Aaron Katz, co-founder and chief executive of the open-source database company ClickHouse, sat down with Harry Stebbings at Craven Cottage, the Fulham ground whose shirt his company now sponsors. They covered where the AI cycle actually is, what an agent buying infrastructure does to software design, why he limits open-weight models to code review, and the wager the two of them struck on when ClickHouse reaches a billion dollars of annual recurring revenue.
👤 Guest: Aaron Katz, co-founder and CEO of ClickHouse, previously chief revenue officer at Elastic and before that twelve years at Salesforce
🎙️ Host: Harry Stebbings, founder of 20VC and an investor in ClickHouse
📰 Published: 31 August 2026
🔴 YouTube | 🟢 Spotify | 🟣 Apple Podcasts | 🔗 Episode page | ⏱️ 1 hr 4 min | ✅ Time saved: 28 min
Key Takeaways
The belief he thinks is wrong is the bubble consensus itself
"A widely held belief around AI that's generally agreed upon but is wrong is that it's overblown and that we're in a hype cycle, that we're in a bubble."
He has been through the internet, mobile and social cycles, and says "Those cycles, in my experience, were much more gradual."
Gross margins are not the number to watch; durability of revenue is
"The switching costs can be very low for agentic applications." — model providers leapfrog each other every other week
Infrastructure software is the opposite case, which is why he will not price his own business off a margin line
His own Anthropic bill went up a hundredfold and he is not capping it
"That's not showing up in our use of Claude Code because our Anthropic spend is up 100 times from what it was at the beginning of the year."
"And the cost of tokens, as we know, is going down, not up."
The mistake he would undo is being too lean on sales, not too loose on spending
About 100 quota-carrying salespeople against competitors running thousands
"Our salespeople wake up every morning thinking about 10 different use cases."
Agents have no persona, so latency becomes the product
"And so what's the number one requirement for agentic query patterns? Low latency."
He says Tesla ingests a billion events a second into ClickHouse
Before agents can buy infrastructure they need an identity, a budget and an approver
"We can talk about the fact that agents will need to have an identity that they don't have today."
Anthropic picked ClickHouse, he said, because it asked Claude what to use
Enterprises fear open-weight models more than frontier ones, and he thinks they are right for now
"I think when you need the legal protection that most enterprises do, you're going to want to work with one of the frontier lab providers."
He splits the difference on token share three years out: "Well, the easy answer is 50-50."
The AI customer base is the smallest part of his revenue, which he treats as the point
"So we've got very little concentration risk." — anything above a tenth of revenue from one customer, category or industry gets his attention
The front of Fulham's shirt is a hospitality line item, not a branding one
"We had 20 executives last night attend an intimate Michelin-grade dinner." — half customers, half prospects
Awareness is the harder half to measure: "That's very difficult to assess."
He and Stebbings bet money on when ClickHouse hits a billion dollars of ARR
"I'd put the over under at December 2027, and I would take the under."
"I'm going to go for Feb 28. I think now that's like 18 months. That's more than 18 months." — Stebbings, taking the other side
ClickHouse could list next year and he would rather not
"I think I take the under on being a public company. We could take the company public next year if we wanted to. There's no rush."
"Like your stock can trade down 40, 50% on a slight miss in a quarter."
Craven Cottage, and Why This Cycle Does Not Feel Like the Last Three
Stebbings opened by noting where they were sitting. "Aaron, dude, I've done over a thousand shows. I've never had a setting quite like this for a show." Katz's reply: "Yeah, total joy."
Katz described the business first. ClickHouse is the world's most popular open-source database, he said, used by nearly every AI-native company including Anthropic, OpenAI and Weights and Biases, and "It's known for its lightning-fast query execution, extreme resource efficiency in terms of storing vast volumes of data."
Stebbings laid out his own contradiction before asking anything. He said he has seen this before, that "The levels of debt that we're seeing is insane. The prices and the valuations are so exuberant." — and set that against adoption and revenue scaling as two paradoxical data points
Katz's answer was three words and he repeated them at the end of the episode. "We're just getting started."
He has been through the internet, mobile and social cycles, and says this one moves differently. "Those cycles, in my experience, were much more gradual."
"This seems to be accelerating at an unprecedented pace in terms of how quickly these agentic experiences are maturing and how quickly these companies are growing."
"I mean, we haven't seen revenue growth like this in our lifetime."
Low Gross Margins Are Not the Risk. Durability of Revenue Is
Stebbings put the standard bear case to him: the hole people pick in AI revenue scaling is the gross margin profile, and he cited Fireworks, whose chief executive told the show on a previous episode the company was at 30% to 35% and hoped to improve. Does the industry live with lower margins, or scale into traditional software margins over time?
Katz said cheap capital buys these companies the room to run at margins public investors would reject. "You know, a lot of these companies just seem to have unlimited access to capital right now."
