Why Anthropic, Meta, and Tesla All Chose the Same Database | Aaron Katz, ClickHouse

ELI5/TLDR

Aaron Katz took a fast Russian database called ClickHouse, plucked it out of Yandex, moved the engineers to Amsterdam, and built a cloud service around it. Five years later it is the database under the hood of Anthropic, Meta, Tesla, OpenAI, Vercel, and a long list of others when they need to chew through huge piles of analytical data quickly. Along the way he raised the largest pre-seed round in enterprise software history, wired $100 million out of Silicon Valley Bank thirty minutes before it collapsed, and built a sales motion that refuses to do golf dinners. The conversation is half origin story, half a working theory of where infrastructure goes when AI agents start picking the tools.

The Full Story

A database with a head start

ClickHouse was written in 2009 by Alexey Milovidov inside Yandex, the Russian search engine. The company was drowning in clickstream data — petabytes of web-analytics events — and nothing on the market could ingest and query it fast enough. So Alexey built one. The name is short for “clickstream data warehouse.” Yandex open-sourced it in 2016, and within a few years it was quietly winning workloads at Uber (logging), eBay (metrics), Disney, Comcast, Deutsche Bank. Microsoft says some of its largest analytical workloads, including Titan and Clarity, run on it.

By the time Katz got the call in early 2021, the database was already popular without a company behind it. That is the unusual starting position. Most startups have to invent both the product and the demand. ClickHouse Inc. inherited a feature-rich engine, the core team of committers, and thousands of companies already using it for free.

The pre-seed that wasn’t a pre-seed

Katz raised $50 million before there was a product, customers, or revenue — possibly the largest pre-seed in enterprise software history. He had no pitch deck. The thesis was: take a popular open-source database, hire the people who actually write the code, and build a managed cloud version. A $250 million Series A from Coatue and Altimeter followed almost immediately. This was peak ZIRP, when capital was looking for a home, but the underlying logic was sound:

“We had this very feature-rich database, and we had the core team of committers. As you know, in open source, you can have hundreds or thousands of contributors, but to really control the road map and the direction of the project, you really need to employ the committers.”

Plus 50 design partners — companies already running the open-source version — telling him exactly what a managed service would have to look like for them to switch.

Getting the IP out of Russia

Forming the company took nine months because Yandex was a public company with a $30 billion market cap, and you cannot just walk intellectual property out of a public company without lawyers. Once the deal closed, Katz had a harder problem. A condition of the formation was that the company had no ties to Russia — no engineers on Russian networks, no offices in Moscow.

So he convinced Alexey and the core engineering team to relocate from Moscow to Amsterdam. They moved in January 2022. Russia invaded Ukraine in February, six weeks later. The company spent the next year untangling itself from Yandex while the engineers, now in the Netherlands, watched their home country become unrecognizable. Katz is careful when he talks about this, but the texture is clear — celebrating the heritage of the software while sympathizing with engineers who have not been home since.

The $100 million wire

A year later came the SVB collapse. ClickHouse had $300 million sitting at Silicon Valley Bank. Katz was reasonably sure $200 million was in custody at US Bank and therefore safe, but $100 million was on SVB’s balance sheet. On the Thursday that everything started to wobble, nobody at SVB could give him a clear answer.

“I called my head of Europe, Arnaud, and I said, ‘Wake up the bankers at our bank in the Netherlands cuz there’s a $100 million wire that’s coming through in the next 30 minutes, and it needs to clear.’”

The wire cleared. Three minutes later SVB’s banking system went down and other founders started wiring themselves money from their own companies, which, as he puts it, “is not a great look.” The government eventually backstopped SVB on Sunday, but for 72 hours nobody could confirm whether the remaining $200 million was safe.

The Datadog model versus the Snowflake model

When Katz designed the go-to-market, he studied two reference companies. Snowflake won by hiring an enormous enterprise sales force — golf, dinners, six-month evaluations, the whole machine. It worked, but it took years and a lot of capital. Datadog won by building a service a developer could sign up for, deploy, and push to production without ever speaking to a salesperson. If they did talk to anyone, it would be another engineer.

