Databricks agrees to acquire MosaicML for $1.3bn

Databricks announced a definitive agreement to acquire MosaicML, a start-up that built tools for training large language models efficiently and had released the open-weight MPT family of models, for approximately $1.3 billion including retention packages — roughly six times MosaicML’s $222 million valuation at its previous funding round, in which it had raised about $64 million from investors including DCVC, AME Cloud Ventures and Samsung Next. MosaicML’s MPT-7B, released in May 2023, had by the time of the deal been downloaded more than 3.3 million times, and the company had shortly before released a larger MPT-30B model.

Databricks framed the acquisition around MosaicML’s training-efficiency technology, which it said delivered two to seven times faster training than standard approaches, and around a shared goal of making it economical for companies to train their own custom models rather than relying solely on third-party APIs — cutting the cost of training a multi-billion-parameter model from millions of dollars to, Databricks said, the thousands, and the time from days to hours. Chief executive Ali Ghodsi said the combination would help “democratize AI and make the Lakehouse the best place to build generative AI and LLMs,” referring to Databricks’ existing data-platform product. MosaicML’s research team, including the group behind MPT, was expected to join Databricks on completion.

The deal was among the largest acquisitions of an open-model company during the 2023 generative-AI funding boom, and set Databricks — a data-infrastructure company without a frontier model of its own — up as a vendor selling the tooling to build and fine-tune open models in-house, an alternative business model to the closed, API-only approach of OpenAI and Anthropic.