Buying snapshot
Scroll sideways to see all columns. With a keyboard, focus this table and use the arrow keys.
| Decision point | What Palantir states |
|---|---|
Platform | AIP, Foundry (data operations) and Apollo (autonomous software deployment) together form the Palantir platform |
AIP features | Availability is subject to change and may differ between customers |
Compute | Autoscaling runtimes: Spark for batch, Flink for streaming, and single-node engines such as DataFusion, Polars and DuckDB |
Language models | Commercial LLMs and custom models through the Model Catalog, or your own models and provider accounts (Bring Your Own Model) |
Python SDK | Public on GitHub under the Apache-2.0 license; a client for the Foundry API, not the platform itself |
What Foundry measures
Foundry reports all computational work in compute-seconds, a unit of work rather than of time, and all storage in gigabyte-months. These are the units to ask Palantir to quote against.
Scroll sideways to see all columns. With a keyboard, focus this table and use the arrow keys.
| Usage type | How it is measured |
|---|---|
Batch compute (Transforms, Pipeline Builder, syncs, Data Health checks) | Compute-seconds from when resources are provisioned until they are released; nothing while the job is not running |
Interactive compute (Contour, Ontology and timeseries queries, Code Workspaces) | Compute-seconds per query, rolled up to the Project; these systems keep idle compute running so they can answer fast |
Continuous compute (Streaming, Foundry Machine Learning) | Time the job is ready to receive messages; not available on all Foundry environments |
Foundry storage | Gigabyte-months; files removed with a DELETE transaction keep accruing storage, and only Retention reduces size |
Ontology volume | Gigabyte-months for the indexes of object types |
How to track it. Each item's usage accrues to its parent Project, and each Project belongs to a usage account. By default there is a default account and a user folders account; administrators can add accounts, for example by department or business unit. Resource Management compares usage with enrollment-level limits and lets administrators set limits per Project. Usage is recorded in UTC, and a job counts on the date it finished.
Workload choices that affect compute
Job size. Compute-seconds grow with the number and size of executors, the data size and job complexity. More parallel executors finish most jobs sooner but use more compute-seconds per unit time. Usage counts allocated vCPU cores, not how busy they are, and includes spin-up and cleanup time.
Batch or incremental pipelines. A batch pipeline fully recomputes all datasets on each run. Incremental transforms process only changed rows or files, which Palantir says reduces latency and compute costs for large-scale datasets.
Engine choice. Python transforms can run on single-node engines (Pandas, Polars, DuckDB) or on distributed Spark. Match the engine to the data size.
Model usage. With Bring Your Own Model, you connect your own LLMs or model provider accounts, with rate limits and usage observability through Resource Management. Palantir recommends it when Palantir-provided models cannot be used for legal or compliance reasons, or for fine-tuned models.
Data copies. Virtual tables connect Databricks, Snowflake and cloud providers without duplicating the data, so you can decide which data to move into Foundry and which to read in place.
Your own compute. Compute Modules let you bring containerized runtimes, models and applications.
What to confirm in an offer
Which AIP features you get. Because AIP features can differ between customers, list the ones your workflow needs and check each is covered.
AIP Bootcamps. Palantir describes bootcamps where customers work hands-on and reach outcomes with AI in a matter of hours. Ask whether one is part of your offer and on what terms.
Work your team keeps. Incremental pipelines need careful handling of schema changes. Plan the integrations, permissions and pipeline upkeep your own staff will own.
Our take
Palantir makes sense to price when you need integrations, applications and AI agents on one platform. Run one real workflow first and read its compute-seconds and gigabyte-months in Resource Management, so the scope you request rests on measured use rather than guesses.