The order comes straight from Brussels, and it’s not a recommendation: Google must open up its Android-integrated AI features and make Search data available to rivals. The Digital Markets Act (DMA) hits one of the most sensitive nodes of the digital ecosystem, demanding a mix of transparency, interoperability, and forced sharing of strategic assets.

This is about more than market dynamics. The measure strikes a nerve for anyone handling sensitive data or developing proprietary models. The obligation to share Search data — and to expose AI functionalities previously locked inside Google’s walled garden — signals a paradigm shift: regulation can decide what becomes a common good, even when it’s built on private infrastructure.

The sovereignty knot: what changes for AI practitioners

For organizations evaluating where to run their Large Language Models (LLMs), this move has disruptive effects. If a company entrusts data and models to a dominant cloud provider that is later forced by law to open access to third parties, the boundary between protected and exposed data blurs. You don’t need to invoke GDPR: DMA-imposed sharing introduces an indirect exposure risk that no contractual clause can fully neutralize.

In practice, those who currently use cloud AI services for inference on proprietary data may need to rethink their architecture. On-premise stacks — with self-hosted models, dedicated servers, and physical memory control — return to the center of discussion precisely because they ensure data never leaves the company perimeter. This is not a technical detail but a governance choice: when regulation mandates openness, owning the hardware becomes the only way to preserve confidentiality.

Hardware and incentives: the silent push toward local deployments

The DMA order doesn’t mention GPUs, VRAM, or quantization. But it redefines the incentives. If compliance becomes more complex and expensive in cloud environments subject to data-sharing obligations, the value of local, air-gapped architectures rises. Those already investing in on-premise clusters for LLM training or inference suddenly gain a competitive edge: the infrastructure is no longer just a compute tool but a legal asset that shields against regulatory intrusion.

This explains why, in recent months, the market for multi-GPU workstations and bare-metal AI servers has drawn interest not only from research labs but also from legal departments. Data sovereignty isn’t negotiated via contract — it’s implemented at the silicon level. And the DMA, unintentionally, is accelerating this awakening.