He picked the Datadog model. The principle is that nobody enjoys being sold to.

“People like buying things, but nobody likes being sold to.”

So ClickHouse Cloud has free trials, Slack channels with the actual engineers (including Alexey and the committers), and zero migration fees. If you are leaving Postgres or Snowflake or BigQuery, ClickHouse helps you migrate for free. The bet is that the friction-free path closes deals faster than the high-touch path, especially with technical buyers. Lucas Bewald (the host, CEO of Weights & Biases) confirms it from the customer side: when his team picked ClickHouse, the experience was “fixing our problems, getting us up and running, waiting to charge us until we were up and running” — no pomp, no golf, no dinners.

Why this matters now: agents pick the tools

The most interesting argument in the conversation is about what happens when AI agents become the buyers. Anthropic asked Claude what database to use for a particular workload, and Claude suggested ClickHouse. The CEO of one of Europe’s biggest fintechs told Katz that “every LLM I ask what I should be using to re-platform our company suggests ClickHouse.” Today this still requires a human to ask the question. Soon, Katz argues, agents will be selecting and provisioning the infrastructure themselves — not just recommending ClickHouse, but spinning up the cloud service and stitching it into the application.

If that future arrives, it has uneven consequences. Companies whose value lives in a polished human interface — his example is Datadog — face a strange new world where nobody is logging in to admire the UI. The agents are. And agents do not care about pretty dashboards. They care about price, performance, and a clean API. Katz expects developers to start using ClickHouse directly as their observability layer, building custom interfaces for their specific application, instead of buying Datadog’s pre-packaged one. He is careful to say Datadog is a formidable, premium product and will survive. But the multiples will not come back.

This is also why ClickHouse is now building a managed Postgres service alongside its analytical engine. The pitch is a unified data stack — analytical workloads on ClickHouse, transactional workloads on Postgres — that an agent can provision in one shot.

”Ate our own dog food, then mutiny”

A small story that reveals the culture. Early on, ClickHouse used Datadog internally for monitoring its own service. It grew, as it does, until Katz was spending seven figures on Datadog before the company had any revenue. Eventually he decided they had to migrate off — partly cost, partly the principle that an observability vendor that does not use its own product cannot be taken seriously.

“It was almost like we had a mutiny in the company. Like our developers — you can’t take Datadog. It’s like I was taking heroin away from a drug addict.”

They eventually got off Datadog and published a blog post on the cost savings and performance gains. The discomfort of the migration is the whole point of the story — even ClickHouse’s own engineers did not want to leave Datadog. That is what makes Datadog formidable, and also what makes Katz so confident that the only thing that displaces it is a fundamental change in who the buyer is.

Hiring through AI

ClickHouse is 500 people today and plans to be 1,000 by year-end. AI is not shrinking the headcount plan; if anything it is accelerating it, because each engineer becomes 10x more productive. But AI is removing categories of jobs Katz never wanted in the first place — SDRs, CSMs — which he sees as roles that “mask other issues around product quality and sales efficacy.” Lead qualification, demand generation, contextual help in the product: all good candidates for agents. Engineering is more conservative. Roughly 50% of code commits today are AI-assisted; he expects 80% within six months. But blind agent commits to a database codebase remain off the table.

“We’re building a database. We’re not building a mobile app over the weekend, you know, a dating app. Move fast and break things doesn’t apply to a database.”

The competitive map

ClickHouse competes on a wide front because it is, in Katz’s framing, a platform rather than a single product. In data warehousing: Snowflake, BigQuery, Redshift, and increasingly Databricks (a partner with growing overlap). In observability: Datadog, Grafana, Elastic. In real-time analytics — the original use case — he claims ClickHouse is the market leader, with no strong alternative for the customer-facing B2B SaaS workloads at companies like Vercel, Lovable, Sierra, Decagon, Ramp, Klaviyo, Attentive, LangChain, and Weights & Biases itself.

