Databricks Strikes $1.3 Billion Deal for Generative AI Startup MosaicML

Databricks has agreed to acquire generative artificial-intelligence startup MosaicML in a deal valued at roughly $1.3 billion, a move aimed at capturing the fast-growing demand from businesses to build their own ChatGPT-like tools.

Databricks, a data storage and management startup based in San Francisco, says the deal combines its AI-ready data-management technology with MosaicML’s language-model platform, enabling businesses to build low-cost language models themselves with proprietary data. Right now, most businesses rely on third-party language models trained on troves of publicly available data accessed online. 

Also based in San Francisco, MosaicML, which was launched in 2021 and will become a stand-alone service belonging to Databricks, has been focused on bringing down the cost of using generative AI—from tens of millions of dollars to hundreds of thousands of dollars per model, according to Naveen Rao, co-founder and CEO. MosaicML has 62 employees and has raised $64 million to date.

The deal is expected to close during Databricks’s second quarter ending July 31.

Generative AI applications are designed to produce original text, images and computer code based on users’ natural-language prompts. Interest in the technology has surged since November, when AI startup OpenAI launched ChatGPT, an online generative AI chatbot.

Companies like Anthropic and OpenAI license ready-made language models to businesses, which then build generative AI apps on top of them. Driven by strong commercial demand for these models, the generative AI market has expanded dramatically—creating openings for startups like MosaicML that say they can offer similar AI models, but at lower cost and customized with a company’s data.

“If you’re building a model from scratch, you know what you’re feeding it,” said Databricks Chief Executive Ali Ghodsi. Off-the-shelf models, which are ready to use because they have already been trained on internet data, are filled with extraneous information that can skew results, Ghodsi said. Many companies are also wary of privacy and security issues around sharing their data in models built by outside vendors, he said.

Some machine-learning experts and AI vendors say that the computing and synthesis power of large language models like the one powering ChatGPT trumps smaller models, which have powerful, but ultimately limited capabilities in a specific area. Plus, ongoing challenges remain in data management, and determining which models are best suited for certain uses, said Sreekar Krishna, KPMG’s U.S. artificial-intelligence leader.

“Data has always been the key factor to success,” Krishna said, and the need for it has only increased now with large language models coming in.

Corporate technology leaders are facing pressure to get their data ready for AI models. Data serves as the foundation for all algorithms, because it is used to teach them to glean patterns and make predictions from it. 

Companies like Replit, which offers tools for programming, are already using Databricks for their data pipeline, and have piped that information to MosaicML to train a code generation model, Rao said.

Known as lakehouse, Databricks’s technology is designed to prepare and manage business data for AI applications, while unifying data, analytics and AI programming tools in one system. Databricks makes money by renting out analytics, AI and other cloud-based software that taps AI-ready data—what Ghodsi calls the “picks and shovels”—for building enterprise tech systems. Last year, Databricks reported more than $1 billion in annualized revenue. 

Spending in the global generative AI market is expected to reach $42.6 billion by the end of the year, and grow at a compound annual rate of 32% to $98.1 billion by 2026, according to market analytics firm PitchBook Data. It says venture funding in generative AI startups grew to $12.7 billion in the first five months of 2023 from $4.8 billion in all of 2022.

Databricks, founded a decade ago by a group of data scientists in Berkeley, Calif., has a private-market valuation of $38 billion, following a $1.6 billion fundraising round in August 2021. Its investors include Morgan Stanley’s Counterpoint Global, Andreessen Horowitz, Baillie Gifford, UC Investments and ClearBridge Investments. 

Larry Pickett, chief information and digital officer of biopharmaceutical services company Syneos Health, said the current cost of training a model on specialized health data is estimated at $1 million to $2 million. Those kinds of “domain-specific” models can be more useful for companies than ChatGPT, because they have more industry terminology and know-how, analysts say.

But Pickett expects that Syneos Health can spend significantly less than that by using smaller, pre-trained models, “as opposed to building on top of the entire corpus of data that OpenAI has.” Some of those models are already available in open-source libraries like those offered by machine-learning startup Hugging Face, he said.

“Not everybody, every application, requires a GPT-4,” Krishna said, referring to OpenAI’s large language model. Large language models are becoming fine-tuned for very specific applications, he said, and “at that point, it is so small that it could be embedded into any cellphone.”

Write to Angus Loten at [email protected] and Belle Lin at [email protected]