Data Marketplace

← All posts

Compute-to-Data: How to Monetize Sensitive Datasets Without Sharing Raw Files

Data Marketplace Team ·

You have a dataset worth millions. Patient health records. Credit bureau data. Genomic sequences. Customer transaction logs. The insights locked inside could power transformative analytics — fraud detection, drug discovery, personalized marketing, predictive maintenance.

But you can't sell it.

Not because the insights aren't valuable. Because sharing the raw data is either illegal (GDPR, HIPAA), catastrophically risky (re-identification, breach liability), or both. So the data sits unused, the insights remain locked, and both the provider and the potential buyer lose.

Compute-to-data changes this equation entirely.

The Principle: Bring the Computation to the Data

The traditional model of data commerce is: buyer pays, seller sends file, buyer analyzes locally. This model assumes the raw data can leave the provider's control. For sensitive datasets, it cannot.

Compute-to-data inverts the model: the buyer submits a computation (an aggregation query), and the provider executes it server-side. The raw data never leaves. Only the aggregated result is returned. The buyer gets the insight. The provider retains control.

How It Works on Data Marketplace

Our Privacy-Preserved Federated (FED) license archetype enables compute-to-data as a first-class transaction type:

1. Listing

The provider lists a dataset under the FED archetype. The ODRL+DPV policy explicitly permits query and compute actions while prohibiting extract, reproduce, and distribute. The DPV profile declares dataLocalisation: true — the data stays put.

2. Acquisition

The buyer purchases a FED license. This doesn't grant file access — it grants computation rights within the policy's constraints (permitted purposes, usage count limits, time windows).

3. Query Submission

The buyer (human or agent) submits an aggregation query to POST /agent/v1/compute:

{ "datasetId": "ds_abc", "aggregation": "avg", "column": "salary", "groupBy": "department" }

4. K-Anonymity Enforcement

Before returning results, the platform enforces k-anonymity: any group with fewer than the minimum aggregation group size (default: 10) is dropped from the results. This prevents reconstruction attacks where a buyer uses highly specific groupings to isolate individual records.

5. Result + Proof

The buyer receives aggregated results plus a cryptographic execution proof:

  • Query hash: SHA-256 of the query parameters (proves what was asked).
  • Result hash: SHA-256 of the returned data (proves what was received).
  • Input commitment hash: links to the provider's registered dataset commitment (proves which version of the data was queried).

This proof chain means the buyer can demonstrate to auditors exactly what computation ran against exactly which data, without ever possessing the raw records.

What Aggregations Are Supported?

The platform currently supports seven aggregation functions: count, sum, avg, min, max, stddev, and median. All with optional groupBy and filter clauses. This covers the vast majority of analytics use cases — cohort analysis, summary statistics, trend detection, benchmarking.

Arbitrary code execution (custom UDFs, ML training) is on the roadmap but requires a more sophisticated sandboxing model. The current aggregation-only approach is deliberately constrained to minimize information leakage.

Who Should Use FED?

  • Healthcare providers with patient data that can't leave their infrastructure under HIPAA.
  • Financial institutions with transaction data subject to banking secrecy regulations.
  • Telcos with location data that's personal under GDPR but valuable for mobility analytics in aggregate.
  • HR platforms with compensation data that individuals could be identified from.
  • Any data provider whose dataset contains PII that can't be lawfully transferred as raw files.

The Economics

Compute-to-data unlocks revenue from datasets that currently generate zero — because the alternative is not selling the raw file; it's not selling at all. Providers can set per-query pricing, monthly query caps, or flat subscription rates for unlimited compute access.

For buyers, FED access is typically priced lower than raw-file access (since they can't redistribute or derive freely), but the analytics value may be identical or higher — especially for use cases that only need aggregate statistics.

Getting Started

If you're a data provider sitting on sensitive datasets that you've been unable to monetize: list them under the FED archetype. Set your k-anonymity threshold. Define permitted purposes. Let buyers compute against your data without ever seeing a single raw record.

The sensitive data economy doesn't need to be a binary choice between sharing everything and sharing nothing. Compute-to-data is the middle ground — and it's live today.

List a FED dataset at datamarketplace.io. Buyers query; you keep control.

Compute-to-Data: How to Monetize Sensitive Datasets Without Sharing Raw Files — Data Marketplace