting certain information may be explicitly authorized for a mortgage application. For another, collecting that same information may violate internal policy. Even within the same organization, the correct decision may depend on the user’s role, the jurisdiction, the purpose of the interaction and the action being performed.

The challenge is not simply identifying sensitive information or detecting a potential security threat. It is determining whether an AI action is appropriate within its specific operational context and organizational intent.

Traditional classifiers, general-purpose LLMs and rule-based systems each have strengths. But enterprise compliance introduces challenges that cannot always be addressed by a single technique. We believe there is an opportunity to advance more specialized AI capabilities designed for this problem.

And importantly, bigger does not automatically mean better. For compliance, accuracy, reliability, adaptability and the ability to verify decisions matter far more than model size alone.

Why this EuroHPC award matters

The EuroHPC AI Factory Large Scale programme provides advanced computing resources to support ambitious AI research and innovation. Asenion’s proposal was evaluated through a competitive process that included technical assessment and industrial innovation review.

The evaluation recognized the project’s strong innovation potential and credible pathway toward industrial relevance in regulated domains, while identifying important research questions that still require experimental validation.

That distinction matters. The purpose of research is not to assume that a new approach works. It is to establish what works, under which conditions, and how reliably. The new allocation gives our team access to substantial computing infrastructure to pursue these questions through rigorous experimentation.

For Asenion, this is more than a computing milestone. It represents an opportunity to advance specialized AI research informed by years of practical experience implementing AI compliance and assurance across industries and jurisdictions.

From the language of compliance to the language of machines

At Asenion, we have long believed that a requirement is not a control. Regulations, standards and organizational policies establish what must be achieved. But AI systems need operational controls that reflect the circumstances in which they are deployed.

This is the foundation of the Asenion AI Control System:

Requirement → Context → Intent → Control Objective → Control → Verification → Evidence

Our approach connects compliance requirements to operational controls across the AI lifecycle:

ASSESS → VERIFY → WITNESS

The challenge becomes even more important with agentic AI. Security frameworks such as the OWASP Top 10 for LLM Applications provide valuable guidance on common vulnerabilities. But understanding whether an AI agent is authorized to take a particular action requires additional consideration of organizational policies, business context and intended use.

This is where we see an important frontier for AI compliance research.

Building the next generation of compliance intelligence

We are grateful to EuroHPC JU and the AI Factory evaluation teams for recognizing the potential of our research and providing access to this important European computing infrastructure.

As enterprises move toward increasingly autonomous AI systems, compliance must evolve alongside them. The question is no longer simply whether an AI model can produce an acceptable answer. It is whether an AI system can operate within the requirements, permissions and intentions of the organization deploying it - and whether that can be demonstrated with evidence.

Want to understand the bigger picture?

This research is part of our broader mission to bridge the gap between human-written requirements and machine-executable AI controls.

Read our earlier article, From the Language of Compliance to the Language of Machines, to explore how Asenion connects requirements, context and organizational intent to operational AI compliance.