His condition is a demonstrated path to margin expansion over the next few years, alongside the growth and a healthy balance sheet
"I worry less about gross margins like we did five years ago in traditional enterprise software."
Asked what an investor should worry about instead, he named one thing. "If you were to say, what's the single biggest risk would be durability of revenue?"
The switching costs in infrastructure software are very high, and "The switching costs can be very low for agentic applications."
He pointed at the model providers leapfrogging one another, which seems to happen every other week, and said he would call the durability of some AI application revenue into question
Stebbings pressed the point at Anthropic, valued in an IPO he framed at $2 trillion, with Claude Code as the way in. Would that not sit in the low-switching-cost bucket?
Katz put it in that category and then said his own spending contradicts it: "That's not showing up in our use of Claude Code because our Anthropic spend is up 100 times from what it was at the beginning of the year."
The AI Awakening Email, and Why He Will Not Cap the Coding-Agent Bill
Asked how much ClickHouse spends with Anthropic, Katz would say only that it was a significant amount, and then explained how it got there.
He wrote to the whole company at the start of the year. "So I sent an email to the company and the subject was the AI awakening and I said, I think we're moving too slowly and I'm not seeing the adoption of these coding applications, for example, that I would expect to see for a leading database provider like ClickHouse and the company rallied to the call"
On why the frontier labs still hold an advantage: "But I think Anthropic specifically and OpenAI have the benefit of being both a model provider and the harness provider that the open weights models right now are behind in."
Stebbings asked whether independence would beat being tied to one provider, since an independent buyer can route each task to the best model. Katz said some open-weight models get used for code review but not for shipping production code
Stebbings raised the counter-example from his own show: the president of Uber, who was far more skeptical about the return the spending generated internally, and asked how Katz thinks about token budgeting when the instruction is to run free and the bill arrives at the end of the quarter
On token budgeting, he refused the premise that the bill is the thing to manage. The measure he cares about is revenue growth
"And the cost of tokens, as we know, is going down, not up."
"So I'll take that trade any day of the week."
If it ever stops working he can rein consumption back in, but the roadmap is accelerating at a pace the company has not seen
The One Thing He Would Undo: He Ran the Company Too Lean on Sales
Katz volunteered that some people think ClickHouse is too efficient, and Stebbings asked where he should have spent and didn't.
The sales organization is a fraction of what he is competing against. "So we've got about 100 quota-carrying salespeople, for example."
Large data warehousing and observability competitors field thousands of sellers, each thinking about one use case
"Our salespeople wake up every morning thinking about 10 different use cases."
"So I think the one thing, and if I look back over the last two years, that I wish I had done differently was increasing sales capacity."
He wanted the pressure on product and engineering first, and picked a playbook deliberately. He distilled the last decade of infrastructure software down to Datadog and Snowflake — Datadog self-service and developer-led, where a customer can deploy an agent and instrument an application without talking to sales; Snowflake going at the enterprise through expensive sales and marketing
"I just thought it was going to be a lot easier to follow the Datadog playbook than the Snowflake playbook, and it proved to be the case."
The caveat: at some point an enterprise sales motion has to be layered on top of the self-service distribution
What He Took From Twelve Years With Benioff
Katz joined Salesforce 24 years ago, when it was a three-year-old startup selling what he called a glorified contact manager against Act, Goldmine and the spreadsheet
The lesson, in one sentence: "That you can overestimate what you can achieve in one year and underestimate what you can achieve in five."
Marc Benioff's bigger claim came before the product could support it. He said the company was going after Siebel, SAP, Oracle and Microsoft without the product set to do it
Katz credits the marketing with creating the perception that the largest companies in the world could adopt the technology, and a roadmap that would meet their requirements over time — and said he delivered it
Six Acquisitions in Four Years, and When He Buys Instead of Builds
Stebbings cited Jason Lemkin, his co-host on a weekly show with Rory O'Driscoll, who said the week before that anyone not well into their 2027 roadmap is behind. Is development actually accelerating, and how do you measure the return when attribution is hard?
Katz said the company is shipping faster than it ever has, and "We're entering new product categories two years ahead of where we thought we would."
Six acquisitions in four years, and the most recent one is an agent play. Langfuse, out of Berlin, brought three founders and agent observability
"Every enterprise in the world is going to need this technology."
