Paul Copplestone on Supabase: Building for Agents and a Billion Builders

Supabase CEO Paul Copplestone explains why AI agents now drive most new database activity, and what happens when everyone with access to AI can build their own software. Speaking on Y Combinator’s Lightcone podcast, Copplestone says Supabase has grown from about 3 million to 13 million sign-ups in a year. Databases under management went from 8 million to nearly 50 million, more than a million new databases are created each week, and revenue is on track to grow 4x. The hosts say roughly 70% of new activity now comes from agents. The 48-minute conversation covers Supabase’s shift to code-first tooling, error messages written for agents, personal databases, Garry Tan’s GBrain, and why Copplestone is “pretty pumped about 2027.”

TLDW

Supabase has moved through three eras: a dashboard company, then an API company, and now a code-first company built for the way agents work. Growth comes from engineers building faster and from non-engineer “builders” who no longer need to persuade a developer to make their ideas. Copplestone describes a future of personal databases holding people’s health, finances and relationships. To get there, Supabase is building two ends at once: Multigres to scale up to petabytes and a Turso-based tier to scale down to tiny projects. He also explains how Supabase wins agents through self-repairing error messages and evals. The hosts describe models whose capabilities are “going parabolic” while the number of people who know what to build stays flat.

Thoughts

Copplestone’s two-stage model of winning agents is the most practical idea in the episode for anyone who sells developer tools. Stage one is being picked by the model, which depends on training data, defaults and timing. Stage two is not losing the agent when it gets stuck, and it always gets stuck eventually. A human who hits a vague 500 error searches the web and keeps going. An agent that hits one may switch to a different database. So Supabase writes error messages that tell the agent exactly how to fix the problem or which doc to read. That turns a moment where it could lose a customer into a self-repair loop. The wider shift follows from this: Copplestone says much of Supabase’s engineering effort now goes into testing specific use cases with evals rather than building new features. The product is increasingly whatever the agent experiences, and that has to be measured, not assumed.

The “unstoppable force meets immovable object” framing near the end deserves more attention than it gets in the conversation. Model capabilities are rising steeply. The number of people who know what to build, and how to ask for it, stays roughly flat. That is also the honest limit on Supabase’s own story. A billion potential builders is the top of the funnel, but Copplestone’s best examples share a trait. The Brazilian taxi driver who built a local ride app from his phone and the fintech founder writing his own bank integrations both knew their domain deeply. Earlier in the episode a host says the harder thing is now “to know what to build,” and that point decides who benefits. Platforms that help people find and frame problems, not just ship code, will capture more of this wave than platforms that only make building easier.

Copplestone’s dogfooding rule is a smart piece of product discipline. When he builds his own projects, he talks to an agent by voice, refuses to look at the code, and avoids nudging it toward GitHub or anything an engineer would say. He is trying to “code like a normie,” because the normie is who will arrive at Supabase in a year. Most founders test their product as the expert they already are. He is testing it as a customer who doesn’t exist yet. Pair that with his observation that “no one visits dashboards anymore,” which drove a memo requiring every dashboard feature to work in code, and you get a company steering by where its users are heading rather than where they have been.

The Lindy argument is the quiet center of the strategy. Databases have existed for 50 years, and Copplestone expects them to last 50 more no matter what AI does. Garry Tan supplies the consumer version of the same idea when he describes pulling his CRM out of one agent into his own database, so he can attach any agent later. Agents are likely to change hands every few months, while the store of record lasts. The risk is the one Copplestone admits: serving both ends at once, with Multigres for petabytes and Turso for tiny personal apps, is “incredibly difficult.” His own AWS figure, that about 40% of AWS revenue comes from startups that scaled up on it, is one he asked listeners to fact-check. Treat the business model as plausible but unproven.

The aside about vertical SaaS founders who can’t raise a Series A is a useful corrective. A company with 15 people at $2 million in revenue per head and 50% margins does not need venture money, and the hosts tell such founders they may later see the rejection as a blessing. Copplestone admits he envies them, since not fundraising now looks like the real achievement. For the builders Supabase is courting, a small, profitable company may be the more common good outcome, and that is a better future than the funding headlines suggest.

Key Takeaways

Supabase grew from about 3 million to 13 million sign-ups in a year, and from 8 million to nearly 50 million databases under management. It now creates more than a million databases a week, and Copplestone expects revenue to grow 4x this year.
Sign-ups, activation and conversion are all rising at once, which Copplestone calls unusual. Fast top-of-funnel growth normally drags down the other numbers.
The company has had three eras: a dashboard that looked like Airtable on Postgres, an API that suited platforms like Bolt and Lovable, and now code-first, git-native tooling built for agents.
Being a default choice inside LLMs, and being open source so agents can read the code, were critical to Supabase’s growth. Copplestone also expects enterprises to demand source access for security reviews.
“Builders” are a growing segment alongside engineers: domain experts with ideas they could never get an engineer to work on. Copplestone says he can’t yet see a ceiling on product-led growth.
Finance and health are the most common personal apps, because they are where people most want to own their own data.
Copplestone runs simple arithmetic: if ChatGPT’s 1.2 billion weekly users each launched one app a month, that would mean 1.2 billion databases a month. Supabase is redesigning its architecture for that scale.
Supabase acquired Turso, whose engine is compatible with Postgres and SQLite, to power Supabase Spark, a lightweight tier for new projects. Multigres covers the high end, under the tagline “prototype to petabytes.”
Supabase has no internal AI policy. Engineers can use any tool they like, and in six months the share writing most of their code with agents rose from about 10% to 60%.
Error messages now tell agents how to fix the problem. Winning agents takes two stages: getting picked by the model, then keeping the agent from switching away when it gets stuck.
Garry Tan’s GBrain runs on Supabase. It combines Postgres, pgvector and graph-style retrieval to give his agents access to half a million markdown files, and agents continuously try to improve its scores on 40 categories of evals.
The goal is the Stripe and AWS business model: be there on day one, give customers a great experience, and grow with them, because moving a database to another provider is near impossible.
Copplestone’s lessons for founders: pick a durable category, adapt fastest rather than claiming to have every answer, dogfood constantly, and keep spare team capacity for partners’ experiments, since most of those bets fail.

