IBM and Confluent Bring AI Forecasting Directly Into Live Data Streams
A new partnership lets businesses run AI predictions on data as it arrives, without a data science team or months of setup.
Key points
- IBM's Granite Time Series models, which have logged more than 44 million downloads, are now in Early Access on Confluent Cloud.
- The models run inside Apache Flink, a tool that processes streaming data in real time, so forecasts happen at the moment business conditions change.
- Teams can call the models using standard SQL, a common database language, with a single line of code and one parameter to switch between models.
- IBM tested these models in its own operations and with partners in cement, steel, food manufacturing and telecoms before public release.
- Productivity gains from pilot deployments ran 5 to 10 times compared to previous approaches, according to IBM.
Most AI tools work on stored data: you gather it, clean it and read a report hours later. By the time that report lands, the moment it described has already passed.
IBM and Confluent are trying to close that gap. Their new partnership, first detailed on Hugging Face, puts IBM's Granite Time Series models, a family of AI models trained to understand how measurements change over time, directly inside Confluent's data streaming platform. The result is a forecast or an anomaly alert at the exact second the data arrives, not hours later. We covered IBM's broader Granite model family on 25 August in our look at Granite 4.2; this move takes a different branch of that family into live industrial data.
What does this actually do for a business?
It lets one AI model watch thousands of data streams at once and flag problems or predict outcomes without anyone having to build a custom model for each stream.
Take a chocolate factory. A single production line sends temperature and output readings every few seconds. Today a team might watch those readings against fixed thresholds: if the temperature crosses a set number, an alarm fires. Fixed thresholds miss slow drifts, though. A line can creep toward a bad run for hours and never trip an alarm until product quality fails.
Drop a Granite Time Series model into that stream and the picture changes. It forecasts what the line will produce through the evening shift. It scores today's run against how the line normally behaves on dark chocolate, so a gradual drift surfaces early. Plant history gets searched automatically, and the engineer learns how the last similar run ended.
No data scientist required. The demand planner or process engineer who already owns that decision can set it up themselves.
How does the technical side work?
Confluent's platform moves business data continuously between systems, the way pipes move water. Apache Flink, a processing engine inside Confluent Cloud, handles the moment-to-moment computation.
The four Granite Time Series models slot directly into Flink and are called through two existing SQL functions: AI_FORECAST and AI_DETECT_ANOMALIES. SQL is the standard language most database tools already use, so teams don't need to learn anything new. Switching models takes a single parameter change, no pipeline redesign.
| Model focus | What it answers |
|---|---|
| Forecasting (TTM) | What comes next, with a range of outcomes |
| Anomaly detection | When a stream stops behaving normally |
| Similarity search | Which historical pattern looks like this one |
| Optimization | Which settings best hit a target |
All four are in Early Access on Confluent Cloud running on AWS, with an on-premises version via Confluent Platform to follow.
One privacy point: data stays inside Confluent Cloud during inference, so your sensor readings and business signals don't travel to a third-party model-serving service. Confluent's existing access controls apply throughout. On cost, running models this way eliminates dedicated GPU servers, the specialised chips that do heavy AI number-crunching, and there are no fees for moving data in or out of the cloud.
Common questions
Do I need a data science team to use this?
No. IBM and Confluent designed these models for domain experts such as demand planners and process engineers who can call them directly using standard SQL, without writing custom model code.
Is this available right now?
The models are in Early Access on Confluent Cloud on AWS today. Support for Confluent Platform, covering on-premises and hybrid setups, is coming next.
What kinds of data can these models handle?
Anything measured over time: electricity load, sensor telemetry, payment transactions, application metrics. IBM piloted the models in cement, steel, food manufacturing and telecoms before the public release.
