Why Everyone Is About To Have Their Own Database
Paul Copplestone, CEO of Supabase, explains how the open-source Postgres platform grew to 13 million users and nearly 50 million databases by becoming the default backend for agents and a new class of non-engineer builders. He covers the shift from dashboard to API to code-first products, personal databases that individuals own, scaling Postgres to billions of databases, and designing tools so agents can recover from errors on their own.
Supabase’s pivot from developer tool to builder platform
Supabase has grown from a YC summer 2020 startup into a platform with 13 million signups and nearly 50 million databases under management, up from about 8 million a year earlier. Revenue is on track to grow 4x this year, and the company now spins up over a million databases a week. Paul Copplestone attributes much of this inflection to agents, which account for a large share of new activity. He frames the shift as moving from “make something people want” to “make something agents want.”
Three eras of the product
Paul describes three chapters. Supabase began as a dashboard company, with a UI that looked like Airtable on top of Postgres, because even experienced software engineers struggled to use raw databases. The second era was the API company, prompted by a customer who needed to launch many databases programmatically; that bet paid off in 2024 when Bolt, Lovable, and similar tools started building on top of Supabase. The third era is code-first. Paul observed that nobody visits dashboards anymore, so he issued a product memo saying anything in the dashboard must work in lockstep with code and infrastructure-as-code. The CLI was re-architected to run inside sandboxes, since agents prefer the CLI and code over clicking through interfaces.
Open source and default choices in LLMs
Everything Supabase builds is open source, and the company also maintains upstream Postgres tooling so it sprawls across several communities. Paul believes agents favor open source because they can read the source when something breaks, and he expects enterprises to demand source inspection for security, which makes closed-source vendors less attractive. He also credits timing: Supabase became a default database recommendation in LLMs at the moment agents began choosing infrastructure.
Builders versus traditional developers
The growth is coming from a new cohort of builders, people who might have domain expertise and product sense but lack traditional engineering training. For years YC advice held that you needed a technical cofounder no matter what. Paul argues the harder skill now is knowing what to build, since building is increasingly available on tap. The conversion and activation metrics are rising alongside signups, which he finds unusual, because typically top-of-funnel acceleration brings declines elsewhere. He notes that the early product-led growth ceiling predicted by investors has not appeared.
Paul cites examples: a fintech founder who built personal balance sheet and banking-integration software from scratch, and a man in Brazil who built a local mini-Uber from his phone after a conference. He sees these builders as people who previously lacked the means to act on their ideas, now discovering new opportunities.
Personal databases and data ownership
Paul’s own project, which he calls a control tower, shows the direction of travel. He started with the belief that markdown and chat with an agent would suffice, but found that a database and a UI become necessary once personal data accumulates. His system covers email triage, finances, health, and CRM, and it can correlate across domains: a weight increase can be read against calendar travel, and transactions can be categorized using context from his knowledge base. He argues that finance and health are where people most want to own their data, and that syncing devices into one database enables correlations no single-purpose app provides.
This leads to a broader thesis. Databases historically served a small number of large institutions, but if everyone with AI access becomes a builder, everyone could hold a personal database containing their life’s data. Paul wants users to own that store and build or swap agents on top of it, rather than having their information locked inside someone else’s application. He notes the business implication: personal data grows over a lifetime, so a consumer-style “prosumer” segment becomes a durable growth vector alongside the early-stage startup business.
Scaling to billions
If each weekly user of ChatGPT, reportedly around 1.2 billion, launched one database a month, the load would reach billions of databases monthly, which Paul acknowledges has never been required of a database system. The answer is a tiered architecture. Multigres scales Postgres up to petabytes, while a recent acquisition, Tersu, provides a Postgres- and SQLite-compatible tier for prototyping and small workloads, offered as Supabase Spark. Paul describes this as the classic tension between enterprise scale and the small-business or consumer end, and says the company is trying to serve both at once, which is difficult with a limited headcount.
He contrasts this with AWS, where a large share of revenue reportedly comes from startups that grew up on the platform. Winning day one, when customers choose infrastructure, is the central strategy, and Paul notes that day one is itself changing as everyone becomes a builder.
AI adoption inside the company
Supabase has no formal internal AI policy. Engineers are given access to whichever tools they want, and the share of engineers writing most of their code with coding agents rose sharply, from roughly 10% to around 60% in six months. Paul links this to models that have become convincing enough that engineers spend time with them and notice what they can achieve. Testing has become the dominant engineering investment, since AI-generated code must be verified, and he sees similar testing-driven patterns across other YC startups.
Designing for agents
Paul’s most concrete advice concerns error messages. Developer errors often read as a bare 500 with little guidance. Supabase rewrote its errors so they tell the agent how to fix the problem or where to read documentation, which lets agents recover without a human searching for answers. He describes winning agents in two stages: first being selected by the model, then ensuring agents don’t abandon the tool when they get stuck. Supabase runs evals across models, publishes an eval site comparing which models build best, and maintains an internal suite that measures how well models use Postgres. Much of engineering effort now goes to testing specific use cases end to end.
In one example, Gary Tan’s personal knowledge system, Gbrain, runs on Supabase. It began as a retrieval system for an agentic newsroom and was distilled into open-source retrieval for OpenClaw. Gbrain combines pgvector-based RAG with a graph layer that structures unstructured data, and its dynamic schemas resemble older knowledge-graph systems. Its evals span around 40 categories covering retrieval quality, cost, and latency, and an agent loop automatically ships improvements that pass them. Failed experiments are written up as markdown explaining why they failed, which feeds future iterations. Gary sees this as a primitive form of recursive self-improvement for a specific use case.
The future of coding
Paul builds his own products by talking to an agent, mostly without reading code and without steering it toward GitHub or other conventions, specifically to dogfood how a non-engineer would experience Supabase a year from now. He recounts building a developer site for a domain he owned, where a few conversational instructions produced a well-structured site with SEO handled, all from his phone at bedtime. He notes that models eventually absorb such patterns into their weights, which removes the need for prompting, but that until then small amounts of guidance help.
The group discusses a tension: capabilities are accelerating while the number of people who know what to build and how to ask stays roughly constant. Paul is optimistic and sees the platform enabling far more entrepreneurship and a fulfilling lifestyle for many people, which he sums up as expecting 2027 to be a good year for that.
Startup economics and advice
The discussion turns to vertical SaaS founders who are struggling to raise Series A rounds even at $5 to $10 million ARR, while the same companies can reach $10 to $15 million with about 10 to 15 people and reach profitability quickly. Paul’s office-hours advice to founders who raised little is that they may later see it as a blessing, since they may not need the capital that once required hundreds of millions to reach $30 million in revenue.
His final advice rests on durability. Databases have existed for 50 years and will likely persist through the AI era, so choosing a durable category is key when betting a career on it. Beyond that, he emphasizes staying adaptable, watching the newest ways of building, and keeping team capacity free for fast shifts. Dogfooding is central because getting stuck on your own tooling gives the best feedback. Partnering with ecosystem players also helps, though many bets fail, and the ones that work create the most value.
** Tersu is a SQLite-based database company acquired by Supabase in October 2026. It specializes in provisioning a high volume of isolated databases optimized for AI agent workloads, and its architecture supports on-demand databases at low cost. 2. Multigres: