Data engineering services in 2026: what a pipeline and warehouse actually cost to run

Summary. A data platform is quoted as a build and paid for as a subscription. As of 10 August 2026, Snowflake's on-demand list price is $2.00 per credit on Standard, $3.00 on Enterprise and $4.00 on Business Critical, identical in AWS US East (N. Virginia) and AWS Asia Pacific (Mumbai), with storage at $23.00 per TB per month. Google BigQuery charges $6.25 per TiB scanned on demand in us-central1, or $0.04 to $0.10 per slot-hour depending on edition. AWS Redshift Serverless is $0.375 per RPU-hour with a 4-RPU floor, so the cheapest hour you can buy is $1.50. dbt Labs' 2026 State of Analytics Engineering report, published 14 April 2026 from 363 practitioners, found 57% reporting increased warehouse and compute spend against 36% reporting increased team budgets. That gap is the whole problem. eCorpIT builds data platforms where the run cost is designed in, not discovered in month four.

What we build

Four things, in roughly this order of frequency.

Ingestion into a warehouse. Source systems to a landing layer, with change-data-capture where the source supports it and scheduled extraction where it does not. Managed connectors where the connector exists and the volume justifies the licence; custom extractors where it does not.

Transformation layers with tests. Staging, intermediate and mart models with data contracts and assertions that fail the build rather than the dashboard. This is where most of the engineering hours go and where most of the value sits.

Orchestration and observability. Scheduled and event-driven DAGs, retries with backoff, freshness SLAs, and alerting that names the upstream owner rather than paging the data team for a vendor outage.

Cost governance. Warehouse sizing, auto-suspend, query tagging by team, and a monthly attribution report. Optional on most engagements. It should not be.

What it costs to run: the August 2026 list prices

The single most useful thing a buyer can do before signing a build contract is model twelve months of run cost. Here are the published numbers we model against.

Platform	Unit	List price (Aug 2026)	Region cited
Snowflake Standard	credit	$2.00	AWS us-east-1 and AWS ap-south-1 (Mumbai)
Snowflake Enterprise	credit	$3.00	AWS us-east-1 and AWS ap-south-1
Snowflake Business Critical	credit	$4.00	AWS us-east-1 and AWS ap-south-1
Snowflake storage (on demand)	TB/month	$23.00	AWS us-east-1 and AWS ap-south-1
BigQuery on-demand analysis	TiB scanned	$6.25, first 1 TiB/month free	us-central1 (Iowa)
BigQuery Enterprise edition	slot-hour	$0.06 pay-as-you-go	us-central1
Redshift Serverless	RPU-hour	$0.375, 4-RPU minimum	us-east-1
Redshift ra3.4xlarge	node-hour	$3.26	us-east-1
Databricks Jobs Compute (Premium)	DBU	$0.30	Azure US East
Databricks Serverless SQL (Premium)	DBU	$0.70	Azure US East
dbt Starter	user/month	$100, 15,000 models/month	n/a

Three observations a rate card will not give you.

Snowflake's credit price does not rise in India. Mumbai and Central India (Pune) are both at the $2.00/$3.00/$4.00/$6.00 tier, the same as N. Virginia, while São Paulo runs 55% higher at $3.10 on Standard. Data residency in India is not a pricing penalty on Snowflake. Storage is also $23.00 per TB per month in Mumbai, matching us-east-1.

BigQuery's on-demand model punishes wide scans, not big tables. At $6.25 per TiB, a single unpartitioned SELECT * over a 4 TiB fact table costs $25 every time an analyst runs it. Partition pruning is not an optimisation here; it is the difference between a $200 and a $20,000 month.

Redshift Serverless has a floor. The 4-RPU minimum means idle capacity still bills at $1.50 an hour once a workload wakes it, which makes it the wrong shape for spiky, low-volume workloads and the right shape for steady ones.

Ingestion pricing is the line item that surprises people

Fivetran bills on Monthly Active Rows. Their own documentation defines it precisely: a MAR is a distinct primary key added, updated or deleted and synced in a calendar month, and "a key synced 30 times counts as 1 MAR". That is more generous than it sounds for slowly-changing dimensions and considerably less generous for a high-churn event table. There is also a $5 base charge on every connection with monthly usage between 1 MAR and 1 million MAR, excluding the Free plan, and new connections get 14 days free.

The design consequence: route high-churn, high-volume event streams away from per-row managed ingestion and into the warehouse's own bulk-load path. Route low-volume, high-schema-churn SaaS sources into managed connectors, because the maintenance you avoid is worth more than the MAR fee. Getting that split wrong is the most common six-figure mistake we see on inherited platforms.

