Granica

Granica positions itself as an efficiency layer for enterprise AI, aiming to lower the cost of running AI workloads from the data layer through agent execution. Its product line comprises Crunch, a data infrastructure service that continuously compresses, organizes and maintains data lakes according to customer-set policies to shrink storage and reduce bytes scanned by queries; Myelin, agent infrastructure that persists agent state across sessions, machines and handoffs so long-running agents resume with prior files, decisions and constraints rather than rebuilding context; and Large Tabular Models (LTMs), a model class the company is developing to apply generative AI natively to governed enterprise tables for predictions and synthetic data generation.

A defining architectural choice is that software and models run inside the customer's own cloud perimeter, with Granica stating it keeps no copy of customer data and that models trained on that data remain in the customer environment. The company frames this as making trust a property of deployment location rather than a contractual promise, and extends the same principle to agents, where humans and agents are described as acting on the same data under the same controls.

Granica also operates the Granica Research Lab, which works on data selection, data augmentation from related or surrogate domains, and lossy/lossless compression limits for large-scale datasets. The lab reports published work at ICML, ICLR, KDD and NeurIPS, including "Towards a statistical theory of data selection under weak supervision" (ICLR 2024), which received an honorable mention for Outstanding Paper, "Scaling laws for learning with real and surrogate data" (NeurIPS 2024), "Scaling training data with lossy image compression" (KDD 2024), "Compressing tabular data via latent variable estimation" (ICML 2023), and "Train on Validation (ToV)" (ICLR 2026). The company also lists two USPTO patents covering inline data detection in large data streams and data deduplication via sketch computation and similarity metrics.

Business model

Granica sells software deployed into the customer's own cloud environment and states that it uses outcome-based pricing tied to the value created, with Crunch positioned around verified cost savings across storage and compute.

Outcome-based pricing tied to the savings or value delivered to the customer, per the company's stated commercial model.

Traction

The company cites annualized ROI of about $200K per petabyte across storage and compute, data lake cost reductions of 20% to 50% with Crunch, hundreds of recurring maintenance jobs eliminated, and a four-week time to value. For Myelin it reports 95.6x context resumed from cache instead of rebuilt, 10x more agent sessions since adoption, and 4.2 billion tokens of agent work run per day. ShareChat is named as a customer, with a distinguished engineer quoted on reduced serving cost per monthly active user.

Latest developments

Granica has published a post and an open GitHub repository on Large Tabular Models describing the first step toward that model class, and lists a Train on Validation (ToV) data-selection paper at ICLR 2026. Myelin is offered as agent infrastructure alongside Crunch.