The cost of indie cloud: my real bill across 7 apps

“What does it actually cost to run all of that?” is the question I get asked most often when someone hears I run seven apps on the side. The answers floating around online are not very useful. Some are aspirational and assume you’ll never leave the AWS free tier. Others are wrong-scale and describe how a Series-A company runs infrastructure. Neither matches the reality of an indie portfolio.

This is what my actual cloud bill looks like in May 2026, broken down by vendor across all seven apps, including the three incidents in the past month that taught me more about cloud cost than any blog post could.

The portfolio

Seven apps share roughly the same stack. SnowPipe is a Shopify product feed orchestration platform. SnowGen does AI content. SnowScrape is a managed web scraping platform. SnowGlobe is a lead intelligence dashboard. SnowFort is a Fortnite shop tracker that runs as a Discord bot. SnowSports is a sports-data analytics tool. TrueIce publishes auto-updated League of Legends patch data.

Some are revenue-bearing. Most are not yet. All of them run on roughly the same infrastructure: Vercel for hosting, Neon for Postgres, AWS for the heavy lifting on a couple of apps, Upstash for queue and rate limit, plus Doppler, Clerk, Stripe, Sentry, and a handful of smaller services. That sameness is intentional. A one-person studio cannot afford a different stack per app.

Vercel

Vercel hosts every app’s frontend and most of its serverless functions. I’m on the Pro plan. For a single solo founder, the most expensive line item on Vercel is not bandwidth or builds, it’s function execution time when something goes wrong. A polling loop that should fire every five minutes but accidentally fires every five seconds will eat your monthly compute allowance in an afternoon. I’ve come close to that exactly once, on a scheduler bug that started double-firing before I caught it. Vercel’s metered usage emails are the early warning system.

Across seven projects on a single team plan, Vercel is currently the largest fixed cost. That’s fine. The alternative is operating my own load balancer and origin servers, and the math does not work for a one-person studio.

Neon

Postgres for almost every app, on Neon’s serverless platform. Each app has its own project under the same org. The pitch with serverless Postgres is auto-suspend: an idle endpoint stops billing compute time after a configurable timeout, typically five minutes.

That’s the pitch. The reality is that the auto-suspend setting is per-endpoint, and I had nine endpoints across the org with auto-suspend disabled entirely. Every one of those was billing compute hours twenty-four hours a day, even when nobody was hitting the database. The bill ramped to about $80 per month before I noticed, and tracing it back showed Neon was the dominant driver of the entire indie cloud bill that month.

The fix took less than ten minutes: set the suspend timeout to 300 seconds on every endpoint. The bill dropped within the first weekly billing cycle. The lesson, beyond the obvious “check your defaults,” is that serverless platforms do not always default to cost-aware settings. Some of those endpoints had auto-suspend turned off because at one point I was running a long-lived migration and didn’t want the connection to drop. Then I forgot.

AWS

I run two AWS accounts. The current one is where SnowPipe’s Lambdas, SES email sending, and SQS queues live. Its monthly bill is essentially zero at indie scale because Lambda compute is metered to actual milliseconds and most workloads are sporadic.

The other account is older. It started life as a side project called etriever and outlived the project. For most of last year I forgot it existed. Then in early May the monthly bill came in around $66, almost all of it from CloudWatch logs.

The cause was a poison-pill loop. An SQS queue had a Lambda source mapping with no functioning consumer, three EventBridge crons were firing into the queue every few minutes, and every failed invocation was logging to CloudWatch, which charges per ingested gigabyte. The actual workload accomplished nothing. The side effect was about $58 per month of CloudWatch ingestion. I disabled the source mapping and the crons, and the burn stopped immediately. Full deletion of the account is pending, but the cost has already gone to near zero.

The lesson is that old AWS accounts do not quietly deactivate themselves. They sit there spending your money until you log in and notice.

Upstash

Upstash provides Redis for queue work (BullMQ) and rate limiting across several apps. It bills by request and by storage, so an idle queue is essentially free.

A few weeks ago I rotated a token that sat in front of an Upstash endpoint. Eleven secret-manager configs needed the new token. Ten of them updated cleanly. The eleventh was an orphaned BullMQ worker that had been disconnected from anything actually running for months but was still polling Redis with the old token, generating about $20 per month of zombie traffic. After the rotation, the worker could not authenticate anymore, and the cost dropped to zero almost overnight.

A zombie worker is, mechanically, just a process nobody is still observing. It usually shows up as a queue with traffic but no consumer, a Redis instance with steady CPU but no human-driven changes, or in this case an account with $20 per month of activity that you cannot trace back to any current app.

The smaller stuff

Doppler costs nothing on the individual plan up to a generous secrets cap, and I’m not over it. Clerk’s free tier covers up to ten thousand monthly active users, and across seven apps I am not close to that aggregate yet. Stripe is purely transactional, so my fixed cost there is zero, with the rake taken from each charge. Sentry’s free tier covers five thousand errors a month, more than enough at indie traffic. Cloudflare for DNS, GitHub on the personal plan, and a handful of other services round out the stack at zero or near-zero cost.

The total

At the peak last month, before any of the incidents above were resolved, the all-in monthly bill across all seven apps was around $240. Most of that was the Neon $80, the etriever $58, the Upstash $20, plus Vercel’s fixed cost and modest Lambda usage.

After the three incidents were fixed, the steady-state bill is somewhere between $60 and $80 per month across seven apps. That’s roughly $10 per app per month. Not free, not expensive, but a real number that I think is roughly representative of what a solo founder with a serious indie portfolio actually pays.

Three lessons

Defaults on serverless platforms are not always cost-aware. Auto-suspend is the easy example, but reserved capacity, log retention, and instance sizing all default in directions that bias toward developer experience over your wallet. Read the dashboard, not the marketing.

Old accounts are not free. AWS will not call you when an account you forgot about starts burning $58 per month. You have to look. I now do a quarterly review of every paid account I have credentials to, just opening the billing dashboard and confirming the number is what I expect.

Zombies cost real money.A worker still polling, a cron still firing, a source mapping still active, all of these will keep generating bills long after the project they served has been shut down. The fix is not better tooling. It’s writing down what you spun up so you can spin it back down later.

Closing

None of this is exotic. The reason I’m publishing the numbers is that the genre of “indie cloud cost” articles is dominated by either AWS-free-tier optimism or scaled-up-startup post-mortems, and the middle ground that solo founders actually live in is undercovered. If you’re running a small portfolio and your bill looks similar to mine, you’re probably doing it right. If it’s much higher, look for your zombies. If it’s much lower, you’re probably one project away from one of these incidents anyway.