ClickHouse's Katz: the AI bubble is wrong; durability is the risk

Summary

• Aaron Katz’s core macro call is that the AI bubble thesis is wrong: “We’re just getting started.” Having lived through internet, mobile and social cycles, he says those “were much more gradual” while this one is “accelerating at an unprecedented pace” — “we haven’t seen revenue growth like this in our lifetime.” His one contrarian close: “I can’t pick the winners and losers, but I think the winners are gonna far offset the losers.”
• The single biggest investor risk is not gross margins but durability of revenue. Switching costs are very high for infrastructure software and “can be very low for agentic applications,” with model providers “leapfrogging one another what seems like every other week” — so he’d “call into question the durability of some of the revenue for some of these AI applications,” explicitly putting Claude Code in the low-switching-cost bucket even as ClickHouse’s own Anthropic spend is “up 100 times from what it was at the beginning of the year.”
• ClickHouse’s numbers are the episode’s hard data: revenue went “zero, 12, 50, 200, and we’ll finish this year north of 500,” with $1B ARR over/under at December 2027 — “I would take the under.” Gross retention is north of 99%, net dollar retention north of 200%, 4,000+ customers with a ~$100k midpoint spend, and the entire AI-native basket (Anthropic, OpenAI, Harvey, Sierra, Decagon) is “less than 12% of revenue” — “even if half of that goes away, the winners are gonna offset the loss from the losers.”
• Katz rejects the consensus that 90% of tokens go through open models: his answer is “50/50,” especially in the enterprise. Enterprises want indemnification and output-inference protections that open-weights models — “especially those that come out of China” — can’t provide today; but Harry’s counter stands too: enterprises distrust frontier labs’ zero-data-retention promises (“It’s basically you saying, ‘Just trust me. I got you’”), which Katz confirms he sees “every day,” including internally. Separately, Katz says ClickHouse uses some open-weight models for code review but not necessarily to ship production code because of inference-output concerns.
• The three-to-five-year thesis: agents become the buyers of infrastructure, selecting the database, compute and storage behind applications — but they’ll need identity, budget and authorization that don’t exist yet. Agentic query patterns demand low latency first and efficiency second (Tesla ingests “a billion events per second into ClickHouse”), and Katz’s tip to investors: figure out “what companies are best positioned to give that agent everything they need to build a software application.”
• His biggest operating regret is being “too efficient” — under-investing in sales capacity. With only ~100 quota reps against data-warehousing incumbents fielding 2,000–3,000 sellers, he deliberately followed “the Datadog playbook” (PLG, developer-led) over “the Snowflake playbook” (expensive enterprise sales), but concedes “at some point, you need to layer in an enterprise sales motion on top.”
• On going public: “We could take the company public next year if we wanted to. There’s no rush.” Private status eliminates two material burdens — employees watching the stock daily and short sellers (“you don’t have people shorting your company when you’re a private company”) — while structured tenders solve liquidity; Harry notes even M&A currency is no longer a reason, citing private-market deals like a reported ~$8B OpenRouter (as stated) acquisition. Katz would still “take the under” on being public within five years.

Deep dive

1. “We’re just getting started” — the bubble call from someone who’s seen three cycles

• Harry frames the paradox he can’t resolve: smart friends telling him “the levels of debt that we’re seeing is insane” and valuations are exuberant, versus adoption and revenue scaling he can see with his own eyes. Katz’s answer, delivered from the operator’s seat rather than a VC pontificating: internet, mobile and social “were much more gradual. This seems to be accelerating at an unprecedented pace” — both in how quickly agentic experiences mature and how fast the companies grow.
• His quickfire version of the same call: the widely held belief that’s wrong is “that it’s overblown and that we’re in a hype cycle, that we’re in a bubble… I can’t pick the winners and losers, but I think the winners are gonna far offset the losers.”
• On systemic debt risk, his hedge is specific: he worries about “the public exposure to these companies when they’re publicly accessible” — right now it’s private capital, and Nvidia at least has public disclosures. Compression? “Possibly.” Back to levels of 10 years ago? “Highly unlikely.”

2. Margins matter less than durability of revenue

• On the low-gross-margin question (Harry cites FireEye’s Slim at 30–35%): these companies have “unlimited access to capital right now,” so as long as they show “a path to margin expansion over the course of the next few years while still growing at these unprecedented levels with very healthy balance sheets, I worry less about gross margins like we did five years ago in traditional enterprise software.”
• What Harry should worry about instead: “the single biggest risk… would be durability of revenue.” Switching costs are very high for infrastructure software but “can be very low for agentic applications,” with model providers “leapfrogging one another what seems like every other week.”
• Harry’s sharpest test case — a hypothetical $2T Anthropic IPO riding Claude Code as the Trojan horse: wouldn’t that fall in the low-switching-cost bucket? Katz doesn’t dodge: “I would put it in that category” — even though it’s not showing up in ClickHouse’s own usage, where “our Anthropic spend is up 100 times from what it was at the beginning of the year.”

