Key points
- A multi-year deal with no dollar figure disclosed
- Harell says raw datasets stay inside its platform
- Harell says data owners earn a share of each training job
- The GPUs are Nvidia A100 and Hopper chips
CoreWeave (CRWV) signed a multi-year deal to power Harell Data's platform for training AI on private biotech datasets, the companies announced Wednesday. Harell says developers can train models on its platform without seeing or extracting the underlying data.
The agreement covers training, fine-tuning and inference on CoreWeave's cloud. The companies disclosed no contract value, exact duration or computing capacity, and named no data providers or model builders using Harell's platform.
CoreWeave shares traded at $87.17 in premarket trading as of 8:57 a.m. Eastern time Wednesday, up about 0.5% from Tuesday's $86.76 close.
How the platform works
Harell's approach is to bring the model to the data. Training runs execute inside Harell's platform on Nvidia (NVDA) A100 and Hopper GPUs, and the raw data never leaves it, according to the release. What the model builder takes away is the trained model and the rights to sell it.
Harell meters compute for each training run, and the owner of the dataset earns a revenue share on every job run against it. Builders can then list their models on the Harell marketplace and collect a fee each time someone uses them. Harell calls this arrangement Models as a Service.
Testing follows the same separation. Harell's data partners keep the test questions hidden, so no builder has seen the answers, and every result is published against one standard, the release says.
Harell describes itself as a platform for training on proprietary data "in biology and beyond," with biotech datasets as the starting point. The company is based in Bellevue, Washington, and was founded by Harlan Robins, a co-founder of Adaptive Biotechnologies.
"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," Robins said in the release.
How the deal compares
The A100 and Hopper are earlier Nvidia generations than the Blackwell and Vera Rubin systems CoreWeave has featured in other recent deals. CoreWeave's August agreement with Hudson River Trading specified access to Vera Rubin NVL72 and HGX B200 systems.
Without a disclosed contract value, investors cannot assess the agreement's contribution to CoreWeave's backlog. The announcement comes a week after CoreWeave launched a convertible notes offering and an at-the-market share program. CoreWeave completed a $4.2 billion convertible notes offering on September 22, including the full exercise of a $500 million purchaser option.
