MCP Dev Summit - Presented by the Agentic AI Foundation | Till Dohmen, MotherDuck

Clips
More from MCP Dev Summit - Presented by the Agentic AI Foundation

Till Döhmen

AI Lead

MotherDuck

play_circle_outline MotherDuck Brings Multiplayer Collaboration to Local DuckDB, Solving File-Sharing Limits as Seattle Startup Scales, Backed by a16z

play_circle_outline Serverless Pay-Per-Query: Granular Billing for Actual Compute Runtime and Storage Consumption

play_circle_outline In-Database AI with Local Execution: Analyzing Unstructured Data Today, Leveraging On-Device Hardware Tomorrow

Info
Transcript

Till Dohmen, MotherDuck

Till Döhmen

AI Lead MotherDuck

In this interview from the MCP Dev Summit 2026, Till Döhmen, head of AI at MotherDuck, joins theCUBE + NYSE Wired's Gemma Allen to discuss how MotherDuck is challenging the industry's obsession with scale by delivering a collaborative, serverless data platform built for the AI agent era. Döhmen traces MotherDuck's origins as a cloud layer built on top of the widely popular DuckDB, transforming a local, single-user database into a multiplayer platform for organizations. He explains why developer conferences tend to overemphasize scale — reinforcing the assumpt... Read more

explore Keep Exploring

Is MotherDuck an evolution of DuckDB, and what is the origin story of MotherDuck? add

How does MotherDuck’s serverless, consumption-based model work and why might startups choose it over other data-warehouse options; what does this trend imply for SaaS pricing (e.g., seat-based licensing) and product strategy? add

How does MotherDuck use local (client-side) as well as server-side query execution, and how can that capability be combined with AI provider integrations to analyze unstructured data and reduce cloud compute costs? add

bolt Powered by CUBE AI

Till Dohmen, MotherDuck

search

>> Welcome back to theCUBE and NYSE Wired programming. We're coming to you today from the MCP Dev Summit here in the heart of Manhattan, right in Times Square. We were walking the floor earlier, looking at some very cool booths and talking to some very cool folks in this landscape. And I came across the company, MotherDuck, who's branding, quite frankly, I kind of loved, perhaps because I have a kid who was obsessed with ducks. But joining me now is Till Döhmen, Head of AI at MotherDuck. Welcome to.

>> Thank you.

>> So let's start at the very start. MotherDuck. I'm familiar with DuckDB. I know devs loved it. It had a great story. Is this an evolution of that? Fill me in. Take me on the origin journey from MotherDuck.

>> Yeah. I think where MotherDuck originated was the idea that DuckDB is a database that people use mostly locally, kind of as an in place replacement for things like Pandas and so on. And the idea behind MotherDuck is to bring a multiplayer experience to using DuckDB basically.

>> Okay.

>> Yeah. If you use DuckDB in an organizational context, you will get the limitation that you have to deal with local DuckDB files, and people have to send them to each other and so on. And with MotherDuck, we basically solve this problem for users and do a lot more to help basically organizations deploy DuckDB or DuckDB based data warehouse internally, but also for externally facing use cases.

>> Okay. So help me profile this company a little bit. Based in Seattle. How long are you guys around? I think you've a series B. I know you've got some big enough backers. a16z is on your top table, as well as a few others. How old is this company?

>> I think we are a little bit more than three years old by now.

>> Wow.

>> Yeah.

>> Okay.

>> Yeah, we've grown quite a bit since the days where I joined. I joined two and a half years ago. We were around 20 people back then. Now we are closer to 100.

>> Okay. Wow. We love that. We love startups on theCUBE. So I had an interesting chat with some of your marketing team earlier. I have to give them a call out too. Connor and Carly, they were very, very succinct in how they described this product to me. And one thing I understood from the conversation I had with them while walking the shop floor was that this is somewhat of a contrarian problem. A lot of companies are obsessed with data at scale. What you were essentially saying is that your data problems might be smaller and easier to solve than you think. Am I correct in that?

>> Yeah.

>> Talk to me about a typical use case for this.

>> Yeah. I went to the keynote this morning and the keynote talks, two of them, one was from Uber, one from AWS. They're talking about how to use MCP at scale, for example. So if you go to a data conference, you see the same thing, like big companies talking about how to operate a data warehouse at scale. So it's very exciting for people to see and people feel almost the need to show they're successful as an engineer when they can deploy and operate these large scale systems. But they come with a lot of complexity and overheads that smaller companies actually don't need. But there's not so much awareness of that when you go to a typical developer conference, let's say, because always the scale is more in front and center. So we are trying to counterbalance that a little bit. We had organized now the second edition, for example, of the small data conference in San Francisco. And I think people are getting behind the idea more and more that actually as a smaller team, you're better off using a product that is built for your scale.