Stebbings framed the question through Nikesh Arora at Palo Alto, whom he called a good friend and a brilliant M&A machine, and asked where Katz draws the line between buying, building and being distracted
The buy-versus-build rule is about where the founders already are. If his own product and engineering teams can innovate into an area, he lets it run organically
If he sees founders building on top of ClickHouse in a category he thinks belongs in the eventual data platform, he looks at doing it inorganically
One of ClickHouse's own organic shipments is stateless workers, which he described as essentially infinite compute
Software Built for Agents Has No Persona, and Latency Becomes the Product
Four or five years ago, applications were built around a person with a role and predictable query patterns — a report, a dashboard
"While you would traditionally use a specific application for observability or data warehousing or CRM, agents expect that they're going to traverse across all these applications and they're not going to be constrained by access." The experience gets defined by the slowest point in the chain
"And so what's the number one requirement for agentic query patterns? Low latency." Agents run dozens of SQL queries simultaneously across systems, and the patterns are unpredictable and exploratory rather than reportorial
Stebbings asked whether cost-aware agents drive a race to the bottom on pricing. Katz did not accept the premise: "Well, I don't see agents necessarily being cost-efficient." He does not see them thinking about consumption and budget the way people do
Stebbings' aside: "My mother had that in common."
Volume is what breaks the pricing models. The number of queries is exploding at a rate he has not seen, forcing systems to rethink pricing, consumption and access patterns
Efficiency is second on the agent's list after latency. "I mean, Tesla, for example, is ingesting a billion events per second into ClickHouse."
"There just isn't another technology in the world other than ClickHouse that can satisfy those requirements."
Agents Will Need an Identity, a Budget and Somebody Watching the Spend
Asked what looks insane today and will be commonplace in three to five years, Katz went to the plumbing of agent purchasing.
"We can talk about the fact that agents will need to have an identity that they don't have today." They will need a budget, and a way to be authorized to consume services
The precedent is already in his customer base. "If you ask Anthropic how they chose to use ClickHouse, they'll tell you they asked Claude, what technology should we use for the specific observability use case? And Claude suggested ClickHouse."
He sketched the next step: an agent told to build an application and provision the stack underneath it — database, networking, compute, storage — making the selection itself
His advice to Stebbings as an investor was to work out which companies are best positioned to give an agent everything it needs to build a software application
Stebbings asked why governance is not simply a setting inside the harness. Katz's answer was that companies will not allow it: "Because most companies aren't just gonna let their agents run wild and build whatever they want and consume as many resources as the agent deems fit."
Today there is a human watching the consumption, monitoring agentic spend and putting controls on what data is touched and how much is spent
"If you look out at three to five years, those agents are gonna be fully autonomous."
On security and deployment, he said the answer is optionality rather than a single model. Cloud through any of the three major hyperscalers, and the ability to run on-premises behind a customer's own firewall
Regulated industries, data privacy and compliance make that flexibility a requirement rather than a feature
It is a hard roadmap to maintain, and most companies pick one lane — but forcing a customer into a lane limits the addressable market
Specialized Models Will Coexist With the Frontier Labs, and He Sells to Both Sides
Asked whether the future holds millions of specialized models, Katz said both shapes survive.
Specialized models for specific use cases such as legal technology, where he named Harvey as the leader in specialization
"But I think the large frontier labs are still gonna be the dominant providers in the space."
Stebbings, an investor in Legora, asked why Harvey's approach beats a competitor that has not committed the resources to building its own specialized models. Katz declined the comparison and answered as a supplier
"They consume less ClickHouse is the short answer to the question."
"So we're not like picking winners in these categories. Legora could be a bigger company than Harvey a year from now."
The Customer Economics: Hundreds of New Accounts a Month, Net Retention Over 200%
He counts a customer only once it is in production at a certain revenue scale, with a long tail of smaller consumers below that line
Revenue is growing slightly faster than the customer count because existing customers grow faster than new ones arrive — and new customer acquisition is the function he calls most important for a company of this size
Expansion is running ahead of anything in his experience. "We had over 130% net dollar retention at previous companies. We're over 200% because these use cases expand."
A customer starts with data warehousing, adds real-time analytics, then exposes the same repository to its own customers and partners
On average spend, he warned the average misleads — some customers spend tens of millions a year, others thousands a month
"If I were to look at the midpoint, probably around $100,000."
The Competitor He Fears Is the One That Does Not Exist Yet
"The one that isn't in the market yet." The competitors in front of him are visible and their strengths and weaknesses are known
"I worry about the next ClickHouse." He first found the technology as an open-source project ten years ago; thousands of companies adopted it while there was no company behind it, and it was dismissed as another popular open-source database
The worry, stated plainly, is what disrupts ClickHouse the way ClickHouse is disrupting the incumbents in front of it
Open Weights Are Not Open Source, and the Enterprise Fear Runs the Other Way
Stebbings said the share of tokens going through open models is rising almost exponentially and asked for the split three years out, offering the conventional wisdom that it lands at 90% open and 10% frontier, with frontier models reserved for cancer and climate change.
Katz would not go past the midpoint. "Well, the easy answer is 50-50." His analogy was the existing split between open-source and proprietary enterprise software, which he called a fairly even distribution
He worked the analogy through three pairs and lost the argument twice on purpose. Snowflake is a closed ecosystem, Databricks is built around open source — Stebbings picked Databricks. Add Datadog, primarily proprietary, against open observability tools, and Katz said he would take Datadog
On Databricks: "I don't want to underestimate Databricks. It's a great company."
