Sherpa.ai, a startup based in the Basque Country, has closed an $18 million funding round to advance an ambitious and necessary promise: building AI systems that never see users’ raw data. The news, reported by tech.eu, signals growing interest in solutions capable of delivering the power of LLMs without forcing banks, hospitals, and governments to hand their most sensitive information to third parties.

The Sherpa.ai approach is part of a wave that is redefining the boundaries of machine learning: training models without moving data. Techniques like federated learning and differential privacy allow information to remain on the customer’s systems, sending only encrypted gradients or parameters to servers. In this way, the company never accesses account numbers, medical records, or confidential documents, yet the model continues to improve.

The cited customer sectors – finance, healthcare, and public administration – are precisely those where data sovereignty is non-negotiable. European regulations like GDPR, along with national laws, require that certain data must not leave jurisdictional borders. Putting everything in a public cloud, even if encrypted, does not satisfy the requirements of some entities. That’s why Sherpa.ai’s proposition appeals to those seeking to leverage AI without surrendering direct control.

From a deployment perspective, this kind of architecture fits well with on-premise or hybrid infrastructures. Distributed training does not require the entire dataset to be centralized on a single node: each organization can process its data locally, contributing to a shared model. For IT leaders, this means evaluating on-site compute capacity, latency, and TCO, while also reducing exposure risk.

Of course, the privacy path has its trade-offs. Federated training can be slower than a centralized approach, and model convergence is not always guaranteed with the same accuracy. Moreover, differential privacy introduces noise that can degrade performance on specific tasks. Those adopting these technologies must weigh the potential loss of accuracy against the gains in compliance and security.

In a landscape where cloud providers promise “sovereign” environments but often fall short on transparency, the idea of AI that never touches data represents a paradigm shift. For organizations evaluating the shift to large language models, the choice is not only about GPUs and software licenses: it also hinges on defining where and how data is processed. AI-RADAR tracks the evolution of these stacks, offering independent analysis on the trade-offs between hardware, privacy, and operational costs.

The $18 million round, led by European investors, gives Sherpa.ai the opportunity to scale and refine its technology. It remains to be seen whether the promise of data-“blind” AI will convince companies increasingly wary of centralized solutions. The direction is certainly toward machine learning that respects organizational boundaries, and the coming months will tell if this Basque startup can turn the principle into a scalable product.