>> And when we talk scale and we talk about team size, are we talking from the perspective of enterprise workloads, siloed workloads, or are we talking about particular buyer cases in terms of enterprise size? What would be a typical scenario whereby this might actually be a problem that you didn't know could quite easily be solved by MotherDuck?

>> Yeah. I guess there are companies, for example, out there at some point, there's young startups, they want to do analytics, and then at some point they need their first data warehouse. And so they could choose MotherDuck to do that, but it could also choose other products that are more built to scale indefinitely with them, let's say. But that comes also at a cost. And so one thing we are quite particular about is the kind of serverless model of how we run queries or let users run queries on MotherDuck. So we are fully serverless, which means users only pay for the time of compute they actually use and they also only pay for the storage they actually use. So there's no need to dimension things for specific peak workloads. But yeah, you just pay for what you use essentially at the scale that is very, very granular. There's just one user doing one query per day, you only pay for that 10 seconds where this query runs as such.

>> And I think that consumption based modeling is like such an interesting point in time, Because if we think what's happening in SaaS broadly, whether or not we're going to live in the world of seats 5, 10 years from now, seat based licensing across the board, it's really TBD. Things are changing so quickly. So I think having tangible, measurable value from everything you spend, it's so, so key. In terms of the product broadly, talk to me about what's happening in the industry and where this is gearing towards. We hear a lot about inference, the difference between model weights and training models versus what's happening right now in that space, people also developing AI solutions on the edge. Where are you guys really lining up to be most impactful?

>> So the pattern I mentioned earlier with this queries running, just occasionally maybe, this pattern really gets amplified when you put agents into the loop. So we launched an MCP server for MotherDuck so users can query, don't need to write SQL queries anymore. They can connect with us through Claude or ChatGPT or any other AI agent they're using. And that enables business users to ask questions against the data warehouse where they previously needed maybe an analyst to do the analysis for them. But it also means you have done an agent that is very fast at writing queries and that is executing a much bigger amount of queries because a lot of exploratory queries, the agent has to understand the schema and how the tables connect and so on. So you have to basically account or traditionally you would have to size your data warehouse for this new scale of workloads. And with this serverless scaling model, that works better. Another thing is risk of agents doing something wrong.

>> Yeah. And we're seeing and hearing a lot about that. So it's a real, real reality, right?

>> Exactly. And there's another property of our compute model, which is per client compute isolation, which means every user that queries MotherDuck gets its own compute unit. The data layer is shared for all users, but if your agent runs a query, and it's a super inefficient query for whatever reason, it will only affect you as user. It will not affect others that are using the same, querying the same database.

>> Wow. Well, so I can certainly see a lot of value and a lot of opportunity in the roadmap for you guys, especially as we enter this agenda here where we're seeing more and more opportunities to have this hybrid solution. Agility has become so key to the business requirements across the board. So wish you guys all the best. Finish us out here. Tell me what is ahead. What does the next 12 months look like for you and the team at motherduck?

>> So you mentioned the local side of things, so I think that's a very interesting topic. We have integrations in MotherDuck with AI providers that help your users analyze unstructured data in the database. But another interesting property of MotherDuck is it executes query locally as well as on the server. So that means if you have a powerful client device, like the one you're having there, I would say you could probably process a lot of data locally. And so that's what we are in part exploiting also that comes with all the cost savings of not having to run compute in the cloud for query processing. And we do that for traditional query processing, but I could also see a future where that becomes relevant for processing AI workloads where we say, okay, there's this part in the query that needs an AI model to do something to reason about a specific piece of text in a database. And in the future, I would hope that we can also use local hardware to some extent to do that for our users. But that's likely more than 12 months out for now and more, looking at all the things that the MCP ecosystem basically can offer us because it's really an exciting, very rapidly growing product surface for us.

>> Well, it's certainly an exciting time here today and what a time in this industry as well. So wish you guys all the best and stay close to theCUBE.

>> Thank you. Thanks for having me.

>> I'm Gemma Allen. We are here at the MCP Dev Summit in New York City. This is theCube and NYSE Wired. Stay tuned.