He would not rule the frontier labs out of infrastructure either. "I don't see them as competition today. But you see how they're entering new categories so quickly." They are building their own chips
On the 90/10 conventional wisdom, he disagreed, and specifically in the enterprise. Buyers want provisions and protections he does not think open-weight models, especially Chinese ones, can offer — indemnification among them
He insisted on a distinction Stebbings then asked him to explain. "Now, again, it's very important that we distinguish between open weight models and open source software, okay?" ClickHouse is in the latter category
"So open source is essentially where anybody can inspect the source code, anybody can modify it depending on the license that it's governed by." It can be deployed with no attribution, modified, monetized, contributed to or forked, with licensing having evolved significantly over the last five years
Stebbings put the counter-intuitive finding from other guests to him: the largest enterprises tell him they are more afraid of the frontier providers than of open-source Chinese models, because zero data retention is a promise rather than a proof, and they worry about sending source code to a frontier lab
Katz said he sees it every day, including inside his own company. "You worry about third party indemnification."
His fix is scope rather than vendor: "You limit the use cases." Code review, yes; pushing production code into a customer environment, no
On which way he would actually go: "I think when you need the legal protection that most enterprises do, you're going to want to work with one of the frontier lab providers." He thinks the security concern around open-weight models is justified today and largely addressed a year or two from now
He would not follow Stebbings toward the espionage framing. "I personally don't like to speculate that by using some of the software, you're somehow going to be engaging in some nefarious" — and when Stebbings supplied the phrase, Katz said "It's quite fantastic to go there."
Asked what he would do to make American open models leapfrog China, he answered against intervention and against the investment. "So I don't think that any sort of government intervention is wise in terms of limiting the adoption of these technologies, the distribution of these technologies, because I do think open source and open weights models are the future."
"I personally wouldn't take your capital and say, I'm going to deploy it to develop an open weight model."
Enterprises Are Going Back On Prem, and the Sales Cycle Has Compressed to Quarters
"And even companies that I thought would never be going back on prem are talking about going back on prem." He named the most innovative digital-native companies in Silicon Valley as among those considering moving off a hyperscaler
Asked how wide the gap is between what he sees and what large enterprises understand, he answered in one word: narrow. "I mean, I was at Canary Wharf yesterday meeting with some of the largest financial services companies in the world." He said the banks are leading on adoption in a way he has not seen before
"A lot of the times, historically, the sales cycles into these big firms would be measured in years, not quarters."
Open source is part of why the cycle compressed. A buyer can start without a vendor relationship at all, and a self-service product lets them evaluate, deploy and scale before anyone in sales is involved
What Investors Get Wrong: the Moat Question and the Hyperscaler Threat
"Investors often say, like, where's the moat?" The version he hears is how hard it would be for somebody to simply redistribute ClickHouse, and what happens if a hyperscaler offers it as a managed service
He acknowledged that has happened before — AWS, Google or Microsoft taking an open-source project and reselling it as a managed service, leaving the company competing against its own core database
The defense is the cloud offering and proprietary features that are hard to replicate. "And I think very few open source companies get that right."
On whether agents choosing the stack devalues developer relations and brand, he said no, because the enterprise buyer does not go away. A weekend application can have its stack picked by an agent — a managed Postgres service for transactions, ClickHouse for analytics
A large data warehousing project at a bank or a telecommunications company is still driven by an engineer choosing between ClickHouse, Snowflake and Databricks, or an observability workload between Splunk, Datadog and open source
"For the next decade, there will still be a human involved in that decision loop."
"Not at all, because budgets are still controlled by people." Budgets, he said, correlate to technology decisions
Why a Database Company Bought the Front of Fulham's Shirt
Katz traced the sponsorship back to an awareness problem at industry conferences.
He started by buying a party. For AWS re:Invent and Google Next he partnered with the Chainsmokers, who performed in support of ClickHouse
His reasoning was that the alternative was a set of restaurant buyouts and happy hours: "I want one party that 60,000 software engineers are jumping over themselves to get access to."
The financing that paid for the sponsorship was raised for the names on it, not the money. ClickHouse did a round earlier this year and extended it to bring in strategic investors including Stebbings, David Sacks, Michael Dell and JPMorgan
"We didn't need the capital. We had a billion dollars on a balance sheet."
"I quite like the affiliation to those names."
"So I thought what better way to spend this new investor money than to sponsor an English Premier Football Club in London."
Stebbings asked why football rather than Formula One or golf. Katz set his own affection for the sport aside and gave two forms of value
Awareness: the match against Chelsea the previous night was televised globally, putting the brand in front of millions of viewers
Hospitality, which he called the easier one to quantify: "I think this is arguably the best sports experience in the world, and I've been to many."
