NVIDIA and Adobe Advance Sovereign AI and Video Search for Media
The EBU is expanding its framework for broadcaster controlled AI, while Adobe reports faster video search for production teams using NVIDIA GPUs.
Media organizations are tackling two practical challenges as AI enters production: keeping control over the infrastructure and models behind services, and finding useful moments in growing video archives. At a creative industry event in Berlin, the European Broadcasting Union (EBU) outlined an expanded framework with NVIDIA for broadcasters, while Adobe described GPU accelerated search work for Frame.io, including results relevant to German producer UFA.
A deployment framework for broadcasters
The EBU and NVIDIA are extending standards they introduced last year for European public service media. The expanded initiative brings sovereign cloud providers, open model developers and systems integrators into the effort. Its aim is to turn principles around data control and security into requirements and reference architectures that broadcasters can use when deploying AI agents.
The framework is intended to cover several deployment decisions: where workloads run, how organizations govern data, which models support services, and how systems remain secure throughout their lifecycle. That scope matters because an AI service can depend on infrastructure, data handling and model choices spread across multiple suppliers. Shared requirements could give broadcasters a common basis for evaluating those dependencies.
The EBU describes itself as an alliance of public service media with more than 115 member organizations across over 55 countries, reaching more than 1 billion people. As broadcasters consider agents for newsrooms, production and audience services, the initiative positions control over data and infrastructure as part of the operating model. The stated goals include protecting editorial independence and supporting resilience.
“These standards give EBU members a shared foundation to strengthen resilience, retain control and deliver accountable advanced services that sustain audience trust,” said Annsofi Eriksson, the EBU’s chief technology officer.
Two cloud providers are mapping the framework to deployment options. Microsoft is contributing expertise from Microsoft Sovereign Cloud, including Azure Local and Foundry Local. The company says these offerings can let broadcasters process sensitive editorial material locally and keep critical services available when external connectivity is down. Nebius offers a platform for broadcasters to train, fine-tune and run their own models on European infrastructure.
Open models are another part of the approach. They can give broadcasters more ability to inspect and adapt systems for local languages, archive material and editorial policies. NVIDIA’s Nemotron models are presented as a starting point for domain specific applications and public service requirements. The source does not specify which models or services members will deploy, or provide performance results for broadcaster use.
France Télévisions CEO and EBU president Delphine Ernotte said sovereign cloud infrastructure and open models can support innovation while protecting data, editorial independence and cultural diversity. Qvest and other systems integrators are expected to help make the reference architectures workable in broadcast environments and interoperable across systems. That implementation work will shape how much of the framework can move from shared standards into services broadcasters can operate.
GPU accelerated search for video archives
Adobe is also using NVIDIA accelerated computing to speed up video search in Frame.io. The pipeline combines NVIDIA NVDEC for decoding, TensorRT for optimized AI inference, and CUDA for preprocessing and generating embeddings. Adobe is also applying NVIDIA NVENC and NVDEC to media ingest, using GPU acceleration for encoding and decoding as production teams process larger libraries.
Adobe reports about five times faster embedding generation, up to four times higher throughput, and up to five times lower latency for search processing. It has also introduced a proxy running at two frames per second. The proxy is about 80% smaller than the 360p version it replaces, reducing storage and transfer needs while retaining visual detail used for search. The source does not provide the test setup or a baseline for these figures, so they describe Adobe’s reported results rather than a general performance guarantee.
The use case is clear at UFA, which produces more than 3,500 hours of programming each year, including the daily drama “Gute Zeiten, schlechte Zeiten.” Editors assembling a recap or flashback may need a particular reaction or an unresolved exchange from as far back as 15 years. A folder name or file title may not describe that moment, leaving editors to search through footage manually.
Frame.io’s semantic search is designed to find moments based on concepts, actions or settings within video. Hybrid search can also use notes, transcripts and approval status. An editor could search for “a character looking concerned,” narrow results by character name, and filter for approved footage before moving into review. The workflow connects natural language search with production metadata, which can make an archive easier to query without relying on filenames alone.
These gains depend on the quality of the indexed material and the usefulness of its embeddings, transcripts and metadata. Search can narrow the candidate footage, but editors still need to judge whether a result fits the scene and production context. The reported speed figures address processing performance; they do not establish search accuracy or how consistently the system finds a specific moment across different archives.
What the announcements leave open
The two efforts address different parts of media AI deployment. The EBU initiative focuses on how broadcasters select infrastructure, govern data and assess models. Adobe’s Frame.io work targets the time and compute involved in processing and searching video. Together, they point to practical concerns teams will need to weigh: where workloads run, what material enters a service, how models can be adapted, and whether faster indexing translates into quicker editorial work.
The next measure of progress will be operational: whether broadcasters can deploy the proposed architectures across their existing environments, and whether production teams find the search results accurate enough to change archive workflows. NVIDIA and Adobe also describe a wider collaboration spanning Adobe Substance 3D, Firefly Foundry and enterprise 3D digital twin workflows, though the event announcement provides no additional technical details on those areas.
