Nikita Shamgunov has spent his career making databases faster and cheaper. He co-founded SingleStore, then Neon — a Postgres platform that was acquired by Databricks in 2025, where it now ships as Lakebase. He sat down with The Infinite Loop to talk about what agents are doing to infrastructure, why SaaS is on a conveyor belt heading toward an ax, and what he thinks comes after software.
TIL: As AI systems become more agentic, the way they interact with data is changing. What is actually shifting right now?
Shamgunov: When you build software or services, you always ask: who is the user? That's a very profound product question. And a new user has arrived: the agent. There are agents that build things — think Replit, Lovable — and agents that do work. Everyone is now talking about Claude's computer use capability. People ran out of Mac Minis because they were installing and using it to accomplish tasks autonomously.
In the infrastructure world, this new user has new requirements. Those systems need to be autonomous. They need to be very fast. The volume of things they create and consume is jumping ten to a hundred times. And so things that are expensive need to become cheaper, or at least more granular. Things need to be more elastic. And then there are entirely new requirements around security and governance. If work is being done not by a human but by a machine, who's responsible? That machine is making intelligent calls, performing actions either on your behalf or completely autonomously. And if it does something it wasn't supposed to do, who's at fault and how do we catch it? That's arriving at us at cosmic speed.
What else is the agentic era changing right now?
I'm a software developer at heart, and I've watched how our teams' approach to building software changed completely compared to a year ago. Nobody types code anymore. Pretty much all the code being produced is generated by agents. GitHub commits went up ten times between 2025 and 2026. For a startup, ten times is not a big deal. But GitHub is a platform with a billion users, the world's repository for code. The amount of changes going up ten times for something at that scale is remarkable. We're seeing remarkable productivity gains, and remarkably more people participating in the creation of software. That's the rise of the citizen developer.
From my personal lens as the founder of Neon: a year ago, up to 80% of traffic was already being driven by agents, up from around 10% the year before. Right now, I think it's 99.9%. We built our platform so that agents can not only consume infrastructure, but keep calling into the platform as they build new software. If you count that consumption as well, you get to 99%.
It was a transformational year for building new applications, too. If you need a new app or website today, it's so much easier to do.
I was on a panel with the chief AI officer at a major platform, and he said something very interesting: we might reach peak apps. We probably created ten to a hundred times the number of new applications in the past year, but many of them are so small that we may start pulling them into generic gigantic platforms like ChatGPT. The definition of what an app is may transform. We might have something completely ephemeral — existing only temporarily while you're talking to AI, created on demand, used on demand. Maybe you pin it so you can come back to it. But otherwise, it doesn't feel like a standalone app anymore.
Nikita Shamgunov, VP at Databricks
Human attention is such that you can't keep track of that many. There are super apps we all use every day, but you don't have space for hundreds of thousands of them. And we're churning out new ones constantly. Something needs to happen.
What does separating compute from storage actually unlock?
The idea of separating compute and storage is not new, but it's very relevant for anything cloud. In the cloud, you pay for everything. If you couple compute and storage and your workload doesn't need that ratio — say, every terabyte of storage comes with 24 CPUs, but your workload only needs eight — then 16 CPUs are running idle, and you're paying for them.
Separating them allows you to build a more efficient architecture with higher utilization. From there, you start pulling the thread:multi-tenancy, the ability to split infrastructure into small pieces and give each to a user, then expand and contract it based on actual consumption. If you don't do this, you pass the cost to the user. In a competitive market, that's dangerous. If someone delivers the same service ten times cheaper, they win.
As agents increasingly create and interact with databases autonomously, from a security angle, what keeps you up at night?
A couple of things. Anthropic released a model called Mythos that is just incredible at finding vulnerabilities. It creates very interesting dilemmas for people building open-source software, because now anyone can run a security scan and find bugs. Any software will have vulnerabilities. The question is what the time is between discovering a vulnerability and someone exploiting it. And that time is shrinking. Neon depends on Postgres. Postgres is an incredibly popular open-source database that has existed for many years. There are vulnerabilities in Postgres, and we're discovering them at an increasing rate. The time between discovery and malicious application is shrinking fast. That keeps me up at night.
The second thing is reliability at the new scale. We're launching eight million instances a day. That's not what AWS RDS or Aurora is experiencing. They have a lot of revenue and large instances, but not this insane volume of things going up and down all the time. Spinning up, shutting down, spinning up again — that's a new dimension. We actually ran out of IP addresses in a region once. AWS was not able to supply enough because of the consumption volume from agents. I think we solved that particular problem. But the security one is very real, not just for us, but for anyone building internet services.
We’re witnessing what many call the SaaS apocalypse. Does that threat extend to Databricks?