Chapters

01:18 Growth, 4x Revenue and Three Eras of Supabase

Copplestone lays out the numbers: 13 million sign-ups, almost 50 million databases under management and more than a million new databases a week. He credits engineers building at a much higher rate, plus non-engineers coming online. Supabase started as a dashboard company, became an API company, and is now code-first because agents prefer CLIs. Being chosen as a default database by LLMs at the right moment helped hugely, and so did open source, which lets agents inspect the code when something breaks.

04:34 Builders vs. Traditional Developers

Unusually, sign-ups, activation and conversion are all climbing together. The new “builder” may not be an engineer at all, just someone with an idea who opened an AI coding tool and found that plain files only go so far. The hosts describe a fintech founder who exited for more than a billion dollars and now writes his own bank integrations. Copplestone’s favorite example is a Brazilian taxi driver of about 50 who built a mini Uber for his local market from his phone.

12:12 Personal Databases and Data Ownership

Garry Tan describes a personal “control tower” covering email triage, finances, health and workouts. He says its finance section matches Mint or Monarch and is better personalized, because it can look up who a payee is in his own data. He is moving his CRM out of a single agent into his own database so any agent can use it. The hosts note how strange this is historically: the world once had a few very large databases, and now individuals want their own.

15:56 Scaling to Billions of Databases

Copplestone’s projection from ChatGPT’s 1.2 billion weekly users points to databases launched by the billion. Since compute is scarce and most of these projects don’t need “such a Ferrari,” Supabase acquired Turso to build Supabase Spark, a light tier for new projects. Multigres covers the petabyte end. Doing both while hiring fast is “incredibly difficult,” he says.

17:58 AI Adoption Inside Supabase

Supabase has no AI policy. Engineers use whatever tools they want, and the share writing about 80% of their code with agents went from 10% to 60% in six months. Copplestone says engineers needed convincing, and better models plus visible results did it. As more code is written by AI, the company has invested heavily in testing tools. The hosts say YC startups in CI and build infrastructure are booming for the same reason.

19:58 Lessons From Each Era

The dashboard came from seeing that even engineers found databases hard. The API came from a 2021 partner request that paid off in 2024, when Bolt and Lovable built on top of it. The code-first era began when Copplestone noticed nobody was visiting the dashboard. He credits Parse and Firebase with showing that a database company could exist outside the hyperscalers. The model he wants is AWS and Stripe: be there on day one and grow with customers.

27:29 Building for Personal Agents and Dogfooding

A week without his executive assistant made Copplestone “very bullish” on personal agents. He now builds by voice, sitting on a call with an agent, and never looks at the code unless he is checking the architecture. One host, back from Supabase Select, imagines sending his agent to the conference to learn the new features for him. That prompts a joke about a conference for agents.

30:01 Agent-First Design, Evals and GBrain

Supabase’s error messages now explain how to fix each problem, so agents can repair themselves instead of switching providers. The company publishes evals on which models build best with Supabase and Postgres, and much of its engineering effort now goes into testing use cases. Garry Tan explains that GBrain grew out of an agentic newsroom he built. It turns half a million markdown files into a retrieval and graph layer on Postgres, and agents loop on 40 eval categories to keep improving it.

38:42 The Future of Coding

Will builders need to understand evals? Probably not, Tan says, since people are still inventing the basic tools and techniques. The hosts recall Claude Code creator Boris Cherny’s view that you will simply talk to the model like a friend. Copplestone describes rebuilding the information architecture of a developer site by voice from bed, with the agent finishing in 20 seconds. He deliberately prompts like a non-engineer to see what Supabase will feel like for next year’s users.

42:36 More Entrepreneurship and Lessons for Founders

Copplestone, whose earlier startups served blue-collar workers, says he sees a fulfilling entrepreneurial life opening up for many more people. The hosts note that vertical SaaS companies with $5 million to $10 million in revenue struggle to raise a Series A, yet can now reach $25 million or more with 15 people. His advice is to pick durable categories (the Lindy effect), adapt faster than anyone, dogfood constantly, and keep spare capacity for partners’ experiments.

Notable Quotes

“This category of people have all these spinned up ideas that they could never convince a software engineer to like help them build and now they don’t have to convince anyone.”
Paul Copplestone, on the rise of non-engineer builders

“The much harder thing to do is to know what to build. You know, being able to build is something that anyone can sort of have on tap now.”
A Lightcone host, on what changed at YC in the past year

“No one visits dashboards anymore.”
Paul Copplestone, on why Supabase went code-first

“We just want to be there on day one, make sure that they have a great time, and then make sure that they actually don’t want to leave.”
Paul Copplestone, on the AWS and Stripe business model

“There’s like two stages to winning the agents. The first is being picked by the model, and then the second is like if they get stuck, you don’t want them to switch away.”
Paul Copplestone, on designing for agents

“The capabilities of the models are like going parabolic and then the number of people who know how to actually have an impact on society with it is like more or less constant.”
A Lightcone host, on the gap between AI capability and builders

“Databases have been around for 50 years. I think they’ll be around for another 50 years.”
Paul Copplestone, on the Lindy effect and choosing a durable category

Watch the full conversation here.