The economics of NeonDB serverless Postgres pricing

Branching workflows and Databricks-era economics

Neon uses a disaggregated architecture that separates compute from storage. This separation enables Copy-on-Write branching, which creates isolated Postgres instances in under a second without duplicating data. Databricks acquired Neon for $1 billion in May 2025, which lowered storage costs to $0.35 per GB-month and reduced compute prices by 15-25% compared to previous rates. For teams running branch-per-PR workflows, this O(1) branching model avoids the O(data size) costs of snapshotting an RDS instance or replaying migrations in Supabase. Creating a branch from a 50 GB production database takes the same time as creating one from a 1 GB database. This functionality allows every pull request to include its own isolated database, pre-seeded with the production schema, that deletes automatically when the PR closes. Neon’s Scale plan supports up to 56 CU and provides SOC 2 Type 2 and HIPAA eligibility. Neon also provides connection pooling via a PgBouncer-based pooler that supports up to 10,000 pooled connections per project. For hobbyists, the free tier includes 0.5 GB of storage and 100 compute-hours per month.

Comparing provisioned and serverless scaling

Neon scales compute to zero after five minutes of inactivity. This behavior contrasts with Google Cloud SQL, which requires paying for provisioned vCPU and memory for the entire duration of an instance’s life. A Google Cloud SQL Enterprise instance in us-central1 with 4 vCPUs and 15 GiB of memory costs $197.25 per month for compute, and high availability configurations double that cost to $394.49. For those needing a baseline, Cloud SQL offers 1-year committed use discounts of 25% and 3-year discounts of 52% for eligible compute. CockroachDB’s Basic tier provides a $15 monthly credit for 50 million Request Units and 10 GiB of storage, but its Standard tier scales starting at $0.18 per vCPU-hour. For engineering teams managing hundreds of isolated databases for different developers or preview environments, Neon’s ability to scale compute to zero provides a significant financial advantage over Google Cloud SQL’s fixed provisioning model, which charges for idle resources regardless of actual usage patterns. Neon’s Launch plan costs $0.106 per CU-hour with no monthly minimum, whereas a steady 24/7 workload with 2 CU and 50 GB of storage costs roughly $170 per month. Cloud SQL storage costs depend on the region and machine series, but standard SSD in us-central1 costs approximately $0.17 per GiB-month.

The reality of idle costs and latency

The check_availability operation in the Neon control plane creates a "wake-up tax" that prevents true idleness. This periodic ping restarts the compute, which triggers a billing cycle that can keep a low-traffic project from ever reaching a zero-cost state. A project with minimal traffic might still face a floor of 0.25 CU-hour per hour due to these automated pings. You should audit your frontend polling patterns, because these connections can prevent scale-to-zero from activating. While Neon handles up to 27,000 queries per second, the network hop to the Pageserver adds 3 to 10 milliseconds of latency for warm compute. This latency remains negligible for queries that take 20ms or more, but it affects applications requiring single-digit millisecond responses. Typical cold starts range from 300ms to 800ms, with total time-to-first-query landing between 500ms and 1 second. Neon wakes up notably faster than Aurora Serverless v2, which requires 15 seconds to wake up. Does the 500ms cold start latency in a serverless environment outweigh the savings for your specific user experience?