Edelbyte Raises $70M for a Data Platform Built for AI Agent Queries

Edelbyte wants to be the place AI agents go to search and analyze their data, and it now has $70 million to try. The Tel Aviv company is pitching one platform that handles conventional records and vector data, and it claims query latency under 100 milliseconds on multipetabyte datasets. That's a demanding claim. My read is that the performance numbers will matter more to buyers than the funding headline.

The funding and the team behind it

The $70 million came in two rounds. Sequoia and Quiet Capital led them, with Greenoaks and Swish Ventures also participating. CTech reports that Edelbyte is valued at $500 million.

The company was founded last year by four former Microsoft executives: Avner Aharoni, Evgeney Ryzhyk, Alexander Sloutsky and Yoni Leibowitz. Aharoni is the chief executive. The founders say they helped build Azure Data Explorer, the real-time analytics service that many engineers informally call Kusto after its query language. That lineage explains a lot about the product's design priorities. Azure Data Explorer was built for fast scans over large streams of time-stamped telemetry, and Edelbyte is extending that engine toward the mixed workloads that AI systems generate.

Where it sits against Snowflake and Databricks

Edelbyte positions itself as an alternative to Snowflake and Databricks, the two platforms most enterprise data teams already consider for warehousing, lakehouse work and analytics. Those incumbents have spent years building ecosystems around SQL analytics, data sharing and machine learning pipelines. A newer entrant has to show a clear reason to move data or add another system, and Edelbyte's argument rests on two things: unified storage for vectors and standard records, and fast queries across very large volumes.

The company says a single deployment can hold up to hundreds of petabytes. That's a capacity claim, not a benchmark, and it's worth reading it that way. Storage ceilings rarely decide purchases. Query behavior under real concurrency, cost per query and operational overhead tend to matter far more once a team is running production workloads.

Vectors next to records

The vector support is where the AI framing becomes concrete. AI agents keep much of their working knowledge in vectors, which are numerical representations that let a system find items that are semantically similar rather than identical. A product catalog is a good example: a vector lets a system surface items that look or behave like the one a shopper just viewed, even when the keywords don't match.

Keeping vectors in the same system as conventional records has a practical appeal. Teams running retrieval-augmented setups often end up with one database for transactional or analytical records and another for embeddings, then write glue code to join results. Co-locating the two can simplify that architecture, but it also means the platform has to perform well on both access patterns at once, which is harder than doing either well alone.

Query types and graph processing

Edelbyte says the platform supports several query styles. A search engine can pull records that users request. A business intelligence tool can run aggregate calculations. The platform also includes graph processing, an approach for analyzing data points that relate to one another, such as records created inside the same system or entities linked by shared events. Graph queries are often where general-purpose analytics engines struggle, so whether Edelbyte's implementation holds up will be a useful test for technical evaluators.

Latency claims and how the speed is achieved

The headline performance claim is that the platform can query multipetabyte datasets with under 100 milliseconds of latency. Edelbyte says it gets there with several optimization methods, though the company's description stays at a high level. That leaves open questions a buyer should press on: what the test datasets looked like, how many concurrent queries were running, what query shapes were measured, and whether results hold for cold data rather than cached hot paths. Sub-100ms latency at petabyte scale is achievable for narrow, well-indexed queries. It's a much harder promise for complex joins across messy schemas.

Speed also depends on the index strategy. Edelbyte generates indexes, which act as shortcuts that help the engine locate specific items in a large dataset. An index only pays off when it matches how an application actually reads data, and tuning that match is usually a manual, iterative job for data engineers. The company says its platform automates part of that index customization workflow. If that automation works as described, it removes a common source of performance drift, where queries slow down as access patterns change and nobody has time to re-tune.

Query federation: fewer copies, different trade-offs

The more architecturally interesting feature may be query federation. Analyzing data normally means copying it into a dedicated processing environment first. Edelbyte says it can instead run queries where the data already lives, and that this cuts processing time and reduces the bandwidth spent moving data between systems.

That's a sensible idea, and it reflects a broader shift toward keeping compute close to storage. It also introduces a trade-off. When a query touches systems with different performance characteristics, the slowest source sets the pace, and the platform's latency guarantee becomes harder to state cleanly. Teams running federated setups typically find that governance, access control and consistency across sources take more design time than the query engine itself. Anyone evaluating Edelbyte should ask how it handles those cases, not only how it handles the fast ones.

Customers, hiring and what to watch

Edelbyte says several Fortune 500 customers already use the platform to run petabyte-scale data environments. The company plans to put the new capital toward expanding its installed base and hiring. Those are the typical priorities for a young infrastructure company with a $500 million valuation, and they signal that the next year will be about proving the product in enterprise accounts rather than in demos.

Chief executive Aharoni framed the opportunity in terms of legacy infrastructure. He argued that systems built before AI became central to business operations will struggle as query volume and complexity grow. That's a reasonable thesis. The open question is whether buyers will replace or extend existing platforms to meet that demand, and whether a newcomer can offer enough operational simplicity to justify the migration cost.

Three things will tell the story. Independent benchmarks that publish methodology would move the latency claim from marketing to evidence. Pricing and cost-per-query figures would show whether the architecture pays off economically at scale. And the company's federation and indexing features will need to hold up across the messy, heterogeneous data estates that most large enterprises actually run. Edelbyte has a credible pedigree and a clear pitch. Now it needs numbers that outside engineers can check.