Snowflake, Databricks and the Model Makers: The Battle for the Agentic Client and AI Backend

**Dave Vellante** (0:00)
Agentic AI is being misread as a series of separate battles. For example, Snowflake versus Databricks, Copilots versus Agents, Model Makers versus App Vendors, etc. We think the real story is that the biggest opportunity in software is converging around who owns the new intelligent client and the AI backend that makes it useful. The new client, you should think of that as the agent-based system of engagement.
Snowflake's CoWork & CoCo, for example, or Databricks Genie, think Microsoft Copilot. Google's got Gemini Enterprise, ChatGPT of course has Codex, and of course Claude CoWork with its momentum, and there are of course the others. But that client, we don't think can deliver the business outcomes that people want without a new backend, what we call a System of Intelligence. Now, that represents a model of the enterprise in terms of its business rules, and very importantly, that tacit knowledge, that tribal knowledge that people always talk about.
We don't think you can build one without the other. Now, we frame this premise for this Breaking Analysis using Clay Christensen's integrated innovation concept and Jensen's extreme co-design, and we apply it to enterprise software. That is why Snowflake is the focal point for this Breaking Analysis, but not the whole story. Snowflake is not just competing with Databricks anymore. It's now in the same strategic arena as Microsoft, Google, OpenAI, Anthropic, Salesforce, SAP, ServiceNow. We'll put Celonis in there, and of course, there are others. They're all trying to define where business users, builders, and agents get work done, and where the enterprise context that powers that work gets built.
Now, the key premise again here is that the System of Intelligence, that backend, does not only ingest data from pipelines and catalogs, etc. It does that, but it also learns from business users and builders through the agentic client. Here, we're talking about skills, artifacts. You hear about semantic views. Query history is an important input. The actions that agents take, and of course, the human reasoning traces, they all become inputs that provide direct feedback into the intelligence layer. So, we think a key success factor is having the tightest feedback loop between the agentic client and the enterprise intelligence backend, and that requires deliberate software engineering that tightly couples the pieces of the stack. Once again, George Gilbert and I dig into the emerging AI software stack, and we'll connect the dots from what we learned at Snowflake Summit, and of course, Microsoft Build, which was last week, also in San Francisco, and we'll set up the Databricks Data and AI Summit that's coming in mid-June. George, once again, welcome. Thank you for your time.

**George Gilbert** (3:06)
Good to be with you, Dave.

**Dave Vellante** (3:07)
All right, let's get into it. Alex, bring up the first slide, George. You came up with this concept in previous episodes, and we shared this, that there are more agents than fleas on a camel. But you're calling out Brett Taylor on this slide, who's both the chairman of OpenAI, he's also the CEO of Sierra.
And you put a but on this slide. What's the but?

**George Gilbert** (3:32)
Well, it's that there's this theme that came out of Y Combinator, that vertical agents are going to be 10 times bigger than vertical SaaS. And this is the services software story that all the work that, or much of the work that humans can do can be captured in agents. And if we can just bottle that knowledge up, we can turn these specialized agent companies into giant companies. And that's where I'm like, to make the point, there are more of these than there are fleas on the average camel.
The problem, this is the but, is we're just building more silos, the same silos with a new technology we've been building for 60 years. And the point of agents and what agents are driving for the new data infrastructure is that we have end-to-end infrastructure and visibility across how the enterprise works so that you can achieve business outcomes, that there are not specialized silos, that you can like onboard a customer, you can do a bank can do a know-your-customer process.
And in a broader scope, you can connect planning and operations so that you can realign the activity of an entire enterprise all the way from long-term activities down to, let's say if you're a logistics company, it might be where are we going to build new fulfillment centers to down to what am I going to pick, pack, and ship, and how for this order. All those things need to be aligned. To do that alignment, you can't have silos. And that's why I object, and I think customers will find limiting all these siloed agents.