It’s not often that a founder with no intention of selling, and an acquirer that’s not looking to acquire, find themselves in a whirlwind frenzy of dealmaking, that a mere two months later results in one of the biggest A.I. deals of the year. But that’s exactly what happened with Databricks’ $1.3 billion buy of MosaicML, announced in late June. This is the story of how the two founders—Ali Ghodsi, the chief executive of Databricks, and Naveen Rao, MosaicML’s CEO—met, realized they have lots in common, and in warp speed came to combine their companies.
Ghodsi, 44, and Rao, 48, had long known of each other, but they didn’t actually meet until March 30. That’s when they both attended the Cerebral Valley conference in San Francisco, an AI summit that attracted about 200 people, including some of the biggest names in industry like Cristobal Valenzuela, CEO of Runway; Clem Delangue, CEO of Hugging Face; and Ghodsi, who spoke about the movement towards open source in AI.
After the conference, Ghodsi and Rao went to a VIP dinner that was held at Rich Table restaurant, where the executives were free to roam and chat. Over glasses of sangria and plates of bucatini, the two executives discussed the buying behaviors of customers in the cloud, client perceptions about security, as well as what things have worked out for Databricks, Rao said. At one point, Rao congratulated the Databricks CEO on all his success and Ghodsi responded, “Yeah, but we’re still a little guy. We need to do a lot more to grow.”
Databricks, however, is much bigger than MosaicML. Founded in 2013, Databricks is a data-analytics company that employs more than 5,500 people and was valued at $38 billion in 2021. MosaicML, by comparison, was launched two years ago, has 62 employees and is looking to disrupt the AI sector. MosaicML’s software lets startups use their own data to train and deploy large language models as well as other generative A.I. Instead of having to rely on Big Tech to train their AI models, which would be very expensive, companies can turn to MosaicML and potentially save millions of dollars, as well as weeks to months of time, Fortune has reported.
It was at this dinner that the leaders of Databricks and MosaicML realized they had similar goals, Ghodsi said. Both Ghodsi and Rao wanted to democratize AI and enable lots of people to have access to these models, both knew security was a big issue for their customers, and neither was interested in building a consumer app. This is counter to all the publicity generated by ChatGPT, the AI chatbox from Microsoft or Google’s Bard. “We’re not interested in that. We are focused on enterprises and big organizations, B2B, which is usually something very boring that no one cares about,” Ghodsi said. The CEOs exchanged numbers and agreed to do more together.
The meeting piqued Ghodsi’s interest. Rao is very well-known in the AI community, having sold his first company, Nervana Systems, to Intel in 2016 for $408 million. When he left four years later, in 2020, Rao was Intel’s top AI executive. But in late March, Databricks wasn’t necessarily looking to buy anything and MosaicML was definitely not interested in selling, Ghodsi said.
There were signs that a merger might be a good fit, Ghodsi said. Replit, whose software is popular in high schools to teach coding, is a big Databricks customer. On April 19, Replit announced that it used both MosaicML and Databricks to train large language models. “We click [with MosaicML], and they are stealing customers from us?” Ghodsi said with a laugh. “It seems even our customers were telling us we need to use both [Databricks and MosaicML] together to succeed.”
Ghodsi spent the weeks after Cerebral Valley reading up on Rao and MosaicML. Rao founded MosaicML in 2021 with the goal of making machine learning more efficient. The startup has raised $64 million in funding from venture firms like DCVC, Lux Capital, and Playground Global. Many of Databricks customers, including Replit, were using MosaicML to train their LLM, he said. “It was more or less known that if someone wanted to train an LLM they should talk to MosaicML,” Ghodsi said.
MosaicML seeks to dispel several myths about AI, including that building and training an LLM is vastly expensive. For example, Stable Diffusion was trained, using MosaicML’s software, to create AI-generated images for less than $50,000 in about 35 days, according to the startup’s pitch deck. Big Tech, by comparison, has invested billions in AI, and would likely have spent “orders of magnitude more” on Stable Diffusion, because they don’t have scarcity of funding and could throw more resources at all problems, Ghodsi told Fortune.
MosaicML also gives its customers the models, so they own them, Ghodsi said. This prevents any sharing of data with other customers. “You get full control over your IP,” he said.
On Friday, May 7, at around 5 pm, Ghodsi went to Chris Hecht, Databricks’ SVP, corporate dev and product partnerships, and told him he was interested in the startup. “MosaicML is on fire. If you’re interested, you probably have to move very fast,” Hecht said. He recommended that Ghodsi make an offer during the weekend. Ghodsi was unprepared by the response. Databricks hadn’t even begun discussing a potential MosaicML transaction. “Now I had the weekend to make a bid,” Ghodsi said.
The Databricks CEO didn’t know Rao that well but he texted him, asking “mind doing a zoom?” Ghodsi wanted to hold the call on Sunday, May 9 but Rao couldn’t speak that day so they scheduled it for Monday, May 10. He sent Rao a link to the video call.
Ghodsi, meanwhile, spoke to his team and prepared them for a very likely brushoff. The Databricks CEO is a well-experienced buyer. He expected Rao to say something like “Thank you, it’s very flattering, but now is not a good time,” which Ghodsi said is code for, “We don’t want to sell right now but maybe in a year or two. We’re doing too well.”
