Build a Media Agent: CPU Agent, GPU Models as Tools

​Run expensive GPU models only when your agent needs them - Alexey Grigorev

​Generative media apps don't need a GPU running all the time. The agent itself is cheap: it can run on a CPU, read a brief, and decide what to make. The expensive part - image and speech models - can live on GPUs and be called as tools only when needed.

​Join this hands-on coding workshop where we'll build a small media agent on Nebius: a CPU agent as the control plane, with GPU-hosted models exposed as tools. We'll also compare the two ways to run GPU work - a Serverless Endpoint and a batch Job - by latency and cost per asset.

​We'll cover the following steps:

• ​Setup and why: CPU for the agent, GPUs for the models (10 min)

• ​First tool: wrap a text-to-image model on a Nebius Serverless Endpoint as generate_image(spec), which returns an Object Storage link (20 min)

• ​CPU agent as the control plane: take a brief and decide which tools to call, then add text-to-speech as a second tool via MCP (25 min)

• ​Endpoint vs Job: run the same generation as a batch Job for 50 records and compare latency and cost per asset (20 min)

• ​What's next: the capstone project - a Personalized Media Factory, Q&A (15 min)

​To follow along, you'll need:

• ​A Nebius account with credits

• ​uv installed

• ​The workshop repository: /DataTalksClub/build-media-agent-workshop

​By the end of this workshop, you'll have a working media agent that calls GPU models as tools, and you'll know when to use an Endpoint and when to use a Job.

​This workshop is a preview of our new course "Production AI Agents for Generative Media", built with Nebius. The course starts on Monday, November 16, 2026. Apply by November 30, 2026: https://courses.datatalks.club/register/production-ai-agents-generative-media/

​About the Speaker:

​Alexey Grigorev is the founder of DataTalks.Club and the creator of its free courses, including Machine Learning Zoomcamp, Data Engineering Zoomcamp, and LLM Zoomcamp. He has over 15 years of experience as a software engineer, data scientist, and ML engineer, and is the author of "Machine Learning Bookcamp". These days he spends most of his time building AI agents and teaching others how to put them into production.

​DataTalks.Club is the place to talk about data. Join our Slack community!