IBM is putting its Granite time series foundation models on Confluent Cloud, and the pitch is refreshingly blunt: stop building bespoke forecasting models one at a time. The old way meant months of expert work per model, covering only the few hundred series where the money was obvious while everything else ran on safety margins, extra inventory, and decisions made after the window had already closed.
A time series foundation model flips that economics. Train it once across vast, varied signals and it generalizes to streams it has never seen. Give it a window of measurements and it forecasts what comes next, flags how far behavior sits from normal, finds similar historical patterns, and identifies which settings best hit a target. IBM claims the models have racked up 44 million-plus downloads and were battle-tested internally and with design partners in cement, steel, pulp and paper, food, and telecom before this release. The numbers they’re floating: every point of accuracy worth millions, productivity gains of 5 to 10x, and the work shifting from data scientists to the domain experts who actually own the decisions.
On Confluent Cloud, the models run natively inside Apache Flink, which means no separate ML platform, no data warehouse extraction, and no glue code between pipelines and inference. Confluent handles model serving, infrastructure, scaling, and runtime operations. You call the models directly from Flink SQL for real-time anomaly detection and forecasting. The stateful nature of forecasting—where the next value only means something against recent history—is managed by Flink’s keyed, fault-tolerant state, so each model gets the context it needs without a database hit per call.
The initial rollout is on Confluent Cloud on AWS, with Confluent Platform following for on-premises and hybrid environments. That sequencing matters for the many industrial companies still running plants where cloud connectivity is a negotiation, not a given. IBM’s framing that a signal’s value decays with time—a pump caught drifting today is a work order, the same pump next week is an outage—is the kind of operational truth that has driven real-time analytics adoption in manufacturing for years, but until now the model layer was always the bottleneck. This integration removes that bottleneck, or at least moves it somewhere less expensive.
