IBM puts Granite time-series inference inside Confluent Cloud’s streaming SQL

IBM and Confluent have opened early access to four Granite time-series foundation models inside Confluent Cloud, making forecasting and anomaly detection callable from the same Flink SQL pipelines that process live Kafka data. The IBM announcement says the initial service runs on Confluent Cloud on AWS, with Confluent Platform support for on-premises and hybrid deployments planned later.

What changed

The integration exposes IBM’s models through Confluent’s existing AI_FORECAST and AI_DETECT_ANOMALIES Flink SQL functions. Teams can switch among the supported models without redesigning the surrounding streaming pipeline, while Confluent manages serving infrastructure, scaling and runtime operations.

The early-access set includes PatchTST-FM-r1 for probabilistic forecasts, FlowState-r1.1 for point forecasts, TTM-r3 for efficient processing of many series with control variables, and TSPulse for anomaly detection, classification, similarity search and gap filling. IBM describes the models as ranging from 1 million to 260 million parameters and requiring no GPU.

Inference results are written back to Kafka topics. That lets alerting systems, dashboards, lakehouses and AI agents consume the output without a separate handoff from a machine-learning environment. IBM also says those inference pipelines inherit the platform’s schemas, lineage and access controls, while Kafka’s replayable topics provide a record for audits, troubleshooting and model evaluation.

Who it affects

The immediate audience is data-streaming teams that already use Confluent Cloud and Flink SQL but have treated forecasting as a separate workload. IBM’s examples include scoring payment activity before a transaction completes, forecasting demand across product catalogs, monitoring industrial telemetry and comparing current signals with historical patterns.

The design also matters to platform teams trying to reduce the operational boundary between event processing and model serving. Small CPU-capable models can sit closer to the stream than a general-purpose GPU inference service, although the early-access label means teams should treat the interface and deployment limits as subject to change.

What to do

Confluent Cloud users can enroll in the early-access program and test the SQL functions against non-critical streams. A useful evaluation should compare forecast quality and anomaly sensitivity across the four models, then verify how results, schemas and lineage appear in downstream Kafka topics.

Hybrid and on-premises teams cannot use the announced Confluent Platform path yet. They can still evaluate whether the proposed architecture fits their governance model, but IBM gives no availability date for that support. Production adoption should wait for service-level, pricing and support details that the announcement does not yet provide.