Imagine a conveyor belt, and an ax chopping companies as the belt moves. The ax is AI. If you go on Twitter and search "SaaS market cap," you'll find an image showing the percentage of market cap taken out from those companies.
At Databricks, we ask ourselves: how do we participate in the SaaS apocalypse? But at the same time, maybe we're on that conveyor belt too — just farther out. That's why we need to be extremely paranoid.
Step one: we participate in the SaaS apocalypse by being the infrastructure for whoever is doing the disrupting. Data is much harder to disintermediate than SaaS — we're the system of record for analytical data. But AI is getting exponentially smarter, so maybe even that is possible eventually.
Step two: we have no choice but to move faster than our competitors. We build internal tools that allow us to drive agentic software development, standardize our code repository and run the development loop automatically. We also have platform advantages — the more a customer uses, the harder it becomes to switch. Does that mean we're off the conveyor belt? No. We've just bought ourselves a bit more time.
Is the industry ready for the volume of compute demand that's coming?
Ready or not, you have to do it. Salesforce just announced headless Salesforce — everything is becoming a tool for agents. Salesforce charges on a per-seat basis, and with agents, that model is going away. You need consumption-based pricing. Companies will burn a lot of tokens but not have a lot of headcount. Every SaaS company now faces the same dilemma: do you package yourself to be a tool for AI, or do you fight it?
I'm very bullish on compute demand continuing for a very long time. Here's a human argument for it. If you can hire a colleague who is equal in every way but ten IQ points smarter, you'll choose the smarter one, and you'll pay a premium. If smart starts at 130 and genius starts at 150, I want an army of geniuses in a data center. Intelligence commands a premium, and we will pay for it. Robotics will come, world models will come — all white collar work will be automated before blue collar, because the physical world is much harder. But it will arrive there too.
What did selling Neon cost you?
You're no longer the king of your little world. But the world was very little in the grand scheme of things. Everybody at Neon made a lot of money; the founders made a ton of money.
Neon was fast-growing, so it would have been fine either way. Now I'm watching the Neon property grow inside Databricks as Lakebase, which addresses the enterprise market — a market Neon didn't have at all. Very few enterprise customers were using Neon. Now thousands are using Lakebase.
At a startup, you're raising your stakes and it either works or it doesn't. That risk went away.
What changed is the cadence. At Neon, you could make a decision on Slack in thirty minutes and move on. Now Neon is a piece in a seven-billion-dollar business. The complexity increased.
The talent profile also changed. All the top people at Neon are still here, and we injected what I call discovered talent. Startups are built on undiscovered talent — people who are slightly unusual, who don't quite fit big companies, or who are earlier in their careers. In an established industry like databases, there's a lot of discovered talent — people who just know a lot. And that started to pay off — reliability, performance, all the fundamental things dramatically improved.
It's like Tesla — people used it for a handful of things it did extraordinarily well, even though Audi and Mercedes were better in other dimensions. For Neon, that handful was developer experience. Now we can lift the fundamentals up because of the talent and resources that come with being part of Databricks.
You're also an investor. Where is your focus now?
I write angel checks — fifty-thousand-dollar checks into early-stage startups. My thesis over the last two or three years was software and all the inputs into AI: Lovable, Replit, Modal. But every few years the focus changes.
Right now, I think the fog of war is lifting a little. We know that we're either consumerizing the creation of software or we're doing professional software development: Cursor, Cognition, Claude Code. Those professional systems are at peak market fit right now. Cognition is exploding. Factory AI just raised a hundred and fifty million from Khosla. Cursor is at two billion plus in revenue run rate. Anthropic will try to move up the stack and attack that space because people are making too much money on top of their model.
At seed stage today, it's a little unclear to me where the opportunity is. Maybe the swim lanes are already predefined. If I had an incremental dollar right now, I would give it to Replit, Lovable, Cognition, Cursor, and Anthropic.
You've watched AI move from training to inference to agents. What comes after that?
Writing code in a programming app, I think that will go away. But code written by AI in parallel, using what I'd call AI factories, we're going to see a lot more of that. The more interesting question is what becomes transitional and what has staying power. Perplexity fascinates me. You'd think competing with Google would be the death of it — but they reinvented themselves with Perplexity Computer, hit half a billion in revenue. Whether it's transitional or not — I think the pattern it represents will stay. The bigger thing I'm watching — as a human, not just as a systems professional — is scientific discovery. AI can solve high school math, college-level math, but it's not at the level of making new scientific discoveries yet. It can assist mathematicians but it's not creating the next group theory. I think we might start seeing genuine scientific discovery in mathematics. Math is so well-defined. And from there it will move to biology, to physics. When that moment arrives, it won't just be transformational to software. It will be transformational to humanity.
The interview was edited for clarity and length.
