On August 3, Andy Pavlo announced he's leaving Carnegie Mellon University to join ClickHouse and establish ClickHouse Labs, its first formal research organization. If you're in the database world, you already know why this matters. If you're not — but you run systems that depend on fast analytics — it's worth a few minutes of your time, because this hire is a signal about where real-time data infrastructure is heading, and it's a louder signal than another funding announcement or a benchmark press release.
I've been running hosting infrastructure long enough to have watched databases go through several distinct eras. The late-nineties era of hand-tuned MySQL. The NoSQL decade, where every team reinvented joins by hand. The managed cloud warehouse era, when everyone got drunk on Snowflake and BigQuery and called it modern data infrastructure. And now we're in something genuinely different: the era where AI workloads have become first-class database tenants, and the analytics layer isn't an afterthought anymore — it's load-bearing infrastructure.
ClickHouse is at the center of that shift, and Pavlo just told the industry he believes it.
Who Andy Pavlo Actually Is
The short version of his CV: Associate Professor of Databaseology at CMU since 2013, NSF CAREER Award in 2019, Sloan Research Fellowship in 2018, and the ACM SIGMOD Jim Gray Dissertation Award back in 2014. This year he added the 2026 IEEE TCDE Ramez Elmasri Outstanding Database Education Award, which recognizes a decade of freely available course materials and lectures that have become the de facto database curriculum for engineers who want to actually understand what's happening underneath their ORM.
But the academic resume isn't really the point. What makes Pavlo credible is the work:
- NoisePage, the self-driving DBMS project, which used ML to forecast workloads, model database behavior, and generate tuning actions autonomously — without a DBA in the loop.
- OtterTune, the database tuning startup he co-founded in 2020, which productized the same ideas: feed query performance data into a model, let it tune knobs and indexes the way an expert would, but faster and at scale.
- BusTub, the educational DBMS used in CMU's 15-445 course, which has produced a generation of engineers who actually understand buffer pool managers, B+ trees, and concurrency control instead of just cargo-culting advice from Stack Overflow.
He's also maintained DBDB.io, a community catalog tracking database systems across industry and academia, which means he's spent years mapping the landscape with unusual rigor. He's been watching ClickHouse since its open-source release in 2016, and he's assigned ClickHouse technical articles to his CMU students. This is not a prestige hire. This is someone who has done the homework.
Where ClickHouse Actually Stands
Before we talk about what ClickHouse Labs will do, you need the context of what ClickHouse has become, because the numbers are striking even for people who've been watching the company.
In January 2026, ClickHouse closed a $400 million Series D led by Dragoneer, at a reported valuation around $15 billion, and simultaneously acquired Langfuse, the open-source LLM observability platform. As of their Open House 2026 conference, they're at $250 million ARR — tripled year-over-year — with over 4,000 customers on ClickHouse Cloud. The customer list reads like an AI industry roll call: Anthropic, Meta, Cursor, Tesla, OpenAI, Weights & Biases, Vercel, and over a dozen more showed up on stage at Open House to talk about how they're using it.
That customer base composition is not accidental, and it's the most important thing to understand about ClickHouse's current position. AI companies are their fastest-growing segment. The reason is structural: large language models and the systems built on top of them generate enormous volumes of structured operational data — inference logs, token counts, latency traces, embedding metadata, user interaction records. You need to query all of that, often at sub-second response times, at concurrency levels that would make a traditional warehouse flinch. ClickHouse was designed exactly for this: columnar storage, vectorized execution, aggressive compression, and a query engine that doesn't apologize for being built to go fast.
CEO Aaron Katz said at Open House: "AI workloads demand the performance and cost efficiency ClickHouse was built for." That's a measured understatement. What he's describing is a structural shift in what the analytics database market looks like when the primary workload changes from human analysts running morning dashboards to AI systems querying millions of rows per second to decide what to do next.
What ClickHouse Labs Will Actually Do
The stated ambition is to build something with the research impact of IBM Research or Microsoft Research — which, if you remember what those organizations produced in their primes, is a genuinely large claim. The specific directions Pavlo flagged are worth unpacking:
DBMS Design for Autonomous Agents
This is the most interesting research problem on their list, and it's one that hasn't been seriously worked in academic database research yet. Agents don't access databases the way humans or traditional software do. They issue queries dynamically, often with imprecise schemas of what they're looking for. They may need to iterate across many small queries in tight loops. They're often operating with incomplete information about what's available. They need the database to be more of a collaborator and less of a dumb record store.
Pavlo's background in self-driving databases (NoisePage) and ML-based tuning (OtterTune) puts him in a genuinely good position to think about this. A DBMS that can predict what an agent is likely to query next, pre-materialize results, adjust its execution plan based on observed agent behavior, and surface relevant data proactively — that's a fundamentally different design target than a warehouse optimized for human-authored SQL.
Accelerating Validation of Existing Engineering Work
One thing that often doesn't get enough credit when companies establish research labs: the research function isn't just about inventing new things. It's about giving ideas that engineers have already built — but that have never been formally analyzed or stress-tested — the rigorous treatment they need to become production-grade. ClickHouse has shipped an enormous amount of engineering work in the last three years. A research lab can apply formal methods, benchmarking discipline, and comparative analysis to that work in ways that a product engineering team under delivery pressure usually can't.
New Hardware, Algorithms, and Execution Strategies
The hardware landscape for analytics databases is genuinely exciting right now. CXL memory, disaggregated storage, AI accelerators that are increasingly being co-designed with software — the traditional assumptions about CPU memory bandwidth and cache hierarchy that ClickHouse's execution engine was designed around are being disrupted. Academic database research moves faster than most companies can on these questions, and having someone with Pavlo's hardware awareness on staff means ClickHouse Labs can influence the engineering roadmap before the hardware ships, not after.
The Broader Signal
I've spent time thinking about why this particular hire, at this particular moment, feels important enough to write about. Here's my honest read:
When a company is in growth mode, it hires sales. When it's in product-market-fit mode, it hires engineers. When it establishes a research lab named after itself with a mandate to produce work at the level of IBM Research, it's telling the industry that it expects to be here in twenty years and wants the intellectual infrastructure to match. That's a different kind of confidence than a funding announcement. Funding is about what you can do with money. A research lab is about what you intend to become.
The timing also matters. ClickHouse is growing fast precisely because it's at the intersection of two major structural trends: AI companies need fast analytics infrastructure, and enterprises are looking for alternatives to expensive, tightly coupled cloud data warehouses. That's a genuinely large addressable market. Establishing a research lab now, when the company has the resources to fund it seriously but the competitive dynamics are still forming, is the right move. In five years, the analytics database market is going to look meaningfully different than it does today, and ClickHouse Labs is an attempt to shape what that looks like rather than react to it.
What This Means If You're Running Analytical Infrastructure
My practical take for engineers and architects: this is a reason to take ClickHouse's roadmap more seriously if you haven't already, and it's a reason to look more carefully at the DBMS-for-agents research direction if you're building AI systems that need to access structured data at scale.
The question of how AI agents should interact with databases is genuinely unsolved, and it's going to matter a lot. If you're making architectural decisions about your analytics stack today, the companies that are asking that question seriously — and have the research firepower to explore it — are worth paying attention to.
Pavlo has been watching ClickHouse since 2016. He took a decade to decide it was worth leaving the classroom for. That's the kind of conviction that tends to produce interesting work.