On the transformation side, dbt's Developer tier is free for one seat, 3,000 successful models a month and one project. Starter is $100 per user per month with five developer seats and 15,000 models a month. Enterprise pricing is not published. dbt Labs and Fivetran have completed a merger, which is worth knowing before you architect a stack that assumes the two remain independently priced.

Orchestration: Airflow 2 is already gone

Apache Airflow 3.3.1 is the current stable release. Airflow 3 first shipped on 22 April 2025, and Airflow 2.x reached end of life on 22 April 2026, with limited maintenance having ended on 22 October 2025. The project's own wording: "EOL versions will not get any fixes or support."

If your platform still runs an Airflow 2 scheduler in August 2026, it is unpatched. That is the first thing we check on a takeover engagement, ahead of anything about model quality, because it is the one finding that is both certain and time-bound.

Our five-step delivery process

1. Discovery and cost model, 1 to 2 weeks. Source inventory, volume and cardinality profiling, and a twelve-month run-cost model against the list prices above before any code is written. If the model says the platform costs more than the decisions it supports are worth, we say so at this stage.
2. Landing and contracts, 2 to 3 weeks. Ingestion into a raw layer with schema contracts and freshness expectations agreed per source, plus the MAR-versus-bulk-load split decided deliberately.
3. Transformation and testing, 4 to 8 weeks. Staging, intermediate and mart models with tests that gate the build. Sized by the number of source systems, not the number of dashboards requested.
4. Orchestration and observability, 2 weeks. Airflow 3 or a managed equivalent, with freshness SLAs, retries and ownership-aware alerting.
5. Cost governance and handover, 1 to 2 weeks. Query tagging, warehouse auto-suspend policy, per-team attribution reporting, and runbook documentation your team owns.

Typical first engagement: 10 to 16 weeks for a single-warehouse platform with 4 to 8 source systems. We work on a fixed-scope phase model with a named senior engineer per phase, not a time-and-materials open end. Exact commercials depend on source count and data volume, so we scope before we quote.

The stack we use

Warehouses: Snowflake, BigQuery, Databricks, Redshift. We are not tied to one, and the choice usually falls out of the cost model rather than preference. Steady, predictable query volume favours BigQuery slot commitments or Redshift provisioned nodes; bursty analyst-driven workloads favour Snowflake's per-second credit billing or BigQuery on demand.

Ingestion: managed connectors where the source is a well-known SaaS API, native CDC where the source is a database we control, custom Python extractors where neither applies.

Transformation: dbt for SQL-first teams, with tests and contracts mandatory rather than optional.

Orchestration: Airflow 3, or the cloud-native scheduler where the platform is single-cloud and the DAG count is low.

We are a Microsoft, AWS and Google partner, which matters mainly because it gives us support escalation paths on the platforms above rather than because it changes the price you pay.

Why eCorpIT

Founded in 2021 and based in Gurugram, eCorpIT is CMMI Level 5 appraised, MSME certified and ISO 27001:2022 certified. Data platform work is delivered by senior engineering teams, not staffed to juniors behind a senior name on the proposal. We publish the run-cost model as a deliverable in phase one, which means you can cancel after two weeks with a costed decision in hand rather than a slide deck.

We design data platforms aligned with DPDP Act 2023 requirements, including in-region processing where the classification calls for it. We do not claim certification against frameworks we do not hold.

India-specific considerations

For platforms holding personal data on Indian data principals, the Digital Personal Data Protection Act 2023 sets the frame and the DPDP Rules 2025, notified as G.S.R. 846(E) on 14 November 2025, set the mechanics. Section 16(1) of the Act gives the Central Government power to restrict transfers to notified countries, which is a negative-list mechanism rather than blanket localisation. The harder constraint is Rule 13(4): a Significant Data Fiduciary must ensure that personal data specified by the Central Government, and its traffic data, is not transferred outside India. Rules 5 to 16 commence eighteen months after publication.

Penalties run to ₹250 crore for failure to maintain reasonable security safeguards, per the Press Information Bureau summary issued in November 2025.

Practical effect on architecture: because Snowflake prices Mumbai and Pune identically to N. Virginia, keeping an Indian workload in-country costs nothing on the warehouse line. The cost, where there is one, lands on cross-region replication and on the engineering effort to keep a residency boundary honest across ingestion, transformation and BI. Budget for the boundary, not the region.

What drives your bill, in order

1. Scan volume, not storage. Storage at $23 per TB per month is rounding error next to $6.25 per TiB scanned run forty times a day.
2. Idle compute. Auto-suspend set to 10 minutes instead of 1 minute, multiplied across eight warehouses, is a five-figure annual line.
3. Ingestion row churn. A high-update event table on per-row managed ingestion.
4. Full refreshes. Incremental models that quietly full-refresh because a schema change broke the incremental predicate.
5. Seats. dbt at $100 per user per month is real money at 20 seats and invisible at 3.

