Harell Data Picks CoreWeave to Run AI Training Without Sending Out Raw Data

Harell Data plans to put CoreWeave’s cloud behind a way to train AI on proprietary scientific data without sending the raw datasets to model builders. Under a multi-year agreement, Harell will deploy its platform on CoreWeave Cloud for training, fine-tuning and inference. The deal puts computing capacity behind Harell’s central promise: builders can work with valuable data without taking possession of it.

The model travels to the data

Harell’s platform is built around a separation: model builders can run training against a dataset without receiving the underlying data. CoreWeave says those runs will execute inside Harell’s platform on NVIDIA A100 and Hopper GPUs. The raw data stays there; what a builder takes away is the trained model, its intellectual property and the right to sell it. That is Harell’s description of the design, not an independent security assessment.

Harell argues that handing valuable research data to another party for commercial model training can destroy its value. Keeping the files inside its platform is its proposed answer to that risk. The CoreWeave agreement is intended to supply the cloud infrastructure as Harell scales the service for researchers.

We created Harell Data to connect researchers to proprietary scientific datasets for model training and inference without requiring them to view or extract the underlying data.

Harlan Robins, founder and president of Harell Data

Two ways to earn from one dataset

The agreement also backs a marketplace with payments on both sides of a model’s life. Harell says it meters computing use for each training run and gives the dataset owner a revenue share whenever a job uses its data. That ties payment to training activity rather than requiring the owner to hand over a copy. Harell has not specified the owner’s share.

Builders have a separate route to revenue. They can list models trained through Harell on its marketplace under a Models as a Service arrangement and collect a fee each time someone runs one. A data partner earns when its dataset is used for training; a builder can earn from later use of the resulting model. Those are Harell’s stated payment terms, not earnings figures from the CoreWeave deployment.

The test data stays hidden, too

Harell applies a similar separation when models are evaluated. Its data partners keep test questions hidden from builders, so a builder has not seen the answers before a result is published. Harell says results are then published against one standard. Unlike keeping training data in place, this step is meant to make model evaluations comparable without exposing the test material to builders.

The deployment is still planned, and the announcement includes no results from jobs run under the new agreement. For researchers, the next test is how the CoreWeave-backed platform performs when they use it—not only how its data-handling design is described.