When an AI-Built App Outgrows the Free Tier
An AI-built app outgrows the free tier in a predictable order. Which limit bites first, what each failure looks like, and when upgrading is worth it.
Free tiers do not fail gracefully. An AI-built app outgrows the free tier at a specific limit, in a fairly predictable order, and each limit produces a different symptom. Knowing which one you are about to hit is the difference between a planned twenty-five dollar upgrade and a Saturday morning with a dead app and no idea why.
The order below is what actually happens, which is not the order the pricing page implies.
The order an AI-built app outgrows the free tier limits
1. Project pausing, long before any usage limit. This catches side projects hardest and surprises people most, because it has nothing to do with how big you got. Per the Supabase pricing page, free projects are paused after one week of inactivity. A demo you showed a client a fortnight ago is not slow, it is off. The app returns connection errors and the dashboard says paused.
2. Egress, well before storage. Free egress is 5 GB plus 5 GB cached. That sounds generous until you serve images. A 2 MB product photo delivered to 2,000 visitors is 4 GB on its own. Apps that store user uploads and serve them full-size burn through egress while the database is still nearly empty.
3. Database size, if you log. 500 MB of database is a lot of rows and very little logging. AI-generated backends frequently include an events or audit table that writes a row per request, with a JSON payload, and nothing that prunes it. That table reaches hundreds of megabytes while your actual business data sits under 20 MB.
4. Edge function invocations, only at real volume. 500,000 a month is roughly 16,000 a day. Most apps never approach it. The ones that do are usually calling a function on a timer or polling from the client every few seconds, which is a bug rather than growth.
5. Monthly active users, almost never. 50,000 on the free tier. If you are hitting this, the other four already fell over and the upgrade decision made itself.
Limit | Free tier | What the user sees when you hit it |
|---|---|---|
Inactivity | Paused after 1 week | App completely unreachable |
Egress | 5 GB plus 5 GB cached | Images stop loading, requests fail |
Database | 500 MB | Writes fail, app appears broken on save |
Edge functions | 500,000 per month | Specific features stop responding |
Monthly active users | 50,000 | Sign-ins fail |
Projects | 2 active | Cannot create the next one |
Find out where you actually are
Read your usage dashboard before deciding anything, and read the trend rather than the number. A database at 180 MB growing 40 MB a week is a four-week problem. The same 180 MB flat for a month is not a problem at all.
For the database specifically, find out what is actually taking the space, because it is usually one table:
select relname as table,
pg_size_pretty(pg_total_relation_size(c.oid)) as total
from pg_class c
join pg_namespace n on n.oid = c.relnamespace
where n.nspname = 'public' and c.relkind = 'r'
order by pg_total_relation_size(c.oid) desc
limit 10;If a logging or events table is at the top, you have a retention problem, not a capacity problem, and deleting old rows with a scheduled job is both cheaper and more correct than upgrading.
Fix first, upgrade second
Three cases where paying does not solve the underlying thing:
Egress from unoptimised images. Serving originals when you need thumbnails. Resize on upload and serve appropriately sized variants, and egress typically falls by most of an order of magnitude. Paying for 250 GB to deliver images that should have been 150 KB is buying your way around a missing resize step.
A logging table with no retention. Add a scheduled delete. Thirty days of application logs is almost always enough, and most of what gets logged was never going to be read.
Polling instead of subscribing. Client code that fetches every three seconds generates invocations in proportion to how long people leave the tab open. A subscription or a sensible interval cuts it to a fraction.
Each of those is a one-afternoon fix that also makes the paid tier last longer when you do move.
When upgrading is simply correct
Three situations where you should stop optimising and pay.
Anyone depends on it being up. The moment a paying customer or a client relies on the app, the inactivity pause alone justifies the cost. Reliability is the product at that point. The Pro plan starts at 25 a month and brings 8 GB of disk, 250 GB of egress, and 2 million function invocations, which is a different order of headroom.
You are rate-limited on diagnosis. Free tiers restrict logging retention and observability. When you are debugging production and cannot see yesterday, the upgrade pays for itself in one incident.
Growth is steady rather than spiky. A spike is a scaling event. Steady week-on-week growth means you will be having this conversation again in a month, and doing it once is cheaper.
The planning version of this, before the bill exists, is forecasting your AI bill for next month, and the revenue side sits in adding usage-based billing to an AI-built app and whether an AI app needs a free tier of its own.
Do these before you need them
Two things that cost nothing now and matter enormously later: turn on backups, and know how to restore them. A free tier hitting its database limit mid-write is exactly when people discover their backup situation. Adding backups to an AI-built app covers it, and the traffic failures that tend to arrive in the same week are in what breaks first when an AI-built app gets real traffic.
FAQ
Why did my free project stop working when nobody was using it?
Because nobody was using it. Free projects pause after a week of inactivity, which looks identical to an outage from the outside. Opening the dashboard and resuming brings it back.
Should I upgrade or optimise first?
Optimise if the cause is a fixable inefficiency such as full-size images or an unpruned log table, because those keep costing you on the paid tier too. Upgrade without hesitation once someone is depending on the app being available.
How do I know which limit I am closest to?
Your provider's usage dashboard shows each one. Look at the rate of change rather than the absolute figure, since a limit you reach in three months is a different decision from one you reach on Friday.
Does upgrading fix a database that is full of logs?
It postpones it. A table growing 40 MB a week fills 8 GB eventually too, just later and with a monthly charge attached. Add retention either way.
How did this land?
About the author

Staff Engineer, Platform
Carlo works on the platform that turns prompts into running apps. He writes the engineering deep dives and the changelog notes worth reading.


