Discover more from Supervised
Databricks and Snowflake's rivalry in AI enters a new chapter
Snowflake announced a slew of new services this morning at its summit in Las Vegas. It was overshadowed by a massive $1.3 billion acquisition by its chief rival, Databricks.
Snowflake took to the stage today in Las Vegas for its annual conference to unleash a slew of announcements for new products and enhancements aimed at keeping—and wooing—some of the largest companies in the world with their data platform.
And Databricks did what Databricks often does for its chief rival: takes the air out of the room. Databricks announced on Monday it was acquiring MosaicML, a startup focused on enabling companies to train and deploy their own custom machine learning models in a more cost effective manner, for $1.3 billion. The deal’s been pretty much done for a few weeks, as I understand it.
Snowflake and Databricks often do this dance where they converge on some areas and then diverge in others. Databricks has built out a partner ecosystem (including investing in startups), but it also invests heavily in building its own tools. Snowflake, meanwhile, continues to lean hard into its partner ecosystem despite being able to build many of the same tools. It could easily build a vector database, for example, but it’s handing that responsibility off to a partner like Pinecone.
Databricks’ announcement came the day before before Snowflake’s keynote today, where the latter announced it was launching a large number of new services aimed at enabling companies to use its compute engines in more flexible ways. One of the most important is Snowpark Container Services, which basically allows companies to spin up instances of Snowpark applications—such as the notebook tool Hex—and run them directly within Snowflake’s compute clusters.
The two seemed to be on a direct crash course in machine learning, but we’re seeing once again a bit of distance growing between the two: Databricks focusing on the end-to-end of machine learning, and Snowflake focusing on fine-tuning on customer data and deploying through Streamlit. They’re not direct competitors, but they’re also not not competitors. (In fact, many that use one also use the other.)
“Snowpark Container Services massively changes the opportunity to how we can deeply partner with pretty much anyone out there,” Kleinerman said. “Today we partner with a lot of people, but customers rightly say this is complicated. The native apps and Snowpark Containers removes a lot of that friction. We have less pressure to build things ourselves.”
The benefit of relying on partners is that you hand off a lot of heavy lifting—and go-to-market motions—to other companies assuming you have built a platform for it. But the benefit of running the tools yourselves is you’re the ultimate arbiter of what those products are and customers only have to purchase a single product.
To build or to partner (or buy)
Snowflake launched a long-expected text-to-SQL model where it can leverage its enormous corpus of queries to access data it has on Snowflake—a data set that is likely the envy of companies that don’t have the scale or haven’t set up proper accessible logging to get it. It also launched a fine-tuned model for accessing information within documents. There are some natural continuations of these products, though that’s what we’re getting for now.
They can build their own LLMs. But Snowflake also hosts most of a company’s operational data, ranging from product interaction to sales information pulled out of Salesforce. It’s priming itself as the home to fine-tune models of customer data rather than build those models from scratch via Snowflake’s compute engines.
That is the appeal of fine-tuned models (particularly those in open source)—or, specifically, models not owned by OpenAI. You don’t have to ship your data over to OpenAI to get the most out of the GPT-4 API. But at the same time, OpenAI offers a works-right-out-of-the-box solutions that has pretty much universal adoption with all the companies I talk to (and the ones they talk to, by extension).
“In the context of LLMs, the real big opportunity is fine-tuning,” Snowflake chief product officer Christian Kleinerman told me. “We are custodians of very valuable data, and all of them want better results from models—but they want to be very careful with their data. We’re not gonna compete with OpenAI on zero shot. But the real money is fine-tuning with customers own data.”
Snowflake could have built its own vector database (and potentially an embedding model to go with it), but it’s instead handed that responsibility off to Pinecone. That’s the case for now, and it could change, but that extensive partner ecosystem helps it tackle use cases it might not necessarily want to focus on as it looks to become the underlying compute mesh for company operations.
Snowflake, of course, is not afraid of buying companies to break into new areas. They acquired Streamlit for $800 million early last year to build a front-end way to interface with machine learning models and operations, which has become a key part of their machine learning stack.
But its $150 million acquisition of the LLM-powered search engine Neeva was also basically an acquihire—in large part to get its founder Sridhar Ramaswamy in the door. (Databricks, from what I understand, also evaluated acquiring Neeva at the time but Snowflake offered the highest price.)
While Snowflake wants to become the home of fine-tuning, MosaicML’s acquisition represents Databricks wanting to further own model development and serving from the very beginning of the process by enabling companies to create their own foundation models. MosaicML was best known for trying to drive the cost of training and developing models from scratch for companies as low as possible.
My understanding is that MosaicML earlier this year went out to raise a funding round at a $400 million valuation. During the fundraising process it entered discussions with Databricks for its acquisition. The price for the acquisition has been set for several weeks now, but those discussions began more than a month ago.
But one of the key value propositions of the acquisition was that MosaicML had built a highly reliable training infrastructure that avoided errors that might slow down model development—or interrupt it altogether. MosaicML chief scientist Jonathan Frankle told me earlier that its newest open source model, MPT-30B, completed training with no major convergence or loss issues.
“The key difference for us is really we’ve focused on training and tuning as a repeatable process, the reason it’s important that there’s no intervention on these runs is how you make this to a product,” MosaicML co-founder and CEO Naveen Rao said. “All the folks trying to build foundation models, the process is not the product—it’s a means to an end to build the model. It’s not repeatable, and they push it over the line with duct tape and chewing gum.”
The acquisition is a substantial outcome for some of its investors, including Lux Capital, Maverick Ventures, Future Ventures, and Playground Global. MosaicML, which came out of stealth in 2021, is one of the younger companies in the rapidly-growing AI space.
