One brand running our ingestion framework was paying ~$2,960 a month to get its data into Snowflake: ~$2,500 to Fivetran and a second managed tool, and ~$460 in Snowflake credits to load it. The same sources now run on AI-built connectors in the brand's own cloud account for ~$151 a month, which is $0.95 of compute and ~$150 of Snowflake.
That is ~95% lower, about $34k a year, for the same sources landing in the same warehouse. Until recently the reason to pay a vendor was that building and maintaining connectors was expensive engineering work, and AI has made most of that work cheap.
Why managed ingestion costs so much
Fivetran bills monthly active rows: every row added, changed or deleted in a month, counted separately for each connection. That is a reasonable price for a small source, but two of the most common DTC sources are exactly the wrong shape for it.
Klaviyo sends millions of email events a month, and every one is a new row. In NetSuite, editing an order rewrites its lines, and the line and history tables at a mid-size brand typically hold 20 to 50 million rows, so ordinary order changes keep a large share of them active every month. At this brand the Klaviyo connector alone was ~$1,300 a month on Fivetran.
A lot of teams have already moved to cheaper tools, and the pricing unit barely changes. Airbyte Cloud bills credits per million rows, Hevo bills events per month in tiers, and Estuary bills per gigabyte moved plus a charge per connector. Portable sells flat plans, but in the quotes we have seen they start around $2k a month, run $3-4k once all your sources and destinations are on it, and ~$7k for an enterprise tier. This brand had already moved some sources off Fivetran, and that half was still ~$1,100 a month before warehouse costs.
A job running in your own cloud only gets slower as rows grow, not more expensive. On AWS that is a Lambda function scheduled by EventBridge, and on Google Cloud the same thing is a Cloud Run job scheduled by Cloud Scheduler. Both bill by the second of compute with a free monthly allowance that a handful of daily syncs barely touches, which is how ~$2,500 becomes $0.95.
Why the warehouse bill fell too
Snowflake bills a warehouse for every second it is switched on, with a 60-second minimum every time it wakes up. So the pattern of writes matters as much as their size.
The vendors write often and in small pieces. Fivetran syncs every few hours and merges into each table separately, and streaming tools write continuously, so the warehouse keeps getting woken up and rarely sits idle long enough to suspend. Our framework only sends the records that changed since the last run and loads each table in one step, a few times a day. The warehouse wakes, does one larger piece of work quickly and switches off again, which took the load cost from ~$460 to ~$150.
On BigQuery the same pattern is even cheaper: batch load jobs are free, so loading changed records a few times a day costs nothing beyond storage, where streaming inserts are billed per gigabyte.
We pay for seconds of compute instead of rows of data, and we wake the warehouse a few times a day instead of keeping it awake all day.
New sources in about an hour, including ones the vendors don't support
The other cost of a managed tool is waiting: if a source isn't in its catalog, you don't get the data. With the framework, you paste a link to the source's API documentation and it reads how the API authenticates, its base URL, how it pages through results and how to ask for only what changed. You check what it found and paste the key, and it then finds the endpoints and proposes the tables they land in and the order they load in.
That is how a new ads platform from OpenAI was loading here while its API docs were still thin and before the main ingestion vendors offered a connector for it. Long-tail sources change too: a small survey or reviews tool that would be another connector charge on a managed plan is a short setup and then roughly nothing to run.
Reliability and security
Every table is checked each morning against its source: row counts where the API offers them, plus freshness, coverage, nulls and duplicate keys, with a gap starting a repair run of that table's pipeline. Long jobs stop at a set budget and pick up where they left off on the next run instead of failing. A cheap pipeline that quietly drops rows would cost more than the vendor bill ever did, so the checks are the part we would not skip.
On security, the data never passes through a third party. It goes from the source's API to your own cloud account to your own warehouse, API keys are stored encrypted and are never shown again after they are pasted, and access is managed per person.
