Supabase says AI coding tools and agents now create 70% of its new databases. The figure is real evidence that software has become easier to start, and weak evidence about how much of it is finished, secured or used, because the same company counts accounts rather than active projects. Who builds is shifting; what survives is undocumented.
Supabase says 70% of the new databases on its platform are now created by AI coding agents or AI-driven tools, up from more than 60% in June 2026. The company’s press release of 2 October 2026 adds that it takes in more than 4 million new databases a month, which makes the claim one of the few measurements of AI-built software taken from the infrastructure underneath rather than from a demo on top. It deserves attention and a careful reading, because a database created by a tool records that something was started. Nobody has published how many of those starts became working, secured or used applications.
The best-known airing of the figure is a Lightcone episode from Y Combinator, released at the end of October 2026 as “What Happens When Everyone Can Build Software”, with Supabase chief executive Paul Copplestone as guest. The Lightcone is Y Combinator’s own show and Supabase went through the accelerator in 2020, so the tone is friendly by design. The sections below test the episode’s claims against the company’s releases, an independent study of agent behaviour and a security record that complicates the story.
Numbers the episode puts on the table
The episode’s description lists three headline figures: 13 million users, more than a million databases created every week and roughly 70% of new activity coming from agents. An AI-assisted write-up of the conversation by consultant Sean Weldon adds Copplestone’s account of the past year: sign-ups rising from about 3 million to 13 million, databases under management growing from roughly 8 million toward 50 million, and revenue up fourfold. It also relays two anecdotes, a taxi driver in Brazil who built a ride-hailing style app from his phone and a founder with a billion-dollar exit who now builds personal finance software as a side project.
The write-up is a machine-assisted summary, not a transcript, so its detail is indicative rather than exact. The company’s own documents cover the headline numbers. The release gives more than 1 million new users and 4 million new databases a month, and Copplestone’s blog post of 2 October says Supabase is already launching over one million databases per week. The two are consistent, since a million a week comes to about 4.3 million a month. Revenue is the figure nobody has published, so the fourfold growth is Copplestone’s statement in the episode, relayed second hand.
Thirteen million counts accounts, not apps
The same number travels under four names. The episode calls them users, the release calls them developers, Weldon’s write-up calls them sign-ups, and the Origin Notes newsletter, which took apart the company’s go-to-market, states that the figure means registered accounts and that Supabase does not publish active users. The company also reported more than 250,000 paying customers as of June 2026. Origin Notes divides that by the October user count to get roughly 2 in 100, a rough ratio because the two figures are four months apart.
The 70% has a definition problem of its own. The episode’s description says 70% of new activity; the release says 70% of new databases created by “agents or AI-driven tools”. Those are different denominators, and the second phrase covers app builders as well as autonomous agents. Origin Notes cites Lovable’s lead engineer saying that Supabase’s API handles about 90% of Lovable’s interactions with it, so a person typing a prompt into an app builder can register as an AI-driven tool. The 70% measures the route a database took into existence, not how much of the work an agent did unaided.
Defaults decide which backend agents choose
Supabase did not wait for this to happen. In the fall of 2024 a customer it had never spoken to began creating databases at a rate that made the team think, in Copplestone’s words, “We thought we’re getting DDoSed.” The customer was Lovable. Origin Notes traces the pattern to an API Supabase shipped in August 2022 so that other software could provision backends, later packaged for platforms such as Lovable, Bolt, Vercel’s v0 and Figma Make, and to open-source code and public documentation that models can read. Supabase’s own test, which the company calls an early result on a small sample, found that published Agent Skills lifted one coding agent’s success rate from 46% to 71%.
Choice by an agent is not choice by a market. Amplifying, an evaluation firm, collected 2,430 Claude Code responses across three models and four project types using open-ended prompts, and found PostgreSQL was the most common database pick at 58.4%. Origin Notes, summarising the study, puts Supabase at 24%, mostly as an all-in-one bundle of database, sign-in and storage. Amplifying itself says the results show what the agent favours and are not market adoption data, and that Claude Code built custom solutions instead of choosing a tool in 12 of 20 categories. A handful of products, among them Lovable, Bolt, Claude Code and Codex, now does much of the choosing, a concentration risk Origin Notes names for Supabase.
A database made like a file is cheap to count
Copplestone’s own post states the design goal: agents should be able to create a database as easily as creating a file, with just as little concern about cost. That is the reasoning behind the acquisition of Turso, a company that rebuilt SQLite in Rust so that one server can hold millions of databases and suspend the idle ones. According to FourWeekMBA, the Turso price, the closing date and any valuation for the new round are undisclosed, and the $150 million includes liquidity for employees without a stated split. Supabase’s releases date three rounds in twelve months: $100 million at $5 billion in October 2025, $500 million at $10.5 billion post-money in June 2026, and this one 120 days later.
My reading, which is an inference and not a company statement, is that the acquisition is built for volume that mostly will not last. Origin Notes observes that most of the 4 million monthly databases are free and that nobody outside can say how many ever pay, since revenue is undisclosed. It cites a Sacra estimate of about $170 million in annual recurring revenue in May 2026 and labels it an outside figure. When creating a database costs close to nothing, counting them says more about the tooling than about demand for finished software.
