AI infrastructure that lets enterprises own their data and the intelligence built on it, and scale both efficiently.
Exabyte-scale data infrastructure, generative models for tabular data, and stateful agent infrastructure.
Granica is an AI research and products company, driving efficiency and trust across an AI stack the enterprise owns end to end.
Myelin is the stateful infrastructure for long-running agents.
A long-running agent loses everything the moment its session drops. Myelin keeps its state alive, so the agent resumes exactly where it left off and holds up at scale, efficiently. You build the agents; Myelin makes sure they never start over.
Crunch is infrastructure for exabyte-scale data.
At that scale, the work never stops. Crunch runs continuously on your data, wherever it lives, in the cloud you already own, keeping it optimized, secure, and ready for AI. In production, teams cut data costs by up to half, accelerate their workloads proportionally, and stop babysitting brittle jobs.
And you pay for the outcome, not the software: pricing follows the value Crunch creates, so it pays for itself.
Granica Crunch cut serving cost per MAU, and unit economics improved with it.
$200K / PB
Annualized ROI across storage and compute.
4 weeks
Time to value, from kickoff to verified savings.
Hundreds
Recurring jobs eliminated, freeing engineers.
“Lowering the infrastructure cost to support each MAU is critical for us as the platform scales. Granica handled the heavy lifting, from setup to daily scheduling to verification, so we saw real savings quickly without disrupting our Delta pipelines, and could redirect engineers back to building the platform.”
Arya Ketan
Distinguished Engineer, ShareChat
Granica Research Lab
Rigorous research on the data foundations of AI
At Granica, our dedication to efficiency is rooted in fundamental research on the problems at the heart of understanding and working with large-scale data. In the age of massive data, questions around data selection and compression become paramount in practical workflows:
- How do you subsample a dataset so that training stays tractable without throwing away the signal?
- When labels are scarce, can a small, expensive-to-collect dataset borrow strength from a larger one in a related domain?
- And how far can data be compressed before the structure that makes it useful begins to disappear?
Failure to leverage data properly means models trained at enormous cost, only to learn the wrong signal; pipelines that buckle under scale; and sub-optimal decisions made on data that was never representative in the first place.
The Granica research team has developed answers to such non-trivial questions, leading to published work at top machine learning venues such as ICML, ICLR, KDD, and NeurIPS. One of them, Towards a statistical theory of data selection under weak supervision, earned an honorable mention for Outstanding Paper at ICLR.
The insights behind this work flow directly into the products we ship.
Large Tabular Models: What we are building
Granica develops efficient solutions to hard data problems arising in enterprise. We see the next frontier as bringing generative AI to tabular data.
Data in enterprise workflows naturally comes packaged as tables, which the current text-and-image generative paradigm was never designed for. Bringing the full power of generative AI to enterprise therefore calls for a class of models that works natively with tabular data. We refer to these as Large Tabular Models (LTMs).
Building an LTM comes with a host of new challenges. The same questions we have tackled in our prior research are exactly the ones that prove critical to whether a tabular generative model succeeds: how to select, augment, and compress data without losing what matters. An LTM must be trained on judiciously selected data to learn well, and it must draw on related datasets to augment domains where examples are scarce. For LTM training to be tractable, the data must be compressed enough to fit within computational limits, without discarding the structure that carries the signal.
We bring our expertise on these questions directly to bear on building LTMs. For more on the new frontier, and the first step we’re taking toward building them, see our post and repo below:
Selected publications
- Train on Validation (ToV): Fast data selection with applications to fine-tuningICLR 2026Read paper
- Scaling laws for learning with real and surrogate dataNeurIPS 2024Read paper
- Towards a statistical theory of data selection under weak supervisionICLR 2024Read paper
- Scaling training data with lossy image compressionKDD 2024Read paper
- Compressing tabular data via latent variable estimationICML 2023Read paper
- Sampling, diffusion, and stochastic localizationPreprintRead paper
- Inline data detection in large data streamsUSPTO patentRead patent
- Efficient data deduplication through sketch computation and similarity metricsUSPTO patentRead patent
AI should belong to the enterprise that runs it.
Everything we build follows from that.
- 001.
Efficiency
You cannot scale what you cannot afford to run.
The cost of intelligence is the cost of moving data and computing on it. You do not get past that by spending more; you get past it by spending less per unit of intelligence. So we compute where the data already lives. The most capable intelligence we know runs on twenty watts. We treat that as the standard, not the exception.
- 002.
Trust
What's yours stays in your house.
Data stays in the customer's environment, and so do the models trained on it. We keep no copy. Trust earned this way isn't a promise in a contract. It's a property of where the system runs.
- 003.
Control
It answers to the people who own it.
The system belongs to the enterprise. The infrastructure we provision for agents makes that literal: humans and agents act on the same data under the same controls, and authority does not fork when the work passes from a person to an agent.
If this is the work you want to be doing, join us.
We are a lean team of builders from Stanford, Google, Amazon, and Snowflake; backed by NEA and Bain Capital, with over $60M raised.