ClickHouse CEO: AI Margins Need to Improve | Revenue Concentration Should be a Concern

Why does it matter? Because the AI revenue boom is real — and parts of it will vanish when the next model ships.

The CEO of a $15B database company has a simple message for everyone celebrating AI growth: you may be counting money that's about to disappear. Aaron Kass built ClickHouse into the infrastructure layer behind nearly every top AI company — Anthropic, OpenAI, Harvey, Sierra, Decagon — and his view from that position is not reassuring for application-layer investors.

• Revenue durability, not gross margin, is the single biggest risk in AI investing today
• The identity and budget layer that will govern autonomous agents doesn't exist yet — and it's the most important infrastructure primitive no one is building
• Enterprise legal requirements will keep frontier labs dominant far longer than benchmark timelines suggest
• Any customer, category, or sector above 10% of revenue is a structural red flag — even if that category is "every AI company in the world"

Switching costs are infrastructure's moat. For most AI applications, those costs don't exist.

A model leapfrogs its predecessor, and the application built on top of it loses its edge overnight. That's the actual risk investors are underpricing in today's AI boom, from the person watching it through the infrastructure layer.

"The switching costs that you and I talked about are very high for infrastructure software. The switching costs can be very low for agentic applications."

Model providers are leapfrogging each other what seems like every other week. The applications riding on top carry the same fragility. When asked directly whether Claude Code — potentially the cornerstone of Anthropic's case for a $2 trillion valuation — falls into the low-switching-cost bucket, Kass doesn't hesitate: "I would put it in that category."

That's not showing up yet in the numbers. ClickHouse's own Anthropic spend is up 100x from the start of the year — the growth is unambiguous. But infrastructure migration is painful in ways that model-switching isn't. Data bakes into systems. Query patterns evolve around them. Architectures depend on them. A coding harness? You change a configuration file.

Before writing a check at an AI application's growth multiple, one question needs an honest answer: if a better model ships tomorrow, does this revenue survive? For a lot of AI applications, the truthful answer is no — and the multiple isn't justified.

Agents will become the buyers of infrastructure — but no one has built what they need to pay

Agents don't think about cost. Kass is direct about this: "I don't see agents necessarily being cost-efficient. I don't see them thinking about consumption and budget like humans do." That's not the interesting part. The consequential shift is where the purchasing decision itself is moving.

Anthropic chose ClickHouse by asking Claude which database to use for an observability use case. Claude recommended ClickHouse. Kass is already projecting three years past that: "I'm thinking about a future where they say, 'Hey, we need to build an application, provision the underlying stack.' So you've got a database, you've got networking, you've got compute, you've got storage, and the agents actually making that selection process."

Those agents will need identity. They'll need a budget. They'll need authorization to consume services before they can buy anything. "We're not there yet today." A human is currently watching every dollar of agentic spend, monitoring consumption, enforcing limits. "If you look out 3 to 5 years, those agents are going to be fully autonomous."

This is not a problem the harness or model provider will absorb — companies won't let agents run wild and consume whatever resources they choose. The governance layer is a standalone unsolved problem. Whoever builds it sits upstream of every database, compute, and storage decision in the agentic economy. The category does not exist today.

Ninety percent of enterprise tokens through open weights? The legal department will stop that.

Enterprise procurement doesn't run on benchmark timelines. It runs on legal liability timelines — and that distinction quietly protects frontier labs from the displacement most AI investors have already priced in.

The consensus circulating through most AI conversations: frontier labs handle cancer and climate change, open-weight models take everything else, roughly 90/10. Kass disagrees with the ratio. "In the enterprise they want provisions and protections that potentially open-weight models, especially those that come out of China, cannot provide around indemnification for example and output inference."

ClickHouse itself draws the line at production code. Open-weight models are fine for code review; they don't push to environments that customers use. "I think there's too much security concern around some of these open-weight models. Is it justified today? I think it is."

The debate gets complicated: some enterprises use frontier labs for less sensitive tasks and open-weight Chinese models for the sensitive stuff, because they distrust frontier labs' zero-data-retention claims. Kass acknowledges the dynamic without endorsing the logic.

One distinction he insists on separating: open-weight and open-source are not the same. ClickHouse is open-source — inspectable, forkable, monetizable by anyone under its license. Open-weight means the model weights are public; legal accountability doesn't follow. Do not underwrite frontier lab revenue erosion based on open-weight adoption curves. Enterprise procurement moves on liability, not leaderboards.

Internet, mobile, social — he's been through all three. None moved like this.