What the $1 leaves out
The bill doesn't count engineering time, and that is the real trade. When a source API changes, someone has to fix the connector instead of the vendor doing it. A recent example was a NetSuite reconcile job that started timing out and needed rework. With a vendor, that is their problem and part of what the price pays for.
That is the part we take on. A Meridian engagement starts from a few thousand a month, roughly what this brand was paying for managed ingestion, and the pipelines are one piece of it, alongside owning them when they break, adding sources as you need them, and the reporting, measurement and AI work that sits on top of the data.
When it's worth switching
Your ingestion bill is past ~$1k a month. Below that, the vendor's convenience is usually worth what it costs.
Most of the bill comes from a few high-volume sources. Event streams like Klaviyo and ERPs like NetSuite are where volume pricing hurts most, and they are where a batch job saves the most.
You're waiting on a connector that doesn't exist yet. A new channel or a niche tool is where building from the docs pays off fastest.
You don't need data within minutes. A few loads a day covers most reporting, finance and marketing work. If something genuinely needs streaming, keep that one source on a streaming tool and move the rest.
Someone owns it. Your team or ours, someone has to be on the hook when an API changes.
Happy to walk through what your sources would cost on a setup like this and which ones are worth moving first.
Common questions
Why is Fivetran so expensive?
Fivetran bills by monthly active rows, every row added, changed or deleted in a month, counted separately for each connection. High-volume sources such as email events or ERP order lines generate millions of active rows a month, so the bill grows with your data even when nothing about how you use it changes.
What is the cheapest alternative to Fivetran?
Managed tools like Airbyte Cloud, Hevo, Estuary and Portable are often cheaper than Fivetran, but they still bill for rows, events, data moved or a plan tier that typically runs into the thousands a month. The cheapest option is ingestion that runs in your own AWS or Google Cloud account and is billed by compute time. At one brand that cost $0.95 a month in compute against ~$2,500 for the managed tools it replaced.
Can AI build data connectors?
Yes, for most REST APIs. Given a source's API documentation, AI can work out its authentication, pagination and incremental filters and propose the endpoints and tables to load. A person still checks what it found, supplies the credentials and owns the connector once it runs.
Is building your own ingestion reliable?
It can be more reliable than a managed tool, because you can see and check everything. The setup we run checks every table against its source each morning (row counts, freshness, coverage, nulls and duplicate keys) and starts a repair run when something is missing. The trade is that someone has to own it when a source API changes, which a vendor would otherwise handle.
Is it secure to run ingestion in your own cloud?
It removes a third party from the data path. Data moves from the source's API to your own AWS or Google Cloud account to your own warehouse, so it sits under your existing cloud security and access controls rather than a vendor's. API keys should be stored encrypted, and access should be granted per person.
Does this work on Google Cloud and BigQuery, not just AWS?
Yes. The AWS setup uses Lambda functions scheduled by EventBridge, and the Google Cloud equivalent is Cloud Run jobs scheduled by Cloud Scheduler, both billed by the second with a free monthly allowance. On BigQuery, batch load jobs are free, so loading changed records a few times a day is cheaper still.
Why does Fivetran make my Snowflake bill go up?
Snowflake bills warehouse time per second with a 60-second minimum every time the warehouse wakes. Frequent syncs that merge into each table separately keep waking the warehouse. Loading changed records in a few large batches a day lets it switch off between runs, which cut one brand's load cost from ~$460 to ~$150 a month.
For about what a managed ingestion bill costs, we run ingestion inside your own AWS or Google Cloud account and warehouse, add new sources as you need them, own it when an API changes, and build the reporting and measurement on top.
See how it works →Bills are one brand's actual monthly invoices, rounded, with sources split across Fivetran and Estuary before the switch; engineering time is excluded from both sides. Vendor and cloud pricing is from published pricing as of November 2026, and Portable's ranges are from quotes we have seen rather than a public price list.