Databricks and MosaicML CEOs say $1.3 billion deal is about helping enterprises create AI IP

The news that Databricks had signed a definitive agreement to acquire AI startup MosaicML for an estimated $1.3 billion made headlines yesterday. But for the two companies’ CEOs, the deal was a natural fit and all about helping speed up enterprise adoption of generative AI. 

Databricks CEO Ali Ghodsi told VentureBeat that for months, customers had been asking him the same questions: “How can I create my own generative AI models? How do I own them and create my intellectual property around this? Can you help me?” 

Over and over again, he explained, MosaicML came up as a company helping enterprises do just that. And more often than not, customers using MosaicML were also using Databricks to build their own generative AI models. 

For example, Replit, a cloud-based collaborative code editing platform, had even published a blog post back in April detailing its efforts to train its own large language models (LLMs) from scratch, using MosaicML, Databricks and Hugging Face. 

“It became clear to us that if we could join forces, we could make it even cheaper and better for the customers,” Ghodsi said. “And if we can make it cheaper, we can grow revenue because we can satisfy more demand and help further democratize generative AI.” 

That requires a tight integration of the two companies: “At some point it becomes clear that if you work closely together, I have to give you the keys to my house and maybe you should have a toothbrush there, and you know, we might as well just essentially get married,” he joked.

Databricks and MosaicML ‘are very aligned’

For MosaicML CEO Naveen Rao, who co-founded the San Francisco-based company in 2021 — which allows businesses to train models on their own data using the company’s model architectures and then deploy the models through its inference API — said Databricks and MosaicML “are very aligned,” saying “going faster is what this deal is all about — can we can we go faster and better serve our customers?”

Databricks, he added, has been an AI company from the beginning. “The data warehousing and lake house is where they found traction and scale,” he told VentureBeat. “Their customers are asking for LLMs and we provide that, it just makes sense. Then we met and discussed our viewpoint on the world — we’re similar in a lot of ways.” 

He pointed to Databricks debut of the ChatGPT-like Dolly in March, which was meant to show that organizations don’t need the latest or largest LLM, but can use a fine-tuned, smaller model. Meanwhile, just last week MosaicML released its own open-source LLM, MPT-30B, which it said was trained at a fraction of the cost of its competitors.

“Dolly is a fine-tuned model, based on a two-year-old open source model, but actually building these models from scratch is a different amount of effort,” Rao explained. “I think what’s interesting is that the Databricks team is smart enough to understand how hard that is, and they said ‘we want this capability because our customers want it.’” They saw us as the leader and it was pretty natural for them.”

AI model IP is key for many enterprises

The most important thing for many enterprises, insists Ghodsi, is model IP.

“The customers themselves are telling me they want to own their models,” he said. “I’ve met many of the Fortune 500 CEOs in the last six months because they are very interested in generative AI, and the number one thing they asked me is, ‘Will I have any intellectual property here?’ ‘Will I be able to own a model that I can then use to gain competitive advantage over my competitors?’ And we want to tell them yes.”

If an organization wants to be competitive and thinks AI is the future of its business, they want to own their model, he explained, they are trading on their own “extremely valuable” data.

“My question is, is there going to be an alternative in this space or not? I think it’s fair to say that if you want to build your own model from scratch, Mosaic is one of the best out there,” he said.

More ‘fuel on the fire’

Rao emphasized that he does not see the $1.3 billion Databricks deal as an “exit.”

“I see this really as an inflection point and accelerant,” he said. “Nothing changes for me. I’m coming to work every day. It’s just now we can go faster. Once we close the deal, we have more resources, more experience. These guys have been at it, servicing enterprise for 10 years. They built up a lot of knowledge — so how do we use that to go faster? It’s just more fuel on the fire in so many ways.”