Tesla is ingesting a billion events per second into ClickHouse. That throughput is unprecedented — and it's the right frame for everything Kass says about the current moment.

Prior cycles were gradual. Internet, mobile, social each felt fast in their time. "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. We haven't seen revenue growth like this in our lifetime."

The infrastructure demands are qualitatively different, not just larger. Human query patterns were predictable — reports, dashboards, personas with defined roles inside an organization. Agents have no persona. They traverse observability, data warehousing, and CRM simultaneously, executing dozens of SQL queries in parallel with exploratory patterns no human query ever produced. The number-one requirement for agent query patterns is low latency, and agents are constrained by neither access nor time zones.

ClickHouse is shipping product two years ahead of its own roadmap, organically and through six acquisitions in four years — the most recent being Langfuse out of Berlin for agent observability. Revenue ran 0, 12, 50, 200, and will finish north of 500 this year. In the database category, that's the fastest growth trajectory in the industry's history, ahead of every named competitor.

Rebase your adoption curve models upward. Agent infrastructure requirements are not incremental on human query patterns — they are orders of magnitude beyond them.

ClickHouse powers nearly every AI company. That whole basket is still under 12% of revenue — by design.

Harvey, Sierra, Decagon, Anthropic, OpenAI — nearly every AI-native company runs on ClickHouse. This looks like concentration risk. It isn't, by design.

"The basket of AI companies that's using us and nearly every AI company's 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."

The operating rule Kass applies: any single customer, category, or industry above 10% of revenue gets disproportionate scrutiny. "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." He describes 10% as the threshold at which a revenue dimension becomes genuine exposure.

The Nvidia comparison surfaces — Mccor reportedly draws 90% of revenue from frontier model providers, and Jensen is doing fine. Kass doesn't find this instructive for operators. "The goal is predictability, sustainability, durable growth. And if I've got one category or sector or customer that can have such a negative effect if they were to leave the platform, that's a concern for me."

Apply the 10% rule across customer, category, and sector when diligencing any infrastructure play. Serving every AI company is a feature only as long as the AI industry stays fragmented. When it consolidates — and it will — concentration collapses fast.

He doesn't lose sleep over Snowflake. He loses sleep over the company no one has heard of yet.

The most feared competitor has no name. "The one that isn't in the market yet. Like our competitors are right in front of me. I can see them. I know their strengths. I know their weaknesses. And what I worry about is the technology coming from the rearview mirror."

ClickHouse was that rearview-mirror threat once — open source a decade ago, dismissed as another database curiosity with no company behind it, adopted by thousands of developers who largely treated it as interesting but not commercial. Then it wasn't. A company formed around it, thousands more adopted it, and the incumbents never saw it coming.

The response isn't a competitive intelligence team. It's a standing mandate to the company itself: "I go to the company and say we need to constantly think about reinventing ourselves to be that disruptor so that we can basically disrupt ourselves." The company most likely to disrupt ClickHouse, in his estimation, is ClickHouse. The alternative is waiting to be surprised by the version of the challenge no benchmark could have identified.

The two traditional reasons anyone went public have quietly disappeared

Employee liquidity is solved by structured tenders. Acquisition currency? "Stripe is buying PayPal for 50 to 60 billion as a private company." Those were the two pillars of the IPO rationale — and both are gone.

What remains is mostly downside. Public company stock can trade down 40-50% on a single slight quarterly miss. Employees watch the price. Shorts appear. None of that exists in the private markets, and none of it is abstract — Kass has been through two IPOs and watched the post-celebration reality closely.

"We could take the company public next year if we wanted to. There's no rush."

He's not anti-IPO. Public markets offer better long-run price discovery, broader investor diversification, and a genuine shared moment for employees, families, and communities. But the IPO is no longer a mandatory milestone for a well-capitalized private company with a functioning liquidity program. Staying private indefinitely is now a legitimate permanent strategy — not a holding pattern between funding rounds. Underwrite private companies on that assumption, and value liquidity timelines accordingly.

The capital is chasing the fragile layer

The dynamic that doesn't fully surface above: infrastructure compounds with the ecosystem. Every new AI company that forms is a potential ClickHouse customer, and that list grows regardless of which models or applications win. Application revenue tied to specific models resets when those models get leapfrogged. The more AI companies form, the more entrenched infrastructure becomes — and the more exposed everything built on top of it gets. Most of the capital chasing AI today is pointed at the wrong layer.

---

Topics: AI infrastructure, database, ClickHouse, venture investing, agentic AI, open source vs open weights, SaaS margins, revenue durability, enterprise software, GTM strategy, IPO, switching costs