Sherpa.ai raises $18M to scale privacy‑preserving, data‑sovereign AI

Sherpa.ai’s raise underscores a critical counterweight to the centralisation of AI: architectures where data stays put and models move. As regulators in Europe and elsewhere push on data sovereignty and cross‑border flows, privacy‑preserving techniques like federated learning and secure aggregation become not just nice‑to‑have, but often the only viable way to train models on sensitive datasets. Sherpa is betting that these constraints will harden over time and that enterprises will prefer platforms designed from the ground up for joint training without data pooling.

In the race to AGI, this matters in two ways. First, it offers a path for smaller countries and institutions with strict privacy regimes to still participate in building powerful models, rather than being locked out by data export rules. Second, it challenges the notion that only massive centralised datasets at a few hyperscalers can power truly capable systems. If data‑sovereign collaborations become routine, we could see a more federated ecosystem of strong but specialised models that collectively rival or complement frontier systems.

It also increases the technical and policy complexity of AI governance: evaluating and auditing models trained across dozens of silos is harder than inspecting a single central pipeline. But that complexity may be the price of aligning AI progress with democratic expectations about privacy and control.