A security record that predates the boom
Row-level security is the database feature that limits which rows a user can read. When it is missing on a Supabase table exposed through a public key, anyone can query that table. In 2025 researcher Matt Palmer found exactly that in apps built with Lovable: 303 endpoints across 170 projects lacked proper protection, according to Superblocks, and a vulnerability database lists the issue as CVE-2025-48757, rated 9.3 critical, for Lovable versions before 15 April 2025. Palmer reported it on 21 March 2025. Lovable shipped a security scanner on 24 April that, Superblocks says, checked whether a policy existed but not whether it worked, and the CVE was published at the end of May.
Superblocks sells a competing governed platform, so its judgments of Lovable deserve discounting; the counts match the vulnerability record. The flaw sits in what Lovable generated, not in Supabase’s database. It still shows what happens when software provisions backends faster than anyone reviews access rules. Neither Supabase’s release nor Copplestone’s post says how many agent-created databases have row-level security switched on, and the sources reviewed do not establish how common the problem is today.
The strongest case against everyone building
The sharpest opposing argument comes from software architect Ivan Turković, who publishes essays while offering contract engineering. His February 2026 piece argues that code is a small part of software, and that debugging, deployment, security and maintenance are where non-engineers stall. He asserts that maintenance accounts for 60 to 80 percent of lifetime cost and cites no study for it, so that figure is his claim, not a measurement. It is opinion, written before the October numbers, and it never mentions Supabase.
The Supabase data neither confirm nor refute him, because they measure different stages. Four million databases a month shows that people and agents are starting software in volume. Turković’s claim is about finishing and keeping it alive, and no source reviewed counts that. Weldon’s write-up reaches the same gap about the episode’s own examples: whether software built by a taxi driver or a founder with a large exit persists as production infrastructure or stays a prototype is not addressed. Both camps argue from the same missing number, the share of projects still running six months on.
Checks a builder or vendor can run now
A person shipping an app through an AI builder on Supabase can test the one thing the record shows going wrong: confirm that row-level security is enabled on every table holding user data, then query those tables without signing in. The vulnerability record’s own advice is to enable it explicitly for tables with sensitive data. A passing scanner is not a pass, as the April 2025 episode showed.
A vendor selling to builders should read the Amplifying work as a distribution warning. The firm says a product the agent does not pick is invisible to a growing share of new projects, and newer models picked newer tools in the study, so the ranking moves with each model release. For anyone reading 13 million as market size, the verdict holds only for starts, not for finished software. Supabase could change that by publishing active projects, the share of agent-created databases with access rules enabled, and an audit by someone outside the company. Until then the 70% shows how quickly software gets started and says little about what is still running.
Author:
Jan Bielik
CEO & Founder of Webiano Digital & Marketing Agency
This article is an original analysis supported by the sources cited below
What Happens When Everyone Can Build Software
Y Combinator’s Lightcone episode with Supabase CEO Paul Copplestone, whose published description names the guest, his 2020 start in Y Combinator’s summer batch and the claim that makes everyone a builder.
Supabase Announces $150M in New Funding and Turso Acquisition
Supabase’s own release of 2 October 2026, the source for 70% of new databases created by agents or AI-driven tools, more than 4 million new databases a month and more than 13 million developers.
Supabase is acquiring Turso
Copplestone’s post stating that Supabase launches over one million databases a week and that agents should create a database as easily as a file.
PT-2025-23236 · Lovable
The vulnerability record for CVE-2025-48757, giving the 9.3 critical rating, the affected Lovable versions before 15 April 2025 and the advice to enable row-level security explicitly.
What Claude Code Actually Chooses | Amplifying
The study deck behind the 2,430 responses, the 58.4% PostgreSQL pick, the 12 of 20 categories where Claude Code builds custom code and the warning that the results are not market adoption data.
Database startup Supabase raises $150M, acquires Turso – SiliconANGLE
Independent trade reporting confirming the investors, the undisclosed Turso price and Supabase’s MCP server that lets outside agents read and edit records.
Supabase Raises $150M, Buys Turso; Terms Undisclosed – FourWeekMBA
An analysis listing what the company did not disclose and dating three funding rounds in twelve months, with every figure marked as unverified company data.
Supabase’s GTM Playbook: How It Reached 13 Million Developers
A teardown that defines the 13 million as registered accounts, reports the Lovable API and the 46% to 71% Agent Skills test, and relays the Amplifying comparison of plain Postgres and Supabase.
Lovable Vulnerability Explained: How 170+ Apps Were Exposed | Superblocks
A competitor’s account of the 2025 disclosure timeline, the 303 endpoints across 170 projects and the limits of Lovable’s first scanner.
No, Average People Will Not Build Their Own Software With AI
The strongest published opinion against everyone building software, resting on debugging, security and maintenance rather than on measured data.
What Happens When Everyone Can Build Software | Sean Weldon
An AI-assisted summary of the episode that relays the 3 million to 13 million sign-up growth, the taxi-driver example and its own note that durability of such software is unaddressed.
| Citing this article? Brief excerpts are welcome. Please credit Webiano.digital, name the author where stated, and include a link to https://webiano.digital and to this original article. Full or substantial republication requires prior written permission. Read our Copyright and Content Use Policy. |
This article was prepared with the assistance of artificial intelligence tools. The content underwent expert human review, and Webiano Digital & Marketing Agency assumes editorial responsibility for its final version and publication.
