Confluent’s Lightning Tables serve live Kafka data to AI agents in about 100 ms

Confluent, now an IBM company, said Lightning Tables is generally available on AWS. In its announcement, it called this the fifth pillar of its data streaming platform, built for a “Streamhouse” that keeps business data current for AI agents.

Lightning Tables turns Kafka streams into queryable tables

Lightning Tables on Confluent Cloud is a read-only serving layer. It stores Apache Kafka data durably and materializes it into tables that update as changes arrive. Applications can query those tables through SQL or a direct API.

Confluent’s own testing, which has not been independently verified, gives these figures:

• About 100 milliseconds from new data being created to being served
• Query latency of 10 to 100 milliseconds
• Throughput above 1,000 queries per second

The company says the service removes the need for separate ETL pipelines, databases or caches to serve applications. Schemas, lineage and access policies carry through from capture to serving. Schema evolution is managed so that changes should not break downstream applications. Flink can also compute features that are stored for machine learning, such as real-time personalization.

Availability is limited to AWS for now. The announcement gives no pricing, no other cloud providers and no timeline for expansion.

Query completes a five-part platform

Confluent defines the Streamhouse as an architecture that captures, processes, governs and serves current business context. It is meant to sit beside existing warehouses and lakehouses as a real-time operational layer, not replace them. The company lists three principles: current, mission-critical and decentralized.

The platform now has five pillars. Stream is built on Kafka and the Kora engine. Connect offers more than 150 managed and custom connectors. Process uses Apache Flink. Govern covers lineage, catalog, quality, encryption and access control. The new fifth pillar, Query, has three parts:

• Tableflow publishes streaming data to lakehouses and warehouses in open table formats.
• Real-Time Context Engine lets AI agents pull up-to-date business context over MCP, the Model Context Protocol.
• Lightning Tables lets applications read the live state of the business with fast SQL queries or an API.

Confluent’s example is an airline handling a canceled flight. Flight operations, scheduling and reservations each hold a different view of the same event. A rebooking agent can then build an itinerary that looks valid in every system but fails in practice, for instance because the aircraft cannot reach a connection in time. In the Streamhouse version, location, departure time and seat availability are checked together.

A product bet inside IBM’s data strategy

The analysis here is on numbers in context. IBM completed its acquisition of Confluent on March 17, 2026, at $31 per share, for an enterprise value of about $11 billion, according to Techzine. The press release says Confluent “sits at the center of IBM’s data strategy.” This launch is among the first major product moves since the deal closed.

Not all of the Query pillar is new. Confluent first presented the Real-Time Context Engine as an early-access MCP service at its Current 2025 conference on October 30, 2025, according to HPCwire’s BigDataWire. What is new on this date is Lightning Tables going generally available and the Streamhouse label that ties the three parts together.

That is also a reason for caution. “Streamhouse” is Confluent’s own term for a category it is trying to create, in the way “lakehouse” was once coined. The press release’s latency figures come from Confluent’s internal testing, and the company says real latency varies with configuration and workload. The only customer results come from Accenture and Extend. Accenture says time-to-insight fell from months to days, with no baseline or method given. The release does not compare Lightning Tables with existing serving databases or caches on cost.

The stakes are in the architecture claim. IBM’s own writers describe a gap between the batch-oriented lakes and warehouses and agents that need one consistent view when they act. As agent identity and control come to the foreground for IBM, as in its push to give every agent an owner, an ID and a kill switch, the data those agents read becomes the other half of the problem. Shaun Clowes, General Manager of Confluent and Data at IBM, put it this way:

The next era of enterprise AI will be won by the agents that know what’s happening in the business the moment they act

What teams can do with it today

Teams already running Confluent Cloud on AWS are the clear first users. They can serve Kafka topics to applications without standing up a second database. Teams on other clouds, or those needing pricing before they plan, will have to wait for details Confluent has not published in this release. The company links a Streamhouse quickstart, a webinar and a working group page from the announcement.

Confluent also says the Streamhouse is meant to work alongside the warehouse a company already has. In practice, a team would keep reporting where it is and add Lightning Tables only for the applications and agents that need the freshest state. Whether that holds up at scale is not shown by the numbers in the release.