Antonio Arcidiacono, EBU Director of Technology & Innovation
AI brings real benefits to media users: personalization, accessibility, new viewing experiences. These benefits, however, tend to imply a huge increase in energy consumption at data centres and across the network, and raise legitimate concerns about data privacy.
We need to think differently – evolving media services towards a distributed AI architecture that is more personalized, efficient and democratic. The fundamental idea is to push key AI processes out of the data centres, to the edge of the network and onto the devices in the users’ homes, referred to in the industry as Customer Premises Equipment (CPE).
Sustainability and privacy
The sustainability case is clear: processing AI tasks locally eliminates the energy needed to transmit large volumes of data to centralized cloud facilities and back, also reducing backbone network traffic and associated energy consumption. CPE devices can run AI tasks during hours of solar generation, using renewable energy directly and avoiding the intensive cooling demands of centralized data centres.
And what about privacy? Running intelligence in CPE devices keeps sensitive personal data on the user’s premises, achieving true data sovereignty. In this way, individuals maintain direct control over what data is processed and shared, which makes it easier for service providers to comply with data protection regulations like GDPR.
In a further advantage, decentralization would eliminate single points of failure and the massive data concentration that is prone to centralized attacks. Models can still be improved but through distributed learning, without centralizing raw data.
From a technical perspective, such a multilayer system would function effectively with edge devices handling routine tasks while cloud resources manage complex training and coordination. This requires the development of efficient, lightweight AI models optimized for edge deployment – like those Apple and Qualcomm are developing for smartphones – and demands coordination between the edge and cloud layers. The vision is made possible by the continued improvements in edge AI chips and CPE capabilities.
Better personalization
The proposed model makes it possible to adjust content recommendations and formatting based on the local context – time of day, device capabilities, user mood or activity – with viewing habits and preferences analysed entirely on-device. Local AI enables real-time interaction with educational content, news shows, or entertainment. Live sports can be enhanced with locally generated overlays, statistics, and multiple viewing angles rendered on-device, while fiction could be adapted based on aggregated local audience responses.
Accessibility improves through locally performed translation, lip-sync, subtitling, audio description, with sign-language interpretation possible on more capable devices. Content authenticity is assured through C2PA-based tracking of origin and modifications.
At the production end, too, new possibilities are enabled. Multiple creators could be working on shared projects with only metadata and deltas synchronized, not raw footage. Broadcast and peer-to-peer content distribution could reduce central infrastructure load and CDN costs.
None of this will happen unless it makes sense for the bottom line. New business models could see users contributing edge computing power in exchange for premium content access. Privacy concerns around lucrative addressable advertising can again be addressed via on-device rendering. Original content can be monetized with AI-managed rights and royalties.
In uncertain times, a failure of infrastructure could see hybrid networks deliver critical information, with disaster warnings and public safety messages tailored to local conditions. Local AI could help organize community responses during emergencies.
This shifts media from a one-to-many broadcast model to a many-to-one, adaptive ecosystem, where the distinction between creators and consumers blurs, all while respecting privacy and optimizing energy use. And it’s a model where public service media can find a natural home.
This article first appeared in the June 2026 issue of tech-i magazine.