Supabase, a company that gives app builders a ready-made backend built around a Postgres database, has announced a product meant to fix a limit its own users kept running into. Supabase Compute lets developers run long jobs, services and AI agents inside the same private network as their project's database. The company announced it on October 27, 2026 as a private alpha.
Supabase co-founders Paul Copplestone and Ant Wilson discussed the launch with swyx, host of the Latent Space podcast, in an episode published October 28. Compute was the first of the company's headlines they covered.
The limit Compute answers
Until now, the way to run custom code on Supabase was Edge Functions: short, serverless functions built on Deno, a JavaScript and TypeScript runtime. "Serverless" means the developer writes the code and the platform decides where and when it runs, without anyone managing a machine. On the podcast, Edge Functions were described as having "basically a two-second limit," and Supabase's community had long been asking to run much longer workloads. The examples given were chunking and embedding: splitting documents into pieces and turning each piece into numbers an AI system can search by meaning.
The two seconds needs a qualification. Supabase's Edge Functions documentation says the limit is two seconds of CPU time per request, meaning time spent actually computing, and it excludes time spent waiting on the network or other asynchronous work. Other limits apply separately. A worker can live for 150 seconds on the free plan and 400 seconds on paid plans, and a request that sits idle times out after 150 seconds. Code that mostly waits for other services therefore hits different walls than code that crunches numbers, though in both cases the ceiling is a few minutes at most.
What Compute is
On the podcast, Compute was described as able to run almost anything. It is long-running and serverless, and it sits in the same private network as the database. That closeness means low latency, so code can open Postgres connections directly. Compute is not tied to Deno: developers can use Rust or Python, or run agents inside it. When the conversation turned to comparable offerings in other clouds, the answer was something like Amazon's Lambda functions, specifically the version that runs a full container. A container is a packaged program together with everything it needs to run.
The alpha announcement is narrower than that description. It pitches Compute for agents, APIs, background processors, embedding backfills and WebSocket services. Each service gets a Linux environment and deploys from Node, Deno or a Dockerfile, and its configuration lives alongside the app's other Supabase files. For now access is limited to evaluation workloads in a single region, us-west-2. Services may restart at any time, and there is no persistent disk, no autoscaling and no GPU. Every service receives all of the project's secrets. Public services suspend when idle, while private ones keep running. Agent sandboxes, isolated spaces where an agent can run code more safely, are described as coming after this first alpha.
The road to durable workflows
Later the conversation returned to durable workflows, the style of programming associated with Temporal, and to the question of why Supabase had not made them a higher priority while Temporal is doing well.
In the exchange that followed, Supabase was described as having moved toward that goal step by step, with compute identified as the missing piece. Members of the community had tried to build workflow engines on top of Supabase functions, and the discussion said those attempts failed because of the two-second timeout. Compute, in that account, is the primitive that could unlock long-running durable workflows.
A long-running process is only part of what durable workflows require. Temporal's own documentation describes a service that records an ordered history of every step a workflow takes. If the worker running a workflow crashes, another one replays the code against that history to reconstruct its progress. Calls to outside systems get retry policies and timeouts, and workflows can wait on durable timers. An onboarding process that waits days for approval therefore does not depend on one program staying alive the whole time. Compute, as announced, supplies containers that can run for a long time, and its alpha services can restart at any moment with no persistent disk. The history, replay and retry machinery would still have to be built on top, which fits the description of Compute as a primitive.
Giving the AI what it wants
The discussion also placed the launch inside a bigger pitch: Supabase as an all-in-one "agentic cloud," an integrated backend that people building with AI can use. Sign in with ChatGPT was cited as part of that, with the argument that a platform has to give AI systems everything they need to work with it.
Sign in with ChatGPT launched in beta on July 29, 2026. It lets people create a Supabase account with their ChatGPT identity, and Supabase says it was an OpenAI launch partner. Logging in is kept separate from letting ChatGPT act on a user's account: users see a consent screen for the integration and can revoke it, and new users set up an organization before creating projects through ChatGPT.
The conversation then turned to partnerships with AI labs. A strategy of partnering with as many frontier labs and models as possible was described as a top priority, alongside rising attach rates. The discussion also asked whether labs outside the United States, such as Mistral, Cohere and the Chinese labs, had significant partnerships; the indication was not many so far, though they might come. Wilson explained why the focus is on American labs: "that's who's delivering the most traffic at the moment, right? That could change."