November 21, 2026
Heard in AI

Supabase says it bought Turso to run billions of small databases

Supabase says agents or AI tools create 70% of its new databases and that it expects to need billions a month. Its plan: start apps on a light Turso-based tier and upgrade them to Postgres. Turso's Postgres layer is unfinished.

A briefing reports one development at a point in time. We may correct or clarify it later; a new development gets a new briefing. How our formats work

Based on Latent Space, episode published October 28, 2026

Supabase sells ready-made backends: the database, user logins, file storage and server code that sit behind an app. More and more of the databases it creates come from AI agents and AI app builders, and the company says they are arriving faster than its usual approach can handle. On an episode of the Latent Space podcast, hosted by swyx with Supabase co-founders Paul Copplestone and Ant Wilson as guests, the discussion put the current rate at about 5 million new databases a month and said the company expects to need billions a month. Wilson described what would happen if each of those got its own dedicated Postgres cluster: "we'll probably run out of physical computers in the world pretty fast."

The company's written figures are a little lower. Supabase's funding release of October 2, 2026, which also announced $150 million in new funding led by GIC, reported more than one million new users and four million new databases per month. It said agents or AI tools created 70% of those new databases. All of these numbers come from the company.

The same release announced Supabase's answer to the problem: it has acquired Turso, a startup that rebuilt the SQLite database. The episode was published on October 28, but it was recorded just before the deal became public, which is why the conversation refers to the announcement coming "tomorrow."

Why a dedicated server per app does not scale

Postgres, the open-source database Supabase is built on, normally runs as a server: a program that stays running, holding memory and processor time even when nobody is using the app. That suits a business with real customers. It is wasteful for millions of prototypes.

SQLite works differently. As the discussion noted, it is essentially a file format: the whole database lives in one file that an application opens directly, without a separate server. That makes it very cheap to create and store. Its limitation is that only one writer can change the database at a time.

What Turso changed

Turso rewrote SQLite in the Rust programming language and added what database engineers call MVCC, or multiversion concurrency control. Instead of locking the whole database for one writer, the engine keeps several versions of each changed row, lets transactions proceed side by side, and checks for conflicts only when each one commits. Transactions that touch different rows can both succeed; conflicting ones can be aborted.

Turso's October 2025 technical preview tested this with a benchmark that inserted 100 rows per transaction. Ordinary SQLite reached roughly 150,000 rows a second, or about 80,000 when each transaction also did one millisecond of computation, and adding threads did not make it faster. With a single thread and no extra computation, SQLite beat Turso. With several threads, Turso was 16% faster, and with eight threads and one millisecond of computation it was four times faster. The preview had clear gaps at the time: it could not create indexes, storing full copies of each row version used a lot of memory, and shared locking limited how far it scaled.

What drew Supabase's interest was a different post. The conversation traced the approach to Turso back to a blog post by Turso's Glauber Costa that showed a route to Postgres compatibility on top of Turso's file format. In that July 2026 post, Costa and Pekka Enberg propose turning Turso into a shared engine with several SQL "front ends." A Postgres front end would read Postgres-style queries and translate them into Turso's own internal instructions, reusing the same storage and concurrency machinery underneath. The authors describe it as a working foundation rather than a finished product, with no published packages yet. They aim for broad compatibility with common applications rather than identical behavior for every Postgres feature, and they say some Postgres extensions may never load.

The plan: Spark, then Postgres, then Multigres

On the podcast, the plan was laid out this way. Turso keeps its own brand and its focus on SQLite. Supabase will add a new, smaller instance type, called Supabase Spark, that works with all of its existing software libraries. The idea is one cheap database for every idea or prototype. When a project takes off, it upgrades to a full Postgres database and, at the top of the ladder, to Multigres, Supabase's system for running Postgres across many machines.

The discussion was frank about where the lightweight design stops. Transactional databases do a lot of reads and writes, and the conversation made clear that Supabase would not run something at massive scale this way; that is the job of Multigres. The small tier is for the billions of projects that need to be cheap first.

Turso's announcement of the deal describes the same graduation path into standard Postgres and later Multigres, without using the Spark name. It frames the goal as inexpensive, isolated databases for each agent, task or user. Existing Turso databases, programming interfaces and workflows continue, Turso Database stays open source, and Costa becomes Supabase's Head of Agentic Services, alongside Enberg and the Turso team.

Asked whether this was a long-held goal, the answer on the podcast was that scaling both up and down was something Supabase had been trying to work out for a long time, and that it saw a way only once newer technology arrived.

"Until they get traction"

When the conversation suggested that most of these databases are tiny, vibe-coded apps (software built by describing what you want to an AI tool), Wilson pushed back. They are small, he said, "until they get traction." A project that suddenly lands on the front page of Hacker News has to keep working far beyond that moment. That is why Wilson called a smooth route between backends, with migrations that never take the app offline, "just like a need to have."

His example was MagicSchool, a Supabase customer. Wilson said its founder was a teacher who saw colleagues spending time on tasks that could be automated. He offered it as one case of startups that now begin without a technical co-founder or a chief technology officer, and said these are no longer isolated cases. "Everyone's like a builder now," Wilson said, calling Supabase's growth over the past 18 months "crazy." He expects a much larger market of almost consumer-like builders: people inside companies and people launching their own startups. That, he said, means far more databases than dedicated Postgres clusters could serve.

Object storage underneath

The newer technology discussed on the podcast is object storage, the kind of cheap, very durable file storage best known from Amazon's S3. The discussion noted that object storage has added features that let it be used a little like a database. The conversation credited two changes in particular: S3's move to strong consistency, which means a read sees the latest write, and S3 Express One Zone, a faster tier.

The write path was described this way on the podcast. A change is first written to a local drive. Before the database tells the app that the write succeeded, it batches the change to something durable, S3 Express. Later, in the background, the data drains into standard S3. The estimate given in the discussion was that writes to Express take one to two milliseconds.

The published measurements are higher. In an article written with AWS, Turso says it confirms a transaction only after batched entries from its write-ahead log (the running record of changes) reach S3 Express One Zone, keeps a local NVMe drive as a read cache, and periodically saves checkpoints to standard S3. Batching writes from many databases together spreads out request costs but adds several milliseconds of waiting. In a test of 1,000 operations on 4 KB objects within one zone, writes averaged 6.4 milliseconds and reads 3.8 milliseconds, not counting batching delay or application time. In return, the servers running the database become disposable, because everything can be recovered from object storage.

The conversation also touched on the difference between durability, meaning data is not lost, and availability, meaning it can be reached at a given moment. In the discussion, Express was described as less durable than standard S3. AWS's storage documentation does not support that: it lists both classes as designed for 99.999999999% durability. The difference is placement and availability. Standard S3 spreads data across at least three availability zones (separate data-center locations) and targets 99.99% availability; Express keeps its copies in one chosen zone and targets 99.95%.

Asked whether fast in-memory caches such as Redis are still needed, the answer in the discussion was yes, because object storage is nowhere near fast enough for that job. The cache layers run from memory, to local NVMe drives, to S3 Express, to standard S3.

The prize: default backend for AI labs

The scale argument ends with the AI companies themselves. Asked what a partnership with a frontier lab would look like, the conversation separated being listed in a lab's app store from being the default backend when a user's chatbot builds an app. The answer was that Supabase hopes to be one of those default providers, and that it could not be before because it could not handle the labs' scale.

The example given was ChatGPT, put in the discussion at about 1.2 billion monthly users. If each of them launched one database a month, Supabase would have to launch 1.2 billion Postgres servers. According to the discussion, Supabase had not been able to reach billions of databases a month until the Turso acquisition gave it the technology to do so.

Share this article

Go to the original

Sources & further reading

  1. 01
  2. 02
  3. 03
  4. 04
  5. 05
  6. 06

Connected ideas and articles

From the conversation

Podcast episodes

Latent Space

Supabase’s Plan for a Billion AI-Built Apps

Episode published This article draws on 7:13–13:00, 21:40–23:18 and 50:04–50:48 (approximate times)

Article history

Updates to this article

Tags

Modal's CEO says the sandbox bet has no obvious successor

Modal CEO Erik Bernhardsson says sandboxes, now about a third of revenue, began as a prototype built in days. He says he sees no similar cheap bet today and calls his next ones, in low-latency inference, training and VMs, bigger and riskier.

8 min read