MosaicML

Platform for enterprises to train and deploy custom large AI models; acquired by Databricks in 2023.

Updated April 2026

Overview

Website: mosaicml.com
Founded: 2021
Headquarters: San Francisco, CA, USA
Segment: MLOps & Experiment Tracking

Product overview

MosaicML built the MPT series of open-source large language models and a platform for efficient training and fine-tuning of LLMs up to 70B parameters using proprietary data in private clouds. Used by companies like Replit and Hippocratic AI; integrated post-acquisition into Databricks Mosaic AI for model serving and training. Distinct for its focus on cost-effective training (2-7x faster), data privacy, multi-cloud orchestration, and full model ownership versus API-dependent competitors.

Revenue model

Pre-acquisition: pay-for-training platform with enterprise tiers ($30M ARR by 2023). Post-2023 Databricks acquisition: integrated into Mosaic AI Model Serving with pay-per-token DBUs (e.g., ~$1.50/M output tokens for Llama 3.3 70B) or provisioned throughput (e.g., $6/hr entry capacity).

Moat

MosaicML's competitive moat centers on its proprietary algorithmic efficiency and integrated platform architecture that dramatically reduces the cost and time of LLM training. The company has built defensible advantages through its specialized research in neural network optimization, cloud-agnostic infrastructure that abstracts away complexity, and deep integration within the Databricks ecosystem—creating switching costs for enterprises that have built their AI workflows around MosaicML's training stack and open-source models like MPT.

Headwinds

Integration challenges and potential loss of independent innovation following Databricks acquisition.