Brussels didn’t hold back. The target is the privileged entrance Google built around Gemini on Android: now it must open it to rival AI assistants, with the same system-level hooks its own model enjoys. And that’s not all—a slice of the search data amassed by the world’s most used engine must be shared with competitors. These twin decisions, formalized by the European Commission on Thursday, are both anchored in the Digital Markets Act (DMA), the law that reins in so-called ‘gatekeepers’.

The technical impact is disruptive even though the official text remains sparse. To grasp its reach, you have to look beyond the legal perimeter and observe what happens when the incumbent mobile OS is forced to level the playing field for artificial intelligence. Until yesterday, Gemini could rely on direct access to sensors, notifications, call management, and deep integration with Google apps—advantages any competitor (from OpenAI’s ChatGPT to Anthropic’s Claude) could only dream of. From today, those same privileges must be granted to anyone meeting the DMA’s technical requirements.

A quake for mobile AI balances – and far more

The obligation goes far beyond app stores or choice screens. The real shift lies in the ability to run language models with the same latency and the same capacity to act on the device that Google had reserved for its own assistant. That means third-party providers can exploit on-chip inference pipelines—NPU, GPU, CPU—without passing through Mountain View’s cloud services. For those building large language models optimized for mobile, the DMA becomes an on-device deployment accelerator no commercial deal could have secured before.

The search data sharing cuts a different front. Competitors will be able to train their LLMs on query patterns and user behavior that were once Google’s exclusive preserve, reducing the information asymmetry that fuels the company’s advertising flywheel. It’s a game about data sovereignty, but also about the capacity to build truly competitive language models without leaning on proprietary, inaccessible datasets. In an ecosystem where on-premise fine-tuning becomes strategic, having access to a knowledge base that reflects European search habits can make the difference between a generic assistant and one that actually understands the user.

Who wins, who loses: the day-after map

The winners are companies that already invested in compact models and local execution frameworks—think firms working on Llama 3.2 or Mistral optimized for mobile, but also businesses crafting voice interfaces tailored for privacy. For them, the DMA provides a fast lane without having to negotiate with Google. The immediate losers are services that based their competitive edge on ecosystem lock-in: Gemini itself, now forced to face rivals offering the same fluid user experience. But device makers that struck exclusive deals with Google might also see that tie-up devalued.

Structurally, the decision signals that AI control is no longer fought only in data centers. When a regulator mandates equal access to system features, the real battleground shifts to silicon. Chips like Google’s Tensor or Snapdragon processors with dedicated NPUs become the neutral field where the quality of inference is decided in milliseconds. Hardware makers for training and local execution—from NVIDIA with its Jetson line to smartphone system-on-chips—see additional demand open up that wasn’t a given: a mass market for on-device AI, no longer tied to a single gatekeeper’s whims.

The open question concerns timelines and implementation details: the Commission has not yet outlined technical standards for search data sharing or system API access. But the direction is set. For those evaluating on-premise or edge deployment of language models, this news isn’t just another headline: it’s confirmation that data sovereignty isn’t protected only by pulling the cloud plug, but also by rules that make the device contestable territory. And it forces a rethink of the Total Cost of Ownership for AI solutions: if on-premise becomes a market mandate as well as a technical choice, the numbers have to be redone.