Key Takeaways

ClickHouse was written in 2009 inside Yandex by Alexey Milovidov to handle petabytes of clickstream data; the name literally stands for “clickstream data warehouse.”
It was open-sourced in 2016 and quietly accumulated production users (Uber, eBay, Microsoft, Disney, Comcast, Deutsche Bank) before any company existed to sell it.
ClickHouse Inc. raised a $50M pre-seed in August 2021 with no product, no customers, no revenue, no pitch deck — likely the largest pre-seed in enterprise software history. A $250M Series A from Coatue and Altimeter followed.
Extracting the IP from Yandex (a public company with a $30B market cap) took 9 months of regulatory work.
Katz required that the new entity have zero Russia ties; he relocated the core engineers from Moscow to Amsterdam in January 2022. Russia invaded Ukraine six weeks later.
During the SVB collapse, ClickHouse wired $100M out of SVB three minutes before its banking system froze. Another $200M was custodied at US Bank but unconfirmed for 72 hours.
The go-to-market deliberately copies Datadog (frictionless self-serve, peer-to-peer engineering support, no migration fees) and rejects the Snowflake model (heavy enterprise sales).
Katz refuses to hire SDRs or CSMs, viewing them as roles that mask product or sales-execution problems.
LLMs already recommend ClickHouse by default when asked what to use for analytical workloads — Anthropic, Claude, “every LLM” cited by a European fintech CEO.
Katz’s framework: when AI agents become the buyer, vendors whose value sits in a polished human UI (his example: Datadog) face a structural shift, because agents do not log in to admire dashboards.
ClickHouse is building a managed Postgres service to sit alongside its analytical engine — public beta in months, GA by end of 2026 — pitched as the world’s fastest Postgres in the cloud.
Roughly 50% of ClickHouse’s code commits are AI-assisted today; he expects 80% within 6 months. Database codebases still require human review of agent-generated code.
Headcount target: 500 → 1,000 employees in 2026, despite AI productivity gains.
The infrastructure layer is explicitly framed as “picks and shovels to the gold rush” of AI.
SaaS multiples are not coming back to prior peaks; companies are now valued on free-cash-flow dynamics and terminal value, not on 30-50% YoY growth assumptions.
ClickHouse’s design partners early on were 50 companies already running the open-source version, which is how the cloud service avoided the usual cold-start problem.

Claude’s Take

This is a strong founder interview, and the fermentation is mostly Aaron’s, not the host’s. Lucas Bewald asks good questions and largely lets Katz talk, which works because Katz is articulate and has lived through enough cycles to be useful.

The most valuable idea is the agents-as-buyers thesis. It is a clean reframing of the SaaS slowdown debate. The argument is not “SaaS is dead” — it is that vendors whose moat is a beautiful UI for humans are about to discover that humans were the moat, and agents will not pay UI premiums. ClickHouse-style infrastructure that wins on price and performance benchmarks gets stronger; Datadog-style premium-product positioning gets squeezed even if the underlying tech is excellent. Whether or not you accept the prediction, it is a useful lens.

What is less examined: the entire interview assumes ClickHouse will keep winning on price/performance because it has the committers and the open-source distribution. But databases are unusually durable competitive moats — Postgres, MySQL, and Oracle are all decades old — and the ClickHouse cloud service is now running into the same operational realities (reliability, durability, security) that made the incumbents formidable. Katz alludes to “sleepless nights” on this front but doesn’t go deeper. The SVB story and the Yandex-extraction story are great founder mythology, but they are largely about luck and timing, not strategy. Useful for color, less useful for transferable lessons.

Score 7. It is a good listen if you are interested in infrastructure go-to-market or the agents-pick-the-tools thesis. It is not essential — there is no single insight here that you cannot find in a Datadog blog post or a Mongo earnings call. But the synthesis is clean and the founder is more candid than most. The bits about SDRs being a mask for product problems, and about Datadog being heroin to its own engineers, are the kind of small honest details that earn the listening time.