"We had 20 executives last night attend an intimate Michelin-grade dinner." He said a C-level executive from one of the largest banks in Europe flew from Paris for it, and that those relationships matter more as ClickHouse moves upmarket
On measuring the return, he separated the two halves. Twenty guests, budget owners at some of the largest companies in the world, half customers and half prospects — "So I can very easily measure the spend that I can gather from that basket of accounts over the next 12 months."
Attribution for the awareness half is the harder problem: "That's very difficult to assess." Website visits and new trials improve, and that could be a derivative of the sponsorship
Sports Assets Keep Repricing, and He Thinks They Keep Going
Stebbings mangled the name of a team and Katz corrected him. "It's not the Seagulls, it's the Seahawks."
Katz said he did not love seeing the sale, because Khosla is an investor in ClickHouse and was a minority owner of the 49ers, and Katz is "a San Francisco 49ers lifetime faithful"
The analogy he offered Stebbings, a Fulham supporter: it would be like investing in Chelsea. Stebbings agreed he would not, "unless there was significant monetary gain, in which case I'd do it in a second"
Asked whether sports assets keep repricing upward, Katz said yes and reached for the two most recent American sales. He called the Lakers sale the most expensive sports transaction in history, and added "It was reported that the Seahawks sold for $9.6 billion."
"I mean, it's an experience you simply can't replicate through technology or anything else." He described 28,000 fans on their feet for a five-goal Southwest London derby, Chelsea against Fulham, to open the season
Stebbings agreed and added his own view that in a world of AI, live experiences and music are the two things people crave more
How He Picks Investors, and the Bet on a Billion Dollars of ARR
Stebbings said several of Katz's investors told him they wanted to put more in every round and were always cut back.
He deflected the credit to his co-founders. "The credit really goes to Yury and Alexey." He called them the best engineers he has worked with by a wide margin and described his own role as the interface to the investor community
What decides an investor relationship is the value they bring — customer introductions and advocacy
He runs references on venture firms the way he would on anyone. "I reference them like you would in any other relationship."
He asks portfolio founders what the investor is like to work with, which customer relationships they made, whether they helped with awareness and recruiting, and how they work with other investors
"And those all need to be a unanimous yes before we start working together."
On raising, he said he thinks less about valuation than he probably should. The usual case for a steady step-up in price is real — employee liquidity through tender offers, recruiting, less dilution, a stronger balance sheet, room to invest ahead
"But I'm thinking about a company 20 years from now, not two years from now." Whether the next round is at 15 billion or 25 billion in ten years he called irrelevant
"We need to get to a billion dollars of ARR as quickly as possible." The most important metric is new customer acquisition, which he said runs to hundreds a month
"Our gross retention is north of 99%. Our net dollar retention is north of 200%."
Then they bet on it. Katz: "I'd put the over under at December 2027, and I would take the under."
Stebbings took the other side: "I'm going to go for Feb 28. I think now that's like 18 months. That's more than 18 months." Katz said he would definitely take the under on that timeframe
At the close of the episode Stebbings said he looks forward to wiring a thousand pounds when it happens
Talent, Remote Work and Sixteen to Twenty Offices
Stebbings put his own contested claim to him: that a company trying to hire A-star talent today cannot, because OpenAI and Anthropic simply pay.
Katz said it depends on the category. "I think if you're a digital native AI startup in San Francisco, it's a very difficult employment environment because you're competing against OpenAI and Anthropic and others that are extremely well capitalized and are putting offers that are extraordinarily aggressive into the market."
An infrastructure company hires a different profile — database engineers and distributed-systems people, not the frontier lab's shopping list
"We employ people in 27 different countries, which gives us a competitive advantage." He named Portugal, Germany and Singapore
"You know, we've got single digit attrition, so we've got extraordinarily high retention." ClickHouse has done structured secondaries and will continue to, but less often than younger companies
On remote work he said he contradicts himself constantly. He spent twelve years at Salesforce in the office five or six days a week, ten or twelve hours a day, during what he called the formative stretch of his career, and learned from being around more experienced colleagues
ClickHouse started during COVID and was distributed by design — European co-founders, Katz in the Bay Area, another co-founder in Utah — because the engineers with the right skills were not concentrated anywhere
"We're almost 800 employees. We'll be a thousand by the end of the year." The company is adding in-person options without a return-to-office mandate, with offices in Amsterdam, London, New York, the Bay Area, Singapore, Sydney and Tokyo
"We're going to have 16 to 20 offices around the world within 12 to 18 months.", across both engineering and go-to-market
Asked whether he would put everyone in an office given the choice, he said no, and gave the revenue map as the reason. "Over half of our revenue comes from outside of the US. 40 percent here in EMEA, 10 percent in Asia."
