Neon

What Neon actually is

Neon is managed PostgreSQL rebuilt with storage and compute as separate layers. That split is what makes its three headline features work: compute that suspends when the database is idle (scale-to-zero), copy-on-write branches created in seconds, and point-in-time restore. The company was founded in 2021 by Nikita Shamgunov, Heikki Linnakangas, and Stas Kelvich — Linnakangas is a longtime Postgres core contributor — and Databricks announced its acquisition in May 2025 for about $1 billion. Neon continues to run as a standalone product under Databricks with its own pricing, console, and roadmap.

The offering has grown past "serverless Postgres" into a set of backend primitives: Lakebase Postgres, Managed Better Auth, a Data API, Functions, Object Storage, and an AI Gateway. In our view the Postgres core is the mature part of the platform; the surrounding services are newer and not available in every region yet. If you come for the database and treat the rest as optional, the picture is coherent.

Two properties matter most in practice. First, compute is billed per CU-hour, where 1 CU is roughly 4 GB RAM plus CPU plus SSD, and suspended compute costs nothing — an idle side project on the Free plan costs literally $0. Second, branching behaves like git for your database: every branch is a copy-on-write clone, so branching a large database is instant and child branches only bill for the bytes actually written on them. Together these two traits explain why the platform became a default for preview environments and for AI coding tools that provision databases programmatically — Databricks cited telemetry at acquisition time saying roughly 80% of databases provisioned on Neon were created automatically by AI agents rather than by humans. That is context rather than a selling point for us, but it explains where the product is heading: API-first, per-branch environments, cheap idle state.

Pricing: what you really pay

Neon publishes unit rates instead of opaque plan bundles, which we consider the honest way to sell compute. Prices below are the published list rates.

Free — $0/month. 100 projects, 100 CU-hours per project per month, 1 GB of Postgres storage per project (20 GB per account), 5 GB of Object Storage per project, 5 GB of public network transfer per project, 10 branches per project, one manual snapshot, a 6-hour history window, and Managed Better Auth for up to 60,000 monthly active users per project. Scale-to-zero is always on here: compute suspends after 5 minutes of inactivity and cannot be disabled on this plan.

Launch — pay-as-you-go. $0.106/CU-hour, storage at $0.35/GB-month, extra branches at $1.50/branch-month, instant restore (history) storage at $0.20/GB-month, scheduled snapshots at $0.09/GB-month, Object Storage at $0.023/GB-month. Autoscaling up to 16 CU (64 GB RAM), 500 GB of transfer per project included then $0.10/GB, history window up to 7 days. Support at this tier is billing support only.

Scale — pay-as-you-go at higher unit rates. $0.222/CU-hour with the same $0.35/GB-month storage. Fixed compute sizes up to 56 CU (224 GB RAM), history up to 30 days, IP allow rules, private networking at $0.01/GB, metrics and logs export, HIPAA and SOC 2 availability, an uptime SLA, and standard, business, or production support options.

Small details we like: invoices under $0.50 are not collected; you can set a hard cap on autoscaling to put a ceiling on compute spend; spending notifications exist on paid plans. Details to watch: egress beyond the per-project allowance runs at $0.10/GB, and history storage bills at $0.20/GB-month until you shorten the window. Because pricing is metered, a runaway workload moves your bill in real time — there is no fixed monthly number to hide behind. For a flat-price comparison against VPS-style hosts, see /benchmarks and our /category/database-hosting listings.

Performance & infrastructure

Neon runs on public cloud regions rather than its own data centers. The docs list eight AWS regions: N. Virginia, Ohio, Oregon, Frankfurt, London, Singapore, Sydney, and São Paulo; the same documentation publishes NAT gateway addresses for Azure regions (eastus2, gwc, westus3), so parts of the platform run on Azure too. The region is fixed at project creation — you cannot move a project, only create a new one and migrate. Storage is multi-AZ on all plans, and read replicas, connection pooling, and Postgres extensions (pgvector, PostGIS, TimescaleDB and others) are included across plans.

