November 21, 2026
Heard in AI

Supabase says its AI bill is about $300,000 a month for 450 staff

Supabase says it spends about $300,000 a month on AI for roughly 450 staff, and each team sets its own AI policy. On Latent Space, it argued the better measure is how many processes now run end to end without a human.

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 3 podcast episodes published October 8–28, 2026

Supabase spends about $300,000 a month on AI across a staff of roughly 450 people, according to the company's own count. No single company-wide rule governs how that money is spent. Ant Wilson, a Supabase co-founder, said each team decides for itself.

The figures came near the end of an episode of the Latent Space podcast, published October 28, 2026. Wilson and his fellow co-founder Paul Copplestone spoke there with the host, swyx. Supabase provides the backend for apps: the database, user sign-in and other server-side machinery that an app needs but its users never see. The closing question noted that the company benefits a great deal from coding agents, the AI systems that write software largely on their own, when outside developers use them. Then it turned that around. What is the policy on AI coding inside Supabase, how much does it spend, and does the spending worry anyone?

Several startups glued together

Wilson said Supabase "has always felt like a bunch of different startups" glued together. A team such as the one behind its auth product, which handles user sign-in, makes its own decisions about tools, how it behaves as a team and even its reporting structure. AI was handled the same way. When people asked what the AI coding policy was, Wilson said, the answer was that it was up to each team. The company has started to consolidate a few things, especially around cost. Wilson said it still wants to let teams decide how they run themselves wherever it can.

A yardstick for the bill

AI providers charge by the token, a small chunk of text that a model reads or writes. A company's AI bill therefore grows with how much its people lean on these tools. The conversation then tried to work out a number that company leaders could take home as a market standard.

The discussion offered rough data points. The average person spends about $20 a month, and the average employee maybe $200. OpenAI employees at a percentile that was not clearly specified were put at perhaps $3,000 to $5,000 a month, with the top users spending into the hundreds of thousands. The measure proposed in the discussion was the ratio of token spending to spending on people. For many companies, it was suggested, the ratio is about 0.1: a tenth of a salary in tokens for every salary paid. By that rough reckoning, a company can probably go up to 1 fairly safely, and above 1.5 it is probably in trouble. The thresholds were offered loosely, as "something like that."

An earlier Latent Space conversation with the AI engineering firm Tenex, published October 21, 2026, used a different scale. That discussion put most companies at perhaps $10 to $200 per employee and more AI-native companies at about 10% of what they spend on people. The average AI startup came in at roughly one to one, and the heaviest users at 5 to 10 times. Tenex estimated its own spending at one to two dollars in tokens per dollar of payroll. The two scales use different thresholds and don't map neatly onto each other.

$300,000 a month

Supabase did not have to dig for its own number. As a remote company it documents nearly everything that happens, so it could simply ask its own AI how much it was spending. The answer was about $300,000 a month. When the conversation guessed a headcount of 600, Wilson corrected it to about 450.

Wilson said that measuring the cost against headcount made sense for Supabase too, because it is hiring very quickly. If the company gets a lot of leverage out of the extra spending, he said, then that is fine. Supabase gave no payroll figure, so the conversation did not place it on the ratio scale discussed earlier.

A few heavy users

The conversation turned to a recent exchange with the chief technology officer of another top startup who was worried about his company's AI spending. The response he got was that the money was really an investment in everyone else, and that usage was probably very skewed: two or three people at the company using AI intensely while everyone else probably was not using it.

The discussion described the same pattern showing up on an internal leaderboard of AI usage. When the leaderboard goes up, people usually end up a little ashamed, and the exchange tied that to wasting money. swyx has drawn his own line on waste. Speaking on The Cognitive Revolution earlier in October, he said of the people he employs: "When you work for me and I'm paying your tokens, you better be like actually producing thoughtful stuff."

Counting processes instead of dollars

The conversation then turned to what a company should measure instead. In Supabase's view, spending is the wrong focus. The better question comes from systems thinking: how many end-to-end processes has the company automated so thoroughly that a human no longer needs to be in the loop, and how many problems that used to take people a long time have shrunk into short ones? According to Supabase, most of its AI investment goes there.

The company's working style helps here. Supabase is asynchronous: people work on their own schedules across time zones instead of gathering for live meetings, and so nearly everything that has ever happened at the company is written down. That makes it easier to build an agent harness for the whole company. A harness is the software around an AI model that gives it the context and tools to carry out tasks. By Supabase's account, building one is mostly a matter of formalizing that archive. In one example from the discussion, you can point AI at the records, ask what the company's principles are, and get back a written account of its decision-making principles across Supabase's entire history.

The job AI hasn't taken yet

Wilson closed with a hiring pitch. Supabase has about 50 open roles, he said. It is fully globally remote and pays the same salary regardless of where people live. Its careers page describes the same remote, asynchronous setup.

Asked to pick the one profile it needs most, Supabase named SREs, or site reliability engineers, the people who keep online services up and running. The reason given was growth at "a ridiculous rate." What stands out in a candidate is a real wish to work remotely. Early on, some hires found the fully asynchronous environment wasn't for them because they needed to see other people. That prompted a joke that the stereotypical SRE, a gray-bearded loner in a dungeon who doesn't want to talk to anyone, would fit right in.

The conversation then turned to AI for site reliability, which had been called on the record one of the hardest categories for agent startups. The assessment was offered provisionally: it might be wrong soon, but so far AI has not cracked SRE. The reply agreed, adding the qualifier that it hasn't happened yet. Building software, the discussion went, is arguably a solved problem. Keeping things running is an entirely different challenge.

Share this article

Go to the original

Sources & further reading

  1. 01

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 1:05:23–1:10:51 (approximate times)

Article history

Updates to this article

Tags

OpenAI's Ari Weinstein on what to hand computer-use agents

OpenAI's Ari Weinstein says computer-use agents can now operate software made for people, citing a two-hour meal order an agent did in 15 minutes. He adds that builders must earn trust through reliability, consent before payments and site limits.

6 min read