Hyperscalers Have the Wrong Primitives for AI | Databricks
MTS
October 30, 2026
Databricks VP and Neon co-founder Nikita Shamgunov explains why cloud hyperscalers offer the wrong primitives for AI agents and how Lakebase solves database state at scale.
Nikita Shamgunov
The set of primitives that you can get off the shelf in hyperscalers are the wrong set of primitives that are built for the agentic use case. Next Generation software is actually agentic, right? It's like new apps, they all call into AI, and that's what makes them agents. They become harnesses.
The set of primitives that these applications need are different. And that's why we hear about sandboxes. That's why we hear about agentic memory. That's why we hear about workflow. That's why we hear about agentic databases, and then the architecture of Lakebase. And so everything, the infrastructure is reshaping itself for the agentic consumption.
SPEAKER_2
Hello, everyone. Welcome back to MTS. We are live with Nikita Shamgunov, who is the co-founder of Neon and now a VP at Databricks, where he is building Lakebase, a Postgres database for AI apps and agents. So Nikita, welcome to MTS.
Nikita Shamgunov
Excited to be here.
SPEAKER_2
Excited to have you. I think one of the big things I was excited to talk about, you had this viral post that you posted last week saying, can a single database architecture satisfy both vibe coded and large scale enterprise apps? Tell me a little bit more about that. What's the answer here?
Nikita Shamgunov
Well, the answer is Lakebase. Lakebase is the architecture that we put together, and we have a product that built with this architecture, we call it Lakebase Postgres. It's all part of the developer platform that is still Neon.
And what makes Lakebase architecture is so special are like the two key features that we introduced to the market pretty much from the beginning. And those features we built for developers, but then agents picked them up. And the first part of this architecture is that everything at the low end needs to be cheap.
That's why Lakebase kind of requires whatever that is Postgres, or we can actually apply Lakebase architecture to other stateful products. But in the specific case of Lakebase Postgres, you start at the low end, because agents are completely voracious. Like we have 150,000 new databases started on the platform, and then we have 15 million starts and stops. Like there's a lot of traffic that is passing through the system.
And if that either takes a long time or costs a lot of money, the cost equation doesn't make sense, and this whole like, agentic experimentation just doesn't work. So being very, very inexpensive, cost efficient at the low end is one of the requirements. The second thing that we did with the Lakebase architecture is like, okay, well, we'll allow the creation of a branch or like a sandbox for the database, database sandbox where an agent can say, hey, this super important workload runs in production, but we just want to play around with it. So I'm going to create a branch and then live in this branch, and whatever the results of my work, I will either push them into production or discard.
And so the database branches was invented by the Neon team back in the days. And now, you know, when we think about this architecture in a category, actually, like we think, we think it's actually so big that it will become a category of stateful products.
And it will like take over the world. And we're getting some reinforcement actually of that, because like not everybody calls it Lakebase, some people do. But they certainly all built around very, very similar principles of how we built Lakebase Postgres. So most recently, and I tweeted about this, Microsoft picked up that architecture with their product HorizonDB. And yeah, there's some other technology that entered the market recently that also picked up this Lakebase architecture. Now, the question stays of, well, are you making any sacrifices? If you're making something really cheap at the low end, you're running micro VMs. What's the tradeoff? Like, what's the tradeoff? Like if you, and then you like put, I don't know, you know, airline reservation or something like really, really classic, what it's called OLTP, online transactional processing workload on the system. Is this going to work at scale?
And my argument, and I also speak about it, on the Tweet, and the reason that it will is we're very, very careful in the places where we're like innovate and change everything. And in the places where just bring all the existing world, right? So Lakebase Postgres is still Postgres. And with that comes a lot of the performance characteristics of Postgres itself. So it's not necessarily worse.
But we also made it strictly better. So we used some of the kind of like important trade-offs of where to put load, right? Like Postgres is a monolith. So all the load goes into this one system. But if you layer the system, and what we did is we separated storage and compute, and we built our storage around an object store, we added branches to our storage. But because our storage is scalable and elastic, and we run it at a very large scale now, with all the traffic coming in, we can actually choose where we burn CPU and how we move data around each one of those tiers, and which tiers are responsible for caching of what data. And the fact that we have choice allows us to move bottlenecks around. So one of the things that we did recently, like databases are like in a grand scheme of things, they support reason rights.
Read the other 7 minutes of this episode here.
$15 gets you this episode and 99 more. Credits never expire. After checkout, one link brings you back here to read it.
Not what you expected? Full refund within 14 days.
Thousands of transcripts fetched by people building searchable podcast archives
Fetch transcripts with code
The same credits work from a script or an AI agent. The demo key returns a sample episode in full, no card needed:
curl -H "x-api-key: pt_demo" \
https://spoken.md/transcripts/1000651996090Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.
From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.
Using your own key:
curl -H "x-api-key: YOUR_KEY" \
https://spoken.md/transcripts/1000798823887