Why we replace Databricks.
Lakehouse was the right idea. Eisberg is the next iteration — open formats on storage you own, agent-native APIs, and governance baked in instead of bolted on.
Spark-era complexity wrapped in a UI
Databricks did Spark a service. The platform still carries Spark's mental model — clusters, jobs, notebooks. Eisberg routes per query, scales pools to zero, and presents the platform as APIs first, not a notebook.
Delta is proprietary, even when it claims to be open
Iceberg is the open standard. Apple, Netflix, and the rest of the industry shipped Iceberg for a reason. Eisberg is Iceberg-only — by intention, not by tolerance.
Notebook-first, not agent-first
Databricks bolted on Genie and a chat layer. Eisberg was designed for the world where agents are the primary users — APIs built to sub-100ms targets, MCP-native, per-action metering, identity-bound agent governance.
Governance is permission inheritance there. Enforcement here.
Databricks agents inherit the acting user's permissions — governance on-behalf-of, scoped to what lives inside their perimeter. The durable difference isn't 'we have agent identity' — the incumbents are adding that too. It's where enforcement sits: Eisberg gates every agent action at a fail-closed query chokepoint on data you own in open format, binds it to a signed Birth Certificate, gates consequential actions on predicted blast radius with two-key human approval, and chains every event into a tamper-evident audit log a regulator can replay. Compliance rule packs for regulated frameworks are in development on top of that enforcement plane.
The honest head-to-head.
| Capability | Eisberg | Snowflake | Databricks |
|---|---|---|---|
| Knowledge Layer — learns your business from operating | |||
| Tribal knowledge extracted + bound (not just linked) | |||
| Cross-domain stitching (business + software + comms entities) | |||
| Agent certificate enforced at the query chokepoint on customer-owned data | |||
| Compute-layer agent enforcement (not metadata certification) | |||
| Automated business ontology — cross-source, on storage you own | |||
| Customer-owned object storage | |||
| Open format (Iceberg) by default | |||
| Composable multi-engine router (embedded + distributed, over Iceberg you own) | |||
| Sub-100ms agent API targets | |||
| MCP-protocol native | |||
| Per-action agent metering | |||
| Policy-as-code governance at every layer | |||
| Impact-gated autonomous actions (blast radius checked pre-execution) | |||
| Autonomous data classification | |||
| Pipelines that resolve their own failures | |||
| Compounding intelligence across customers (k≥3 anonymity) | |||
| Cost ceiling via outcome pricing |
Bring us your Databricks workload.
We will translate your jobs, your tables, and your Unity Catalog policies, then run the same workload on Eisberg. You will see the cost, the latency, and the governance side-by-side.