Back to Home

January 2024

From: Brian and Tobias

Subject: A brief Databricks story

A couple years ago, Databricks was one of the most exciting pre-IPO companies in startup land, but they are now jockeying for the biggest IPO ever because of their position in the Gen AI landscape. The company generates over $1.5B in annual revenue and was last valued at north of $40B. In our view, Databricks is the only private company with enough scale and mindshare to be an “AI Gorilla,” playing alongside the hyperscalers, Meta, NVIDIA, and OpenAI, with the likes of Anthropic, Snowflake, and Apple vying for a place in the jungle. 

The Databricks acquisition of MosaicML may go down as the best pre-IPO acquisition in startup history. This is the case because it propels Databricks into market dominance in AIOps, completing its end-to-end solution. While this acquisition no doubt makes Databricks stronger, we think it will also create opportunities for startups, as Mosaic becomes subsumed into the Databricks platform and therefore much more enterprise-focused. Let’s dive in.

‍

Databricks: how did they end up here?

In 2013, when Databricks raised its Series A from a16z, the story was straightforward, as explained in its Series A deck:

‍

The reason why Databricks was able to raise a $14M Series A (considered huge at the time) was because of the ubiquity, growth, and power of Spark, as well as the caliber of the team. 

Apache Spark is an open-source, distributed processing engine that enables the execution of machine learning, data science, and analytics workloads; Spark helps ML engineers direct data-intensive jobs to many computers at once to increase the speed of training and inference. The beauty of Spark is that it is much more scalable than predecessors, specifically Hadoop, because of novel technology related to parallelizing data tasks and more efficient use of compute resources. As a result, Spark took off in the data community, including at the most tech-savvy companies in the world. 

As with many open source projects however, the solution on its own was hard to use and deploy. The developer experience was pretty terrible. In came a new company, Databricks, to make Spark easier to use by layering on a GUI, notebooks, and other features, including making Spark compatible with languages like SQL and Python. Additionally, Databricks’ Photon product made Spark much less expensive. Spark was already a widely adopted open source project, but Databricks came in and not only made it easier to interact with, but cheaper as well. Databricks could drive three-fold reductions in Spark-related spend. The Databricks team was filled with researchers from the AMPLab who contributed to the development of Spark. 

In 2013 when Databricks raised its Series A, the positioning was “big data,” not AI. However, data, analytics, and AI are all intertwined. Analytics and AI are ways of driving business results from data. AI is more sophisticated than analytics and can help identify deeper insights, but they are both useless without data as the fodder. As Databricks scaled and became the de facto platform for data scientists to perform complex data analysis, it also became the de facto place for ML engineers to collaborate with data scientists to utilize machine learning, specifically at larger organizations.

The current articulation of this new vision for Databricks is what they’ve coined the “Data Lakehouse,” which combines the scale of a data lake with the structure of a data warehouse. In the words of Databricks, the Data Lakehouse is “one architecture for integration, storage, processing, governance, sharing, analytics, and AI.” 

It is important to note that this analysis intentionally leaves out Snowflake, which is in a battle with Databricks to own the analytics and AI data warehouse market. For a great primer on the Snowflake-Databricks contest and the surrounding context around this battle, check out Shomik Ghosh’s podcast on it here (we are big fans and listeners). Lots of companies are fighting for pole position to offer AI services to enterprises. Databricks is one of a few in prime position, but not the only one.

‍

Databricks in 2024: an AI company

In the wake of all of this, Databricks has taken advantage of its position in the AI market to make a fast identity pivot from a data-centric company to an AI-centric one.

Just compare its current landing page…

‍

To its landing page just a year ago:

‍

The 2023 version does have some AI language in it, but it's not the obvious focus. In 2024, Databricks has fully repositioned around AI. It even released Dolly, its own open source LLM. The company’s announcements at its 2023 Data+AI Summit also show this evolution, including:

‍

The MosaicML acquisition and conflicting customer bases

The most significant and discussed development however, has been Databricks’ acquisition of MosaicML, a toolkit for training internal LLMs, for a staggering $1.3B . When this acquisition happened, MosaicML went from 0 to $30m in revenue in just six months, capitalizing on the almost overnight desire for companies to easily train and deploy their own models. 