Over half of customers sit outside North America, and ClickHouse is live in 36 regions across all three hyperscalers
"There's no way that you can centrally manage that from one location."
The partnerships need local people too: the company goes to market with AWS, Google Cloud and Azure, and "I flew to China to launch a partnership with Alibaba."
The Valuation History: a $2 Billion Series B With No Revenue, and Why the Latest Round Felt Cheap
Asked which phase of growth was the most uncomfortable, Katz named the six months after launch — which coincided with ChatGPT's arrival.
Database adoption does not behave like a consumer launch. These services take time to evaluate and adopt, unlike a product that reaches a hundred million users in two months, and the slow start shows up on the company's own revenue chart
"But those first six months, I raised 300 million dollars at a valuation that was hard to justify because we had no revenue. We had no product." Revenue has since inflected, with over 4,000 customers
Stebbings put the pricing history to him: "I mean, at like 50 million in revenue, you priced like six billion dollars." — and made the case that venture pricing is always for the next year rather than the point in time
Katz agreed, from the other side of the table. "You're getting priced for where you're going to be in 12 to 18 months." A track record of overachieving against targets buys a multiple disconnected from the public markets
The round that felt most expensive was the Series B. "The Series B that Coatue and Altimeter jointly led at $2 billion felt expensive."
"We had no revenue, we had no product, we had control of an open source database. We had no customers, we had 15 employees."
It put a target on the company's back, and it took a year to get the product out, with long days spent wondering what the reception would be
The cheapest round, he said, is the current one: "Probably the current one." Stebbings agreed, saying that at $15 billion he would rather pay more for more than less for less
He did not claim foresight about AI. Five years ago he had no idea the wave was coming; what he knew was that ClickHouse was the most resource-efficient and performant database in the world, and that in databases it comes down to price and performance
Triple, Triple, Double, Double Is Not Dead. Durability Is What Matters
Stebbings said he often argues that the old triple-triple-double-double growth ladder is dead and that companies now need to go from zero to a hundred million dollars in a year, and asked whether that is wrong for infrastructure.
Katz went back to switching costs. For any category that goes from zero to a hundred million in a year, he asks what moat preserves that hundred million from moving somewhere else
His own curve was slower and he prefers it that way. "So ours was a bit more gradual. We went zero, 12, 50, 200, and we'll finish this year north of 500." In the database world, he said, that is faster than anything he has seen, including the competitors he named earlier
The AI customer base is the smallest slice of it, which is the point. "The basket of AI companies that's using us and nearly every AI company is built on ClickHouse from Harvey, Sierra, Decagon, Anthropic, OpenAI, etc. represents less than 12% of revenue."
"And so even if half of that goes away, the winners are going to offset the loss from the losers."
"So we've got very little concentration risk."
Stebbings asked whether revenue concentration is still a valid concern at all, noting that he is an investor in Mercor, whose revenue is heavily concentrated in frontier model providers, and that Nvidia's is too and it seems to be working
Katz said it is very valid to him as an operator. "If I've got one customer that or one category or one industry that accounts for more than 10% of revenue, I spend a lot of time thinking about it."
"The goal is predictability, sustainability, durable growth."
On the build-your-own-at-scale assumption, Stebbings offered Shopify still using Stripe as the counter-example and asked how Katz thinks about it given his years at Elastic
"It comes down to customer value, right?" The vendor has to stay ahead on the return and the total cost of ownership calculation the customer runs
Katz defined the term when Stebbings asked: "Total cost of ownership." What it costs to run ClickHouse against a comparable service
On whether a company needs an internal enemy, he said no. "I worry about creating adversity inside of a company." There is enough of it outside — competitors, the competitive landscape, geopolitics, and whatever is happening in someone's own life
Stebbings pushed that an enemy creates fire. Katz: "Like, we only hire people that are insanely competitive."
Hiring in the Mid-to-Late Thirties, and Whether College Still Pays
"The average age of our engineering organization is in the mid to late 30s, which is a little bit different, I would imagine, than in a typical venture-backed company that's only had a product for three years." New graduates get hired, but paired with someone who has been building distributed systems for years
Asked whether the skepticism about college is justified, he said he would still send his two teenage daughters, mainly for the life experience, and did not want to dismiss the academic output of a four-year degree
Stebbings disagreed, arguing curricula update too slowly to stay current. Katz drew the line by subject: medicine is durable and doctors will still be needed, and "If you're studying software engineering, I think you're going to be entering a market that's very uncertain in four years' time."
Stebbings pushed on the doctors point with his own behavior — he uses ChatGPT for most medical queries, which keeps him out of the surgery, and noted that doctors are unaffordable for most people and UK waiting times for a general practitioner run to months
Katz held his position: "If I've got a serious medical concern or I need some very important legal advice, I want to talk to the best lawyer in the world." He allowed that he represents a different generation
He thinks the disruption to medicine comes from robotics rather than from artificial general intelligence. "I think robotics would be probably more disruptive, frankly, to the medical industry."
