See how much of your AI spend comes from your data.
Answer: Six questions about your platform, model, and data using our AI Cost Calculator.
Get Back: An estimate on how much you're overpaying for your queries from data rediscovery.
.svg)
You could be saving $2,534 a year
Based on 5,401 queries a month on Microsoft Fabric with o4-mini. Change any input above and this recalculates instantly.
Summary
Where the savings come from
How every number here is calculated
Every AI query against your data pays up to three costs: ingestion (the LLM re-reading your table, context, or schema), processing (reasoning over the question and generating an answer), and retries (re-asking when the first answer misses). None of that context normally persists between sessions or users, so a naive setup pays all three on every single query. A warehouse, a semantic model, and pre-populated context each chip away at that stack.
Effective input tokens = your average input tokens × the ingestion multiplier for your score. That plus your output tokens is multiplied by the retry multiplier. The cache share sets what fraction of input tokens bill at the model's cached-input rate. Llama 3.1 405B Instruct and GPT-OSS 120B publish no cached-input rate, so cached tokens fall back to the standard rate for those two. That means no caching discount, which is accurate rather than a simplification. The DataOS figure runs the identical formula on the same model with the score fixed at 4, isolating the effect of the architecture.
Each platform bills differently, so the calculator uses the formula that matches how that platform actually charges, then normalizes to a monthly total divided by your query volume:
- Snowflake: credits/hour for your warehouse size × $3.75/credit × active hours
- Databricks: $/DBU for your SQL tier × DBUs/hour × active hours. Excludes the separate underlying cloud infrastructure Databricks bills alongside DBUs, which commonly adds 50–200% on top
- Microsoft Fabric: $0.18/CU-hour × capacity units × active hours, less the reserved-capacity discount if selected
- AWS Athena and Google BigQuery: $/TB (or TiB) scanned × data per query × monthly volume. Billed per query by nature, so active hours don't apply
- Salesforce Data Cloud: $0.01/credit × credits per query × monthly volume
- Palantir Foundry: annual platform fee ÷ 12. Palantir has no per-query pricing, so this is a flat subscription amortized for comparability
- No warehouse: $0, since there's no dedicated compute layer
For DataOS: the hourly rate for your cloud and instance size × your DataOS active hours, because DataOS runs on your own cloud infrastructure rather than charging warehouse credits.
Fabric, Salesforce Data Cloud, and Palantir don't publish per-unit pricing the way Snowflake, Databricks, Athena, and BigQuery do. Rather than leave them blank or invent precise-looking numbers, each uses a clearly-flagged, editable default:
- Microsoft Fabric: ~$0.18/CU-hour, from a third-party Azure pricing breakdown. Confirm against Microsoft's own calculator for your region
- Salesforce Data Cloud: $0.01/credit is Salesforce's published rate, but credits per query is not published. Defaults to a presumed 2
- Palantir Foundry: $300,000/year, the low end of published single-use-case estimates. Palantir is contract-negotiated, so treat it as a placeholder, not a quote
This is a directional estimator built to support a conversation. It is not an audited total cost of ownership, and not a substitute for a proposal built around your actual workloads, contracts, and usage logs. If a number looks off for your situation, change the input. Rates were last verified in August 2026.