On the performance question, we keep our claims narrow. Neon describes reactivation after suspend as "within a few hundred milliseconds", and scale-to-zero applies to computes up to 16 CU — anything larger stays always-on. We have not benchmarked Neon ourselves; our dated sysbench, fio, and iperf3 runs cover VPS instances, not managed databases, so we will not attach numbers here. What we can say architecturally: separating storage from compute removes the usual single-node ceiling, autoscaling adjusts within each plan's CU limit, and the copy-on-write branch model makes ephemeral environments effectively free until they accumulate writes.

Neon's own feature set also flags a real operational nuance: logical replication from Neon keeps compute active while subscribers are connected, so a replica subscriber prevents scale-to-zero. For cost control, that is the kind of detail worth knowing before you wire up replication.

Where it falls short

Cold starts are the price of scale-to-zero. A few hundred milliseconds is fine for a preview branch and annoying for a latency-sensitive production API that sees sporadic traffic. If your workload is spiky but user-facing, either keep compute always-on (paid plans allow disabling scale-to-zero) or accept the p99 hit.

Usage-based billing cuts both ways. There is no overage cliff, but there is also no predictable invoice unless you impose caps yourself. Teams coming from fixed-price Postgres hosting often underestimate egress and history storage, which accrue quietly at $0.10/GB and $0.20/GB-month respectively.

The SLA story is thin below the top tier: the uptime SLA is listed on Scale only, and the published pricing page does not state a percentage, so we cannot verify the number. Support is community-only on Free and billing-only on Launch — production support means Scale or an enterprise conversation.

Region immutability is a practical constraint: changing regions means a new project and a migration, which is fine for greenfield apps and awkward for anything with heavy external connections pointing at a fixed hostname. And while the Postgres core is well understood infrastructure, the newer services (Functions, Object Storage, the AI Gateway) are rolling out region by region — treat them as capabilities to verify per region, not as guarantees. Finally, a Databricks-owned database platform raises a long-term strategic question we cannot answer from the outside: Neon operates standalone today, with its own pricing and roadmap, but the ownership is now inside a much larger company with its own agenda.

Who should pick this

Our pick when your workload is intermittent or your environments are numerous. Developers building on Next.js or serverless backends, teams that want a real Postgres preview branch per pull request, and anyone prototyping with AI coding tools get genuine value from the Free plan and honest marginal costs on Launch. The API-first provisioning model also makes it our default recommendation for platforms that need to create databases programmatically.

One more practical note for teams migrating in: Neon speaks standard Postgres wire protocol, so most drivers, ORMs, and dump-and-restore tooling work unchanged, and connection pooling is built in rather than something you bolt on with an external proxy. That lowers switching friction considerably compared with database platforms that wrap Postgres in a proprietary API. We would look elsewhere when you need a fixed monthly price, when your database must sit in your own country-specific data center under your own network policy, or when you want hands-on production support without paying Scale-tier rates. Traditional managed Postgres (RDS, Azure Database) still wins on operational guardrails for enterprises with compliance-heavy processes; plain VPS Postgres wins on price at steady 24/7 load, since always-on compute billed per CU-hour is more expensive than a small server at full utilization.

Key takeaways

• Storage and compute are separate layers, which makes scale-to-zero, instant branching, and point-in-time restore possible rather than bolted on.
• Free plan is genuinely usable: 100 projects, 100 CU-hours each, 1 GB storage per project, with compute suspending after 5 minutes idle.
• Unit pricing is published openly ($0.106/CU-hour on Launch, $0.222 on Scale, $0.35/GB-month storage) with no overage cliffs — but no fixed monthly cap either.
• Eight AWS regions plus Azure presence; region is fixed per project, so plan placement before you build.
• Uptime SLA only on the Scale plan, and the percentage is not published on the pricing page we reviewed.
• Owned by Databricks since May 2025, still operating with standalone pricing and console.

Sources

• https://neon.com/pricing
• https://neon.com/docs/introduction/regions
• https://neon.com/docs/introduction/scale-to-zero
• https://neon.com/docs/storage/overview
• https://ai.info/companies/neon
• https://techcrunch.com/2025/05/14/databricks-to-buy-open-source-database-startup-neon-for-1b