How Databricks will be integrating MosaicML into their platform remains to be seen, but it has a lot of technical chops for the small team, starting with its founding team that includes well-respected former Intel executive Rao.
Still, the math in these private-companies-acquiring-private-companies deals is always a little fuzzy. They don’t have the explicit value that a public company has, for example, in a live stock price. A $1.3 billion acquisition by a private company that includes stock doesn’t necessarily mean that the final price ends up at $1.3 billion.
Databricks and Snowflake are quickly emerging as the natural acquirers for AI-focused companies like Neeva and MosaicML. There are a range of others on the periphery (I wouldn’t be surprised if Nvidia has kicked the tires on a few machine learning startups), but Snowflake and Databricks have clearly emerged as the primary platforms for data processing and machine learning.
Databricks’ conference is happening this week in San Francisco on literally the same days as the Snowflake summit in Las Vegas. San Francisco is also seen by a lot of people as the epicenter of AI development. Meanwhile, Snowflake brought Nvidia CEO Jensen Huang on stage to do an interview with its CEO Frank Slootman. Most companies you talk to in the machine learning in big data space have representatives and executives at both conferences in some form or another.
But while Snowflake is about creating a suite of products that enable its customers to be highly flexible with the tools they bring in from partners, Databricks is leaning hard into enabling companies to build and deploy models from scratch right on the platform.
Databricks has a natural need to build an end-to-end model production and serving stack. The company has traditionally leaned hard into data science for its operations—and its branding.
Databricks has to differentiate from Snowflake and sell a different kind of story in order to eventually get a successful IPO and an exit for everyone involved. But it also needs to offer a lot of the same tools you’d get at a Snowflake—case in point its SQL business, which passed $100 million in ARR earlier this year, according to Bloomberg.
It needs a similar-but-different story to convince investors that there’s a machine learning-focused alternative to the free cash flow machine that is Snowflake, which has become a go-to data platform.
Meanwhile, Snowflake is continuing to position itself as a massive platform that anyone can bring their data in and build secure products around that data, ideally ending in an application built on top of Streamlit. It’s been trying to drop the difficulty of that, to the point that Streamlit co-founder Adrian Treuille (who now leading broader machine learning efforts at Snowflake) created a Streamlit app using an LLM on stage at the Weights & Biases conference earlier this month.
Snowpark Container Services is probably their most aggressive move in that direction thus far, tripling down on relying on—and empowering—partners to do most of the heavy lifting while focusing on the stuff that makes it possible. They announced a big partnership with Nvidia this week as part of all this to put more GPU oomph behind their products in addition to their CPU availability.
Both philosophies actually represent enormous technical lifts. Kleinerman told me Snowpark Container Services, which is in private preview starting today, was a two-year effort. Model development and serving was also a multi-year effort, in the case of MosaicML and the work Databricks has been doing.
The dance between the two continues, and it may be that both strategies are perfectly viable. But as Databricks inches closer to becoming a public company, we’ll see if one strategy is more effective than the other in machine learning.
AI is killing the old web, and the new web struggles to be born (The Verge): Filed under “who copied have ever seen this coming,” AI chat bots are producing a tsunami of content that is polluting the Web with crap ranging from auto-generated articles to capture programmatic advertising to products on e-commerce websites. Which leads us to….
Some of the World’s Largest Blue Chip Brands Unintentionally Support the Spread of Unreliable AI-Generated News Websites (NewsGuard): An analysis by NewsGuard suggests that 141 brands ran ads in programmatic networks via Google Ads across 55 AI crap-mill sites. Actually evaluating ad effectiveness was already a crap shoot and this is going to make things much more complicated. Or it’ll create new opportunities for startups to build better evaluation metrics!
Stability AI Head of Research Resigns From Startup (Bloomberg): Things continue to trend awkwardly for the commercial entity behind stable diffusion with a slew of high-level departures. Its head of research has resigned and its COO was “let go” in June this year. Its VP of product also left in March this year. This all comes after a huge investigation into the company by Forbes. It’s unclear with all these developments whether Stability AI will be able to close the round it sought at a $4 billion valuation earlier this year.
ThoughtSpot acquires Mode: Empowering data teams to bring Generative AI to BI (ThoughtSpot): In a consolidation of two of the larger business intelligence tools, Thoughtspot is acquiring Mode Analytics for $200 million. A lot of folks I spoke with pegged ThoughtSpot as the IPO candidate for the newer wave of business intelligence tools, though large language models seem to have at least changed that perception. Expect more general consolidation in this space before the dust settles and we figure out what the new new business intelligence flow will look like.
Redpanda nabs $100M Series C as streaming service experiences significant growth (TechCrunch): Streaming data pipelines are still a thing! It’s a good reminder that despite the hype around LLMs there’s still an enormous big data problem to be solved to get real-time data into production. Streaming’s been “around the corner” for years now and after seeming to come into greater focus it looks like LLMs have punted that timeline again, but there’s still plenty of activity happening in the space.
Still on my radar
Will one of the foundation model providers become an acquisition target?
How will MosaicML be integrated into Databricks?
Are larger companies beyond startups getting more access to Nvidia hardware?
How many companies are using Google’s Vertex AI in production?
Supervised is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.
If you have any tips (or answers to any of the above), please send me a note at firstname.lastname@example.org or contact me directly on Signal at +1-415-690-7086. As always, please send any and all feedback my way.
Updates and clarifications: In a previous issue detailing MongoDB’s entry into vector databases, I inadvertently allowed autocorrect (which has been thoroughly shut off) to spell MongoDB PM’s Benjamin Flast’s last name as “Flash.” That’s been corrected in the online version of the issue.