IBM and Confluent bring time-series AI to live data streams on standard CPUs
IBM's time-series foundation models now run natively in Confluent via Apache Flink, allowing real-time forecasting and anomaly detection on CPUs without costly GPUs.
IBM and Confluent have integrated time-series foundation models directly into Confluent’s streaming data platform. The setup allows companies to run forecasting, anomaly detection, and optimization on live data feeds using Apache Flink and standard SQL queries. The models, which have logged over 50 million downloads, run natively on general-purpose CPUs without requiring expensive graphics processing units.
Real-time inference on streaming data
Businesses generate massive volumes of continuous time-series data, including machine sensor logs, payment authorizations, sales records, and inventory shifts. Traditional batch processing often delivers insights too late to stop production errors, stockouts, or failed payments. By embeddingEmbeddingA list of numbers that stands for the meaning of a piece of text. Open in glossary IBM’s models into Confluent Cloud and on-premises systems running Confluent Platform for Apache Flink, teams can analyze signals as data moves through the pipeline.
Time-series foundation models learn general mathematical patterns from vast historical datasets. Instead of forcing data science teams to build and maintain separate models for every product, machine, or store, a single pretrained system serves as a baseline. Teams can still fine-tune these models or join their outputs with business rules and other machine learning systems.
Industrial and retail use cases
The system targets operational use cases across banking, retail, airlines, and manufacturing. In a food manufacturing plant, for example, a pretrained model can ingest factory telemetry every 10 minutes to forecast product quality four hours in advance. Operators can then test and adjust equipment settings before entire production runs fail quality audits.
In retail, store operators can use built-in functions like anomaly detection to track thousands of items across hundreds of locations. Rather than relying on rigid thresholds that misinterpret normal shopping surges, the system builds expected sales ranges based on recent activity, alerting managers when a product underperforms so staff can inspect shelves.
How does this story land?
One tap per reader.
Quick poll
Can CPU-based AI models replace GPUs for most operational enterprise tasks?
One tap. No sign-up.
Ask this story
Still have a question?
Answers use only what is in this story. They are written by AI and can be wrong.