In our view, the acquisition of MosaicML completes Databricks’ enterprise LLM capabilities. For Databricks, Mosaic solves the problem of fine-tuning and training models for companies that don’t want to use API services (e.g. OpenAI), which are easy to use but lack flexibility and control. This is the preference of larger companies and typical Databricks customers, so Mosaic is a strategic product to add to the Databricks suite. The end result should work harmoniously – train your models easily with Mosaic’s tools and then utilize and build on top of those models with Databricks. Databricks speaks to this “unification” of the AI and data stack in its announcement of the acquisition. For Databricks, the ideal state is that customers never have to leave their platform for data science, AI, or ML needs.

More interesting, though, is how Databricks and MosaicML relate in their customer bases. MosaicML’s core value proposition is making training models less expensive, faster, and less complex. The company originally targeted AI researchers but over time saw adoption with scaling tech companies. Replit is a great example. We recently talked to an employee at Replit who’s been there for a number of years – much of their AI-related product development would not have been possible without MosaicML’s platform.

MosaicML pre-Databricks acquisition had diverse customers, including scale-ups and startups. In fact, much of its rise to prominence in the early days of the Gen AI craze was tied to training LLMs at significantly lower costs than was previously believed possible, a move to “democratize” AI. In other words, it was not solely an enterprise product. This is different from Databricks, which sells to more traditional enterprises that will likely be slower to adopt AI but also have a lot of financial resources to pay for a comprehensive platform that you never have to leave, as well as trained engineers to use it.

The divide is stark among companies. The experience of a portfolio founder selling an AI solution to mid-market companies and scale-ups is that he “never hears about Databricks” in customer calls. However, when we spent a couple of months a year ago speaking to AI and ML leaders at larger companies, it seemed like they were all running Databricks shops.

The Databricks acquisition all but assures Mosaic’s focus will be redirected to larger companies. We recently spoke to a MosaicML salesperson who noted that the tools would start being offered only via the Databricks platform, rolling them into one product over time. MosaicML customers have also noted this to us, explaining that Databricks is now pushing them to run Mosaic on Databricks infra.

‍

The implications of the Databricks x MosaicML marriage

Our view is that Databricks’ acquisition of MosaicML has three main effects: (1) it strengthens Databricks’ position in the market by making it a true end-to-end platform for LLM operations, (2) it means that Mosaic’s growth into an enterprise solution will accelerate, and (3) it creates opportunities for newcomers and early-stage companies. 

Let us elaborate on that last point. We believe a gap could emerge in the market created by Databricks’ “enterprise-ation” of Mosaic’s solution. There are lots of companies that have been customers of Mosaic but not of Databricks, and it remains to be seen what happens to them. There is a strong likelihood that these companies become deprioritized for Mosaic as they’re increasingly pushed to conform to the Databricks platform. If this happens, there will be a gap for companies that would have been Mosaic customers in the past but won’t be Databricks customers in the near future. Who fills this gap remains to be seen. Together is a leading early candidate, and we’ve heard from customers that Together is superior to MosaicML in a few key areas such as speed and cost. Other contenders include Modal and Baseten, but there are many more.

We also think there will be opportunities at the even lower end of the market. There are many reasons to want to fine-tune an LLM instead of using an API service. We’ve written about the “POC chasm” in the past and think a lot of the issues preventing companies from productionalizing AI apps stem from the shortcomings of API services. As small companies start wanting to build apps using LLMs, they’ll be left trying to figure out how to do it beyond OpenAI. Where Mosaic may have once been able to help here, they’re unlikely to do so going forward. We’ll be tracking how startups build apps with LLMs using their own data. The answer almost definitely won’t be OpenAI, Mosaic, or internal resources—it will need to be something else. A short-term solution will be fancy prompt engineering and RAG, but that may not scale over time.

And even more broadly, the Databricks acquisition of MosaicML signals an important sign of consolidation in the AI stack for the enterprise customer and use case. However, lots of other companies will struggle to productionalize AI in the years to come and need solutions that serve their needs at a palatable price point and with a degree of flexibility that startups require but enterprises don’t. For large companies that have made architecture decisions at scale, end-to-end platforms work because the tech stack is unlikely to undergo significant or fast change. This is not true for faster growing companies, and they therefore require solutions that are not just less expensive but also more malleable.

As a result, “AI for the 99%” (credit to portfolio company Cake AI for this language) – where that 99% requires not just less expensive options but also more flexible ones – still feels very much up for the taking, especially if Databricks starts to subsume best-in-class AI point solutions via acquisition.

‍

It’s an exciting time to be investing in AI and data products, and we think this is especially true for companies building tooling for overlooked customers of AI (namely non-Databricks customers). We’re continuing to look for companies that fall into this bucket with exceptional founding teams. As always, thoughts and feedback are always welcome.

‍

Until next time,

Brian & Tobias

‍