"We had a family member have a procedure that was done entirely autonomously, with just a doctor overseeing it, but didn't actually touch an instrument."
Quick Fire: the Job That Does Not Exist Yet, and the Belief He Thinks Is Wrong
The job he expects to become common in five years: "An AI finance function, solely dedicated on AI consumption inside of an organization." Token and resource management, and nothing else
"And then that job will be made irrelevant, five years from then, because AI agents will govern themselves."
The job that never becomes irrelevant: "Professional football players." — "Professional athletes aren't going anywhere anytime soon."
Given Meta, Microsoft and Nvidia in a buy-and-sell game, he refused to sell any of them. Meta is a customer; "Microsoft's one of our power users. We power the largest analytical workloads at Microsoft, so I'll probably marry Microsoft." He said he met Satya Nadella a couple of months ago and found it impressive, and that ClickHouse does business with Nvidia and Cerebras
On the chip market he expects more suppliers, not fewer. Asked about Etched and Cerebras, he said the hyperscalers and frontier labs are going to be very relevant providers in the chip ecosystem
The widely held belief he thinks is wrong is the bubble itself. "A widely held belief around AI that's generally agreed upon but is wrong is that it's overblown and that we're in a hype cycle, that we're in a bubble."
"You know, I can't pick the winners and losers, but I think the winners are going to far offset the losers."
Where he does worry is who ends up holding it. "I worry about the public exposure to these companies when they're publicly accessible." Right now it is private capital, and investors stand to lose and to make money
On Nvidia sitting in American retirement accounts, he drew the distinction that Nvidia is a public company with public disclosures
"Well, that's the case with any sort of inflated asset. I'm not suggesting Nvidia is inflated."
Stebbings pressed on concentration in the index. "No, but if you see the concentration of value into a few names in this way, like we've never had 85% of the stock market's value predicated in six companies."
Katz: "Yeah, but look at the value creation that's occurred over the last 10 years." Some compression is possible; a return to the levels of ten years ago he called highly unlikely
The Board: Peter Fenton, Mike Volpi and the Investor He Could Not Get
On Peter Fenton, who sits on his board and backed a previous company of his: Katz called him very philosophical, a career venture capitalist who has helped shape some of the most influential companies in technology, with pattern recognition to match. "He's been great for recruiting."
When he puts Fenton on a call, he tells him whether he is buying or selling — trying to convince the person to join, or evaluating them critically
"I compare and contrast him to Mike Volpi, who's also on my board." Volpi was an operator who spent a long time at Cisco running corporate development, and "I think he did over 100 acquisitions."
There is nobody he wishes he had on the board now, and he said an addition is coming that will be announced
The one he wanted at the start: "I met with Martin Casado at Andreessen, who's a friend of mine, and I would have loved to have him involved because I think he understands what we do in a very unique way, technically."
"He was conflicted at the time."
What worries him is the same thing he named as his most feared competitor. "I worry about the technology that isn't yet in the market and that's going to emerge and how defensible our position is against that because that's what we did."
"We burst onto the scene." Nobody anticipated it, which is why he tells the company to keep reinventing itself and disrupt itself first
Asked which of Snowflake or Databricks he would remove from the market, he removed neither. "I've got a ton of respect for Ali and the company." The markets are adjacent, and "There's plenty of white space on either side of the Venn diagram with us and Databricks."
On being a chief executive and a father, he said the attributes are the same: take the job seriously, commit, look for ways to improve, ask for advice, surround yourself with people who have done it, and copy the best of them
On marriage, after 24 years: "The acceptance that it's not a straight line, the willingness to come together with that understanding and embrace one another's differences, celebrate the achievements of the individuals and the relationship, and realize that you have a shared purpose, especially when you have kids." He drew the parallel to aligning a company behind a purpose
Stebbings said what he is most excited about is breakthroughs in chronic conditions, noting that his mother, whom Katz had just met, has multiple sclerosis
Katz said the same medical advances matter to him, and then listed Champions League qualification, the FA Cup, and the United States going further in the next World Cup in Spain, Portugal and Morocco
Why ClickHouse Could Go Public Next Year and Probably Will Not
"I think I take the under on being a public company. We could take the company public next year if we wanted to. There's no rush."
The reason is the market he would be listing into. "Well, the markets are more irrational now, I think, than they've been in a long time."
"Like your stock can trade down 40, 50% on a slight miss in a quarter." He named the effect on employee morale as the cost
He said he has come around on this, and credited a dinner with Databricks' Ali Ghodsi. "Like I remember having dinner with Ali a few years ago." He put it to Ghodsi that the company seemed ready to list, and was told it was already run like a public company
The two things Katz said do not apply: "You don't have your employees looking at your stock price every day. And you don't have anybody shorting your company."
