Neon Postgres

The average production database on Neon changes size about once every 81 seconds, using only the compute it needs. We wrote a deep dive into how we run autoscaling: https://lnkd.in/gPHnpEzZ

The Neon CLI is moving fast. Just in the last weeks, we added tooling for agents, Postgres debugging, branch-first workflows, and a lot more: https://lnkd.in/gbhScij2

Postgres on S3 is the right design, but even more interesting is making WAL the source of truth. This is how the lakebase storage works - we wrote a deep dive: https://lnkd.in/gHGTPueN

Our AI Gateway is powered by Databricks Foundation Model APIs, a stack with years of inference engineering behind it: https://lnkd.in/gHmNTMFB Just one example: implementing prompt caching took per-replica input-token throughput up 2.5× and P50 latency down 3× on gpt-oss, at only a ~30% cache hit rate This optimization (prompt caching) matters most for agents: coding agents and multi-step tool loops resend the same system prompt, tool schema, and examples on every turn, so caching the shared prefix means the server skips prefill instead of rebuilding it from scratch each time. This is only one piece of a longer list of works that includes continuous batching, KV-cache paging, a TensorRT-LLM-era backend with custom kernels, quantization, per-GPU and per-cloud tuning, and others. AI Gateway is in beta and tokens are free during the period. Take advantage of it and benchmark it on your own agent loop: https://lnkd.in/gAHFagyb

Tokens are a backend primitive in 2026. You can now call LLMs directly from your Neon branch via Neon AI Gateway, backed by Databricks Foundation Model APIs: https://lnkd.in/giy_Ddzt We’re building backends, and LLMs are part of the picture. Neon AI Gateway treats inference as a first-class primitive next to Postgres, Object Storage, Functions, and Auth - model calls live on the same project, same branch, same deploy path as everything else. What you get: - One credential for a wide catalog of frontier and open-weight models - Branch-scoped endpoints, like the rest of the Neon backend - Databricks-scale serving under the hood, the same Foundation Model APIs stack already moving serious trillions of tokens - One Neon bill, with no markup (we pass through lab pricing) AI Gateway is in beta and ready to test. Inference is free during the beta period - take advantage of it: https://lnkd.in/gAHFagyb

Another piece of the Neon backend: functions. For jobs that involve the database, declare a Neon Function on your branch - it runs next to your data and stays long-running enough for agents and realtime: https://lnkd.in/gAicuaiE

PGlite + real-time sync have emerged as key primitives in an era where millions of apps are deployed by agents. We’re excited to announce Electric is joining team Neon at Databricks to build the world's most advanced Postgres backend platform. https://lnkd.in/eyBhT-bi

We're expanding our branching experience to the entire backend, starting with files: you can now branch your database AND your buckets in Neon, without duplicating storage https://lnkd.in/gBbchaJs

A 4B open-source model post-trained with castform matched GPT-5.6 Sol on search accuracy, at ~100x lower cost. The corpus backend for the whole pipeline was Neon + Lakebase Search: https://lnkd.in/gdw3gasf