What a $200-a-month data stack looks like

The quote lands in every founder’s inbox eventually: Fivetran for ingestion, Snowflake for the warehouse, Looker on top — call it $2,000 to $4,000 a month before anyone has answered a single question. For a company with under a few hundred gigabytes of data, which is nearly every company before Series B, that quote is pricing a problem you don’t have. Here is the stack we deploy instead, itemized, with the real monthly numbers.

The bill of materials

Every line below is something we run in production. The prices are steady-state monthly figures, not free-tier gymnastics that collapse the month you get traction.

• Ingestion — $0–75. dlt is an open-source Python ingestion library; jobs pulling your product database, Stripe, and the ad platforms run on the same small VM as everything else, so the marginal cost is zero. If you’d rather buy connectors than own scripts, an Airbyte Cloud starter plan covers seed-stage volumes for roughly $10–75.
• Storage — $2–10. A few hundred gigabytes of Parquet in S3 or Cloudflare R2 costs single-digit dollars. This is the line everyone expects to be big; it is the smallest number on the page.
• Compute — $0–100. DuckDB is free and scans this much data in seconds on a single node. MotherDuck’s small tier adds a shared cloud endpoint your team and your BI tool can both hit, for tens of dollars.
• Orchestration — $20–40. One small VM runs Dagster: schedules, retries, and a UI that answers “did last night’s run succeed.” A serverless cron works too, but you’ll miss the run history the first time something breaks at 3am.
• BI — $0–30. Metabase open source rides on the same VM for nothing extra; Evidence is the code-first alternative if your reports belong in git.
• Monitoring — $0. The orchestrator already knows when a run fails or data goes stale; a Slack webhook is the pager. No observability SaaS required at this scale.

Add it up and the ledger reads roughly $100 a month at the frugal end, $300 with the managed options — against $2,400-plus for the enterprise pattern, an 8–20x difference on infrastructure that recurs every month, forever. The outcome is the same shape: your sources land in one queryable place, models run on a schedule, charts sit on top, and a failed run pings a channel instead of dying silently.

What the expensive stack actually buys

Being fair here matters, because the $2k stack is not a scam — it is mispriced for your stage. It buys three real things:

• Connector breadth and maintenance. Fivetran maintains hundreds of connectors, and when NetSuite changes its API at 2am, that’s their pager. With dlt, an upstream API change is your engineer’s Tuesday morning.
• Governance. Snowflake’s role hierarchies, row-level policies, and audit surface matter enormously with forty data consumers and a compliance program. With four consumers who share a standup, they’re configuration you pay for and never touch.
• Seats and self-serve. Looker’s semantic layer earns its price when dozens of non-analysts need to explore data unaided. Metabase covers a ten-person company fine.

The crossover points are estimable, not mystical. When the engineering hours spent patching connectors cost more than the subscription would — at a loaded $100-plus per hour, half a day a week pays for a lot of Fivetran — buy the connectors. When you pass roughly a terabyte of actively queried data, need cross-team isolation, or an auditor starts asking who can see which rows, the warehouse tier begins justifying itself. Before those points, the premium buys standby capacity you don’t use.

The line item that isn’t on any invoice

Now the honest part. The $200 stack does not run itself. Sources drift, a connector update breaks a sync, a backfill needs to run without double-counting, and someone has to read the morning alert and own the fix. Call it two to four hours a week in steady state, more in a bad week — and that operator is the real cost line, whichever stack you buy. You can pay it three ways: a product engineer’s attention (the most expensive hours you have), a dedicated data hire (a $180k-plus decision you probably shouldn’t make yet), or an outside operator. And one rung further down the honesty ladder: if you’re pre-product-market-fit and every question you have is answerable from one Postgres database, skip all of this — a read replica and a lightweight BI tool are the correct stack, and this post will keep until the day two dashboards disagree about revenue.

We build the stack above in two to three weeks and then run it on retainer — the monitoring, the break-fix, the backfills, the new models as your questions change. The $200 is the infrastructure. The operator is the product.