Employee liquidity has largely been solved by structured tenders. The short seller has not
On acquisitions, he pointed at two deals private companies are doing anyway. "Yeah, well, Stripe is buying PayPal for $50 to $60 billion as a private company."
Stebbings added "And OpenRouter as well at $8 billion." and asked the obvious follow-up: why would anyone go public?
Katz gave four reasons and called the question valid. It raises awareness, it is a financing event, it diversifies the investor base and it is good for employee morale. He has been through two IPOs and called the milestone wonderful, celebrated by family, friends and university classmates
Stebbings said that is a short-term high before a volatile stock market takes over. Katz's answer was a defense of public markets on the merits
"I believe that in terms of price discovery, public markets are generally better than private markets." In the long run they behave more rationally than private investors paying a premium to get in
A public price has some speculation in it, but is more grounded in delivered results
"I mean, when Salesforce went public in 2004, it had a billion dollar market cap."
Katz's bottom line is that the bubble question is the wrong one to ask about AI, and that the number an investor should actually interrogate is how durable a company's revenue is once the switching costs are known — which is why he runs a business whose fastest-growing customers are also its smallest concentration of revenue, and why he is in no hurry to hand that business to the public markets.
Products, Companies & Tools Mentioned
ClickHouse (The open-source analytical database Katz co-founded; he says it is used by nearly every AI-native company and that Tesla ingests "a billion events per second" into it)
Anthropic, Claude Code and OpenAI (Customers, suppliers and the reason his own bill moved: ClickHouse's Anthropic spend is "up 100 times from what it was at the beginning of the year," and he says Anthropic chose ClickHouse by asking Claude)
Datadog and Snowflake (The two playbooks he chose between — self-service and developer-led against expensive enterprise sales; he took the first "and it proved to be the case")
Databricks (The open-source-built data warehousing comparison he used to argue open and proprietary end up roughly even, and the company he would not remove from the market)
Salesforce (Where he spent twelve years; the source of his lesson that you "overestimate what you can achieve in one year and underestimate what you can achieve in five," and his example of a billion-dollar market cap at IPO in 2004)
Elastic (His previous company, invoked when Stebbings raised the assumption that customers build their own at scale)
Langfuse (The Berlin agent-observability company, ClickHouse's most recent of six acquisitions in four years — "Every enterprise in the world is going to need this technology")
Tesla, Microsoft and Meta (Named customers; he says ClickHouse powers "the largest analytical workloads at Microsoft")
Harvey and Legora (The legal-AI pair Stebbings asked him to choose between; he refused, saying "Legora could be a bigger company than Harvey a year from now")
Sierra and Decagon (Named with Harvey, Anthropic and OpenAI in the AI cohort he says "represents less than 12% of revenue")
Weights and Biases (Cited alongside Anthropic and OpenAI as the AI-native companies running on ClickHouse)
Fireworks (Stebbings' example of the low-gross-margin AI business, from a previous episode with its chief executive)
Nvidia and Cerebras (Both do business with ClickHouse; Nvidia is also the stock he declined to call inflated — "I'm not suggesting Nvidia is inflated")
Etched (Raised by Stebbings as evidence of a more distributed chip ecosystem, which Katz agreed is coming)
Mercor (Stebbings' own investment, used to ask whether revenue concentrated in frontier labs is still a concern)
Splunk (Named with Datadog as the off-the-shelf observability choice an enterprise engineer still weighs against open source)
PostgreSQL (The managed transactional database he says an agent would pick alongside ClickHouse when provisioning a weekend app)
AWS, Google Cloud and Microsoft Azure (The three hyperscalers ClickHouse goes to market with, and the same three that can redistribute an open-source project as a managed service)
Alibaba ("I flew to China to launch a partnership with Alibaba" — his example of why partnerships need people in region)
Coatue and Altimeter (Joint leads of the $2 billion Series B struck with no revenue, no product, no customers and fifteen employees)
Stripe, PayPal and OpenRouter (The private-company acquisitions he and Stebbings used to argue a listing is no longer needed to buy things)
Fulham Football Club (Front-of-shirt sponsor deal and the recording venue; he calls Craven Cottage "arguably the best sports experience in the world")
The Chainsmokers (Booked to perform in support of ClickHouse at AWS re:Invent and Google Next — his first attempt at buying awareness at scale)
Cisco (Where board member Mike Volpi ran corporate development and, Katz says, "did over 100 acquisitions")
Andreessen Horowitz (Martin Casado's firm; Katz wanted him involved at the start but "He was conflicted at the time")
Oracle, SAP and Siebel (The incumbents Marc Benioff said Salesforce was going after before it had the product to do it)
Get the latest market chatter as it happens:
If this saved you the listen, send it to one person who would rather have the time back.