Neon
Databricks paid $1 billion for a company with roughly $25M in revenue — because 80% of the databases being created on Neon were provisioned by AI agents, not humans, up from 30% a year earlier.
What is Neon?
Neon is a fully managed, open-source (Apache 2.0) serverless Postgres platform whose defining technical decision explains everything else about it: rather than bolting a proxy layer onto standard Postgres, Neon rebuilt what sits underneath it so the database itself is genuinely stateless, separating storage from compute entirely. Your Postgres compute runs as standard, disposable server processes while your data lives durably in cloud storage — which is precisely why branching costs next to nothing (a branch just points at existing data rather than copying it) and why scale-to-zero is a real architectural property rather than a marketing claim layered on top of an always-on database. Neon can spin up a fresh Postgres instance in under 500 milliseconds, a speed that matters enormously for one specific, fast-growing use case: infrastructure that needs to provision its own databases on demand.
That use case is the whole story behind Neon's most remarkable 2026 headline: on May 14, 2025, Databricks agreed to acquire Neon for approximately $1 billion — a striking valuation for a company reportedly generating around $25 million in annual recurring revenue at the time. The stated rationale, straight from the acquisition announcement, is genuinely fascinating: roughly 80% of the databases being created on Neon were being provisioned automatically by AI agents rather than human developers, a figure that had jumped from just 30% in under a year. As AI coding agents and autonomous systems increasingly spin up their own infrastructure, a database that can be created instantly via API, branched cheaply for testing, and scaled to zero when idle is exactly the primitive that workflow needs — and one year after the deal closed, Neon had grown from roughly 6,000 databases to more than 700,000. The acquisition also funded real, substantial price cuts: compute costs dropped 15-25%, storage fell about 80% (from $1.75 to $0.35 per GB-month), the free tier's compute allowance doubled from 50 to 100 CU-hours a month, and the Scale tier gained SOC 2 Type 2 and HIPAA eligibility.
The Databricks acquisition, its AI-agent rationale, and post-acquisition growth figures are drawn from multiple independent, dated 2026 sources (Tech Insider, BuildMVPFast) corroborating the original 2025 acquisition announcement (AlternativeTo). The architectural explanation and pricing changes are confirmed by an independent technical review (GetAutonoma) and detailed pricing analyses (Vela/Simplyblock, SelfHost.dev).
Key features
Storage/compute separation
A genuinely stateless architecture underlying true scale-to-zero and near-free branching.
Instant, cheap branching
A branch points at existing data rather than copying it, keeping costs minimal.
Sub-500ms provisioning
Spin up a fresh Postgres instance fast enough for on-demand agentic workflows.
Real scale-to-zero
Compute genuinely pauses when idle, rather than running continuously in the background.
Auto-scaling cost ceilings
Set a maximum CU limit per branch to cap costs even during traffic spikes.
SOC 2 Type 2 & HIPAA
Compliance eligibility added to the Scale tier post-acquisition.
Pricing
Free
- Up to 2 CUs, scale-to-zero always active, 1 project
- Doubled from 50 CU-hours following the Databricks acquisition
- Genuinely generous for intermittent-traffic side projects
Launch
- Whether your app ever sleeps determines your real bill, not your data size
- Keeping compute always-on for sub-300ms response runs roughly $19/month
- Set a per-branch auto-scaling limit to cap costs during spikes
Scale / Enterprise
- Added post-acquisition, reflecting Databricks' enterprise customer base
- Best for regulated industries needing serverless Postgres at scale
- Only 8 AWS regions currently available — worth checking for your locale
A real, practical gotcha: frontend polling patterns (React Query intervals, WebSocket keepalives, health checks) can prevent scale-to-zero from ever activating, silently keeping your "serverless" database running continuously — audit what's keeping connections alive if cost optimization matters. Prices reflect Neon's published pricing as of July 2026; older guides citing $5 or $69 minimums predate the December 2025 change and are outdated.
Available models
Integrations & platforms
Pros, cons & best for
Pros
- Genuinely real scale-to-zero, architecturally, not just a pricing gimmick
- Backed by Databricks' scale, funding substantial post-acquisition price cuts
- Sub-500ms provisioning is genuinely ideal for on-demand agentic workflows
Cons
- Usage-based billing means a 10x traffic spike genuinely means a 10x bill
- Frontend polling can silently prevent scale-to-zero from ever activating
- Postgres-only, with no MySQL or MongoDB option, and limited to 8 AWS regions
Best for
- Serverless and edge app stacks (Vercel, Cloudflare Workers) with variable traffic
- AI agents and automated workflows that provision their own databases
- Not the pick if you need a full backend platform bundling auth and storage too
Take a look inside
Alternatives
For a full bundled backend platform or MySQL-based scaling instead:
Our verdict
Neon's core architectural bet — genuinely separating storage from compute rather than faking serverless behavior on top of a traditional database — has been validated in the most concrete way possible: Databricks paid roughly $1 billion for a company generating about $25 million in revenue, specifically because AI agents provisioning their own databases had become the platform's dominant use case. That's a genuinely remarkable signal about where infrastructure demand is heading, and the resulting price cuts (storage down 80%, compute down 15-25%, free tier doubled) are a real, tangible benefit passed on to every user regardless of whether they're an AI agent or a human developer. The honest thing worth internalizing clearly: usage-based billing means your bill genuinely tracks your traffic, for better or worse, and a sneaky frontend polling pattern can quietly keep your database running when it should be scaled to zero. For serverless and edge-native stacks with variable traffic, and especially for AI-driven, on-demand infrastructure provisioning, Neon remains one of the best, most technically credible choices available.
FAQ
Why did Databricks pay $1 billion for Neon?
The stated rationale was that roughly 80% of the databases being created on Neon were provisioned automatically by AI agents rather than human developers, up from 30% a year earlier — a figure reflecting the growing need for infrastructure that AI systems can spin up on demand.
Is Neon's scale-to-zero a real technical feature, or marketing?
Genuinely real — Neon rebuilt Postgres's underlying architecture to separate storage from compute entirely, making the database stateless, which is why compute can actually pause completely when idle rather than just appearing to.
Why does the same database cost $3 one month and $680 the next on Neon?
Because Neon bills purely on usage — the swing has nothing to do with your data size and everything to do with whether your application's compute is scaling to zero when idle or running continuously without realizing it.
What can accidentally prevent Neon's scale-to-zero from working?
Frontend polling patterns like React Query refresh intervals, WebSocket keepalives, or health checks can keep a connection alive continuously, silently preventing your "serverless" database from ever actually scaling to zero — worth auditing if cost matters.
Did Neon's pricing change after the Databricks acquisition?
Yes, substantially — compute costs dropped 15-25%, storage fell about 80% (from $1.75 to $0.35 per GB-month), the free tier's compute allowance doubled, and monthly minimums were removed entirely by December 2025.
Does Neon support MySQL or other databases besides Postgres?
No, Neon is Postgres-only — if your stack requires MySQL or a document database, alternatives like PlanetScale (MySQL) or MongoDB Atlas would be a better fit.