If you want the wider cloud picture rather than the data-platform slice, our cloud FinOps guide for Indian teams covers the same discipline across compute and storage, and the AI and cloud cost breakdown covers what AI workloads add. On the build-versus-buy question for transformation tooling specifically, see our comparison of the BigQuery data engineering agent against dbt and Dataform.

FAQ

What does a data warehouse actually cost per month?

It depends on scan volume more than data volume. Snowflake Standard is $2.00 per credit and $23.00 per TB per month for storage as of August 2026. BigQuery charges $6.25 per TiB scanned on demand in us-central1 with the first TiB free each month. Model twelve months of query patterns before choosing.

Is Snowflake more expensive in India than in the United States?

No. Snowflake's on-demand credit price in AWS Asia Pacific (Mumbai) and Azure Central India (Pune) is $2.00 Standard, $3.00 Enterprise and $4.00 Business Critical, identical to AWS US East. Storage is $23.00 per TB per month in Mumbai, also matching. AWS São Paulo, by contrast, is $3.10 on Standard.

Should we use Fivetran or build our own connectors?

Split by row churn. Fivetran bills Monthly Active Rows, where a key synced thirty times counts as one MAR, which suits slowly-changing SaaS sources. High-churn event tables are better routed through the warehouse's bulk-load path. There is also a $5 base charge per connection between 1 and 1 million MAR.

Is Apache Airflow 2 still supported?

No. Airflow 2.x reached end of life on 22 April 2026, after limited maintenance ended on 22 October 2025. The project states that EOL versions receive no fixes or support. Airflow 3.3.1 is the current stable release, with Airflow 3 having first shipped on 22 April 2025.

How long does a first data platform engagement take?

Typically 10 to 16 weeks for a single warehouse with four to eight source systems, across five phases: discovery and cost model, landing and contracts, transformation and testing, orchestration and observability, then cost governance and handover. Source count drives the timeline more than dashboard count does.

Do we need dbt, and what does it cost?

dbt Developer is free for one seat, 3,000 models a month and one project. Starter is $100 per user per month with five developer seats and 15,000 models a month. Enterprise pricing is not published. dbt Labs and Fivetran have completed a merger, which is worth factoring into a multi-year tooling decision.

Where does data engineering spend usually overrun?

On the run, not the build. dbt Labs' 2026 State of Analytics Engineering report, from 363 practitioners and published 14 April 2026, found 57% reporting increased warehouse and compute spend against only 36% reporting increased team budgets. Scan volume and idle compute are the two largest drivers.

Can our data stay in India under the DPDP rules?

Yes, and on Snowflake it costs no more than a US region. DPDP Rules 2025 Rule 13(4) requires a Significant Data Fiduciary to keep Central Government-specified personal data and its traffic data inside India. Section 16(1) of the Act allows transfer restrictions by notification, which is a negative list rather than blanket localisation.

How eCorpIT can help

eCorpIT builds and runs data platforms for teams in India, the UK and the US, and we start every engagement with a costed twelve-month run model rather than an architecture diagram. If you are choosing a warehouse, inheriting a platform whose bill is growing faster than its usefulness, or still running an unsupported Airflow 2 scheduler, that first phase will tell you what to do in two weeks. Look at how we scope data platform engineering and cloud FinOps as a managed service, then talk to us about your source inventory.

References

1. Snowflake Credit Consumption Table — Snowflake, on-demand platform credit and storage pricing effective 10 August 2026.
2. Snowflake pricing options — Snowflake; region and platform prices are behind a client-side selector.
3. BigQuery pricing — Google Cloud, on-demand and edition slot-hour rates, us-central1.
4. Amazon Redshift pricing — AWS, Serverless RPU-hour and provisioned node rates, US East (N. Virginia).
5. Azure Databricks pricing — Microsoft Azure, $/DBU by workload and tier, US East.
6. dbt pricing — dbt Labs, Developer, Starter and Enterprise tiers.
7. Fivetran and dbt Labs complete merger — dbt Labs.
8. Fivetran usage-based pricing — Fivetran documentation, Monthly Active Rows definition and base charge.
9. Apache Airflow supported versions — Apache Software Foundation, Airflow 2.x EOL 22 April 2026.
10. 2026 State of Analytics Engineering — dbt Labs, 363 respondents, published 14 April 2026.
11. Digital Personal Data Protection Act 2023 — Ministry of Electronics and Information Technology.
12. Digital Personal Data Protection Rules 2025, G.S.R. 846(E) — MeitY, notified 14 November 2025.
13. DPDP Rules 2025 summary and penalties — Press Information Bureau, November 2025.

Last updated: 17 August 2026.