Rao, during this time, said he didn’t want to be part of a big company again. With a startup, 90% of an entrepreneur’s time is spent building, selling and recruiting, Rao said. That’s not what happens at a large business, he said. Executives at a startup can also move fast on issues, but when they’re part of a large company, they have to go through many layers of decision makers, Rao said. “I’m just not a big company person. I don’t enjoy the politics. I don’t enjoy the backstabbing,” he added. (Ghodsi said he often told Rao, “We’re not Intel. We’re not Intel.”)
Rao didn’t want to sell and he needed signs that Databricks was different. One of the most meaningful, and the simplest, was the Zoom link that Ghodsi sent to Rao in May. “The reality is, if I did that with somebody from Intel or Microsoft or Google, it would go through five layers of EA’s before the CEO had sent me an invite,” he said. The Zoom link was a signal that Databricks was still very founder-led and that Ghodsi was super involved directly with many things, Rao said.
There’s also the shared connection to academia. Rao has a PhD in neuroscience from Brown University, while Hanlin Tang, Mosaic ML’s other co-founder and current chief technology officer, has a doctorate in biophysics from Harvard University. By comparison, six of Databricks’ seven co-founders have their PhDs. Ghodsi gained his doctorate in computer science from KTH Royal Institute of Technology and is an adjunct professor at University of California, Berkeley where he teaches computer science. “We’re all scientists,” Rao said simply.
On May 10, Ghodsi spent about 30 minutes speaking to Rao, laying out why MosaicML should join Databricks. He emphasized that the two companies share many of the same customers,that Databricks had done much of the hard work building a sales process and getting the enterprises’ trust, Rao said. “That’s really hard to do. And so no amount of money necessarily gets you that trust,” he said. When Ghodsi finished his pitch, he waited for Rao’s rejection.
Only that didn’t happen.
According to Ghodsi, Rao told him that other companies were interested in buying MosaicML and he’s “been a no to everyone” but Databricks is “the only company it would make sense to do this with. The only question is, can we agree on price?”
Ghodsi was surprised by Rao’s response. He immediately called three of his co-founders, who were excited and wanted to move forward, Ghodsi said.
The next three weeks were spent in negotiations. There were many calls and meetings; neither company used a banker. Ghodsi said his starting point to buy MosaicML was $500 million and he was prepared to walk away at $800 million.
Rao, who has sold a company before, showed his skill as a negotiator. The MosaicML CEO would often tell Ghodsi he loved Databricks but others in his camp, possibly the board and investors, were not as enthused, Ghodsi said. Rao pointed to MosaicML’s valuation. With all the hype around AI, the startup could easily receive a valuation of more than $1 billion if they sought outside funding, Rao said. MosaicML had also received offers from some big publicly traded tech companies, but rejected them, he said. He also impressed upon Databricks the importance of other transformative mergers like Facebook’s $1 billion buy of Instagram in 2012 or the $19 billion it spent on WhatsApp in 2014. Both those acquisitions fueled growth for Facebook over time. “This isn’t an aqua hire. This is a strategic merger…We’re going to work together, and we’re going to go do some amazing things. MosaicML is their WhatsApp or their Instagram,” Rao said.
Rao’s persuasive arguments spurred Databricks to do more diligence. “There were many, many rounds, and I had to figure out a way to get comfortable with this much higher price,” Ghodsi said.
Rao also insisted that the MosaicML transaction be a stock deal because it “puts us all in the same boat.” He wanted there to be strong alignment for his team. None of MosaicML’s VC investors are cashing out and all 62 of MosaicML’s employees were expected to stay. “It’s not just about Mosaic. It’s about Mosaic plus Databricks. It’s about making Databricks the best enterprise AI company in the world,” Rao said.
Near the end of May, Ghodsi traveled to the Bahamas for a family vacation with his wife and two sons. The CEO spent most of the trip consumed by the MosaicML acquisition. He was in constant discussions and meetings with Databricks lawyers, their venture bankers, Rao and then back again. “I have not seen the Bahamas, but my family has,” Ghodsi said.
When Ghodsi flew back to San Francisco on May 29, he signed a Term Sheet for Databricks to buy MosaicML. The price was $1.3 billion.
The companies spent the next several weeks in diligence. On June 26, Databrick’s announced its agreement to buy MosaicML and the transaction closed this week. Orrick, a law firm, provided legal guidance to MosaicML while Fenwick & West represented Databricks.
MosaicML’s 62 employees have joined Databricks, which generated more than $1 billion in revenue for the fiscal year ending in January 2023. None of MosaicML’s VCs are cashing out with the deal. Shahin Farshchi, a general partner at Lux, one of MosaicML’s venture investors, said Rao looks out for his employees and investors. “Naveen is extremely savvy, and he’s extremely fair,” Farshchi said.
MosaicML now operates as a self-contained group within Databricks and is staying in its San Francisco offices, which is down the street from Databrick’s headquarters. Rao is staying put. “I’m here to make this company amazing, whatever that takes to do it,” he said.
The MosaicML founder doesn’t have a title with the combined company. Rao said he prefers to have roles that are a little bit more amorphous, because it allows him to do many things, including “make something really amazing.”
Ghodsi said he doesn’t know what Rao’s title is. “He reports to me,” the CEO said.