3. The AI Awakening email — and how Katz thinks about token budgets

• At the turn of the year Katz emailed the company under the subject “The AI Awakening”: “I think we’re moving too slowly and I’m not seeing the adoption of these coding applications… that I would expect to see for a leading database provider like ClickHouse.” The company rallied; they’re now also evaluating open-weights models, though Anthropic and OpenAI have the advantage of being “both a model provider and the harness provider that the open-weights models right now are behind in.”
• Against the Uber president’s ROI skepticism (Harry’s reference), Katz’s discipline is revenue-first: “as long as we can continue to ship features… then I can always rein in token consumption. And the cost of tokens, as we know, is going down, not up… our revenue’s growing faster than it ever has, so I’ll take that trade any day of the week.”
• The practical limit on open-weights internally: “we’ll use [them] for code review, but we won’t necessarily use [them] to ship production code because we have concerns about the output inference.”

4. The regret: too few salespeople, and the Benioff inheritance

• ClickHouse runs ~100 quota-carrying reps at revenue “significantly more than 100 million,” against data-warehousing and observability incumbents with “thousands of salespeople” who “wake up every morning thinking about one specific use case” while his think about ten. The one thing he’d change over the last two years: “increasing sales capacity.”
• The deliberate design behind it: he distilled the last decade’s infrastructure winners to Datadog (PLG, self-serve, never talk to sales) and Snowflake (heavy enterprise sales), and “thought it was gonna be a lot easier to follow the Datadog playbook than the Snowflake playbook, and it proved to be the case. But at some point, you need to layer in an enterprise sales motion.”
• His biggest lesson from 12 years with Benioff: “you can overestimate what you can achieve in one year and underestimate what you can achieve in five” — Salesforce was “a glorified contact manager” when Marc declared war on Siebel, SAP, Oracle and Microsoft, without the product to back it, and delivered on the roadmap anyway.

5. Agents as the next buyers of infrastructure

• The design shift: software used to serve personas with predictable query patterns; “agents don’t have personas,” traverse across observability, data warehousing and CRM, and “the experience is gonna be defined by the slowest point in that chain.” The agentic requirement stack: low latency first, efficiency second — Tesla is “ingesting a billion events per second into ClickHouse,” throughput Katz calls unprecedented.
• Anthropic’s example as the tell: “they asked Claude, ‘What technology should we use for this specific observability use case?’ And Claude suggested ClickHouse.” Katz’s future: agents that provision the entire stack — but “they need to have authorization. They need an identity. They need to have a budget… We’re not there yet today.” His prompt to Harry: find “what companies are best positioned to give that agent everything they need to build a software application.”
• The counterweight he insists on: “for the next decade, there will still be a human involved in that decision loop,” budgets are still controlled by people, so brand investment doesn’t change. Quickfire corollary: the job that doesn’t exist yet is “an AI finance function solely dedicated to AI consumption” — a role “made irrelevant five years from then because AI agents will govern themselves.”

6. Open vs. frontier: 50/50, and the enterprise trust inversion

• Harry serves the conventional wisdom — 90% of tokens through open models, frontier reserved for cancer and climate. Katz refuses it: “the easy answer is 50/50,” anchored in the analogy that open source vs. proprietary enterprise software is “a pretty even distribution” today. Snowflake is primarily closed while Databricks is built around open source; after Harry picks Databricks in both comparisons, Katz argues for Datadog over Databricks in observability and notes that 3–5 years is a long time.
• He’s careful to separate open-weights models from open-source software — “we’re in the latter category” — and his enterprise-adoption prediction for open weights is far more cautious: indemnification and output-inference gaps “will limit the use cases,” especially for Chinese models.
• Harry’s counter-anecdote — a guest who uses Anthropic for less sensitive work and “an open source Chinese model” for the most sensitive, which Harry assumed was backwards. Katz’s explanation of the frontier-lab fear: zero data retention “is basically you saying, ‘Just trust me. I got you’… a lot of companies worry about sending their source code, for example, to a frontier lab.” Does he see it? “I do. I see it every day,” including internally.
• On the CCP-backdoor thesis he won’t bite: “I personally don’t like to speculate… it’s quite fantastic to go there.” But today’s security concerns about open-weights models? “Today I think it is [justified]. If I look out a year or two from now, I think a lot of these concerns will be addressed.” And he wouldn’t take Harry’s capital to build an American open-weights model. A related enterprise signal: even “some of the most innovative digital native companies in Silicon Valley” are talking about moving back on-prem.

