NVIDIA pushes “agentic” security and faster video search as broadcasters and studios pursue sovereign AI

NVIDIA used an industry event in Berlin to highlight a tightening focus on “agentic security” for media workflows, pairing it with demonstrations of faster semantic search across large video libraries. The company’s message was aimed at public-service broadcasters and commercial producers alike, where AI is increasingly being used to help find moments in long archives, speed editing, and support newsroom operations.

At the center of the announcement is a collaboration between NVIDIA and the European Broadcasting Union (EBU), the alliance representing public service media organizations across Europe and beyond. NVIDIA said the EBU is expanding a set of “sovereign” standards it introduced last year for European public service media, now adding sovereign cloud providers, open model developers, and systems integrators to help define how “agents” can be deployed while maintaining control over data, infrastructure, models, and security.

NVIDIA framed the standards as a response to the growing role of AI agents in newsrooms, production, and audience services. As agents move from prototypes to daily operations, the company said broadcasters need a foundation that protects editorial independence, data sovereignty, and national resilience. EBU membership, NVIDIA noted, reaches more than 1 billion people across 55-plus countries via more than 115 organizations.

The expanded initiative is described as turning principles into deployable services, with the effort intended to define requirements and “reference architectures.” Those architectures are meant to spell out where workloads should run, how data should be governed, which models should power services, and how systems can remain secure across their lifecycle. Antonio Arcidiacono, EBU chief technology and innovation officer, said in the announcement that shared standards give EBU members a foundation to strengthen resilience, retain control, and deliver accountable advanced services that sustain audience trust.

NVIDIA also detailed how different parts of the ecosystem would plug into the framework. Cloud providers, the company said, will map the standards to sovereign deployment options. Open model developers, meanwhile, would bring visibility so that broadcasters can inspect, evaluate, and adapt AI for local languages, archives, and editorial policies. NVIDIA referenced its Nemotron open models as a starting point that broadcasters can customize for domain-specific uses and public-service requirements.

Security and practicality were further addressed through consulting and systems integration, with NVIDIA naming Qvest and “other system integrators” as contributors. The goal, according to the company, is to ensure that reference architectures can be implemented in real broadcast environments, not just defined on paper.

Alongside the sovereign-agent push, Adobe also showcased video search improvements that it says are accelerated by NVIDIA GPUs. NVIDIA described an AI video search pipeline in Adobe that uses NVDEC for video decoding, TensorRT for optimized AI inference, and CUDA for preprocessing and embedding generation. NVIDIA said Adobe is also expanding GPU-accelerated media ingest using NVENC and NVDEC to speed up video encoding and decoding as media libraries grow.

NVIDIA quantified the performance improvements it attributed to the approach, including about 5x faster embedding generation, up to 4x higher throughput, and up to 5x lower latency for search processing. To reduce storage and transfer while keeping enough visual detail for search, NVIDIA said Adobe introduced a new two-frame-per-second proxy, described as roughly 80% smaller than the 360p version it replaces.

The practical value of faster search was illustrated through UFA, a German film and television producer cited by NVIDIA as delivering more than 3,500 hours of programming per year, including the daily drama “Gute Zeiten, schlechte Zeiten.” NVIDIA said semantic search can help editors find specific moments based on concept, action, or setting, addressing the limits of file names and folders. In a scenario described by the company, an editor could search for a “character looking concerned,” narrow results by character name, filter to approved footage, and then move quickly into review.