7. The numbers: zero–12–50–200–500+, and the December 2027 bet

• The growth curve as stated: “We went zero, 12, 50, 200, and we’ll finish this year north of 500” — faster first-three-years growth “than we’ve ever seen” in databases. The billion-ARR bet: “I’d put the over under at December 2027, and I would take the under” (Harry took Feb ‘28; £1,000 rides on it). When Harry asked whether they were at $250 million, Katz said they were “well north of that.”
• Unit economics: 4,000+ production customers, hundreds added monthly, gross retention north of 99%, NDR over 200% (versus “over 130%” at his previous companies), spend ranging from thousands per month to tens of millions per year, midpoint ~$100k. Expansion beats acquisition because siloed use cases — data warehousing and real-time analytics — collapse into one repository.
• On concentration risk, Katz declines Harry’s “seems to be working for Jensen” bait: as an operator, any customer, category or industry over 10% of revenue means “you’ve got exposure.” The basket of AI companies using ClickHouse “includes nearly every AI company built on ClickHouse, from Harvey, Sierra, and Decagon to Anthropic and OpenAI” and represents “less than 12% of revenue. Even if half of that goes away, the winners are gonna offset the loss from the losers.” Same logic answers zero-to-100-in-a-year envy: “durability of revenue is the most underestimated attribute of these companies.”

8. Moats, the competitor in the rear-view mirror, and why Fulham

• The competitor he fears most: “the one that isn’t in the market yet… I worry about the next ClickHouse” — an open-source project that “burst on the scene” a decade ago exactly the way a disruptor would today. Internal mandate: “we need to constantly think about reinventing ourselves… so that we can basically disrupt ourselves.” What investors get most wrong: the hyperscaler-redistribution moat question — survivable “if you can maintain a competitive advantage with your cloud offering or your proprietary features,” which “very few open source companies get right.” And he wouldn’t dismiss Anthropic and OpenAI as future core infrastructure providers: “just imagine the surface area they’re gonna cover in three years.”
• The brand playbook: The Chainsmokers concerts at re:Invent (“I want one party that 60,000 software engineers are falling over themselves to get access to”), then the Fulham front-of-shirt sponsorship — funded from a round extension with David Sacks, Michael Dell and JP Morgan when “we didn’t need the capital. We had a billion dollars on the balance sheet.” ROI splits into global-broadcast awareness and quantifiable hospitality: 20 budget-owning executives at a Michelin-grade dinner last night, half customers, half prospects.
• On sports assets after the Lakers at $12.5B and Seahawks at a reported $9.6B: yes to $20B teams — “it’s an experience you simply can’t replicate through technology… there’s something very visceral about it,” and Harry adds that in an AI world people will crave those experiences more.

9. Rounds, remote work, and why there’s no rush to IPO

• The Series B — Coatue and Altimeter jointly at $2B with “no revenue… no product… 15 employees” — felt most expensive and “put a pretty big target on our back”; the current round at $15B feels cheapest. His fundraising philosophy: “whether or not we raise that 15 billion or 25 billion in 10 years is irrelevant… I’m thinking about a company 20 years from now, not two years from now.” And a candid caveat on timing: five years ago “I had no idea” the AI wave was coming — he just knew ClickHouse had an advantage on price and performance, so it “would satisfy whatever trend was gonna come.”
• On Harry’s anti-remote crusade, Katz is “of two minds, and I contradict myself constantly”: Salesforce’s in-office decade was formative, but ClickHouse was distributed by COVID-era design, employs people in 27 countries with single-digit attrition, over half of revenue outside the US, and will have “16 to 20 offices around the world within 12 to 18 months,” rather than 6 or 7 — no draconian mandate, but rising employee demand for offices.
• On going public: “We could take the company public next year if we wanted to. There’s no rush.” His full-circle framing from a dinner with Oli (who claimed to “basically run a public company”): only two dimensions truly differ — employees watching the stock daily and short sellers — and structured tenders have largely addressed employee liquidity, but “you don’t have people shorting your company when you’re a private company.” Harry pushes further, citing private-company deals like Stripe–PayPal at $50–60B (Katz: “bit of an outlier”) and OpenRouter (as spoken) at $8B, why go public at all? Katz’s case: awareness, financing, investor diversification, morale — and a belief that “in terms of price discovery, public markets are generally better than private markets in the long term.”