Key Takeaways
- Aaron Katz scaled ClickHouse to over $350 million in ARR by relying on the high switching costs native to core database infrastructure.
- Gross margin scrutiny from traditional SaaS does not apply cleanly to AI companies that maintain strong balance sheets and rapid top-line growth.
- The primary existential risk for AI applications is revenue durability, because swapping an agentic layer takes days while replacing a database takes years.
- Fast growth can mask structural fragility: jumping from zero to $100 million in ARR within a single year means little without a defense against model leapfrogging.
The Real Danger in Fast AI Growth
Public market investors spend hours dissecting the compressed gross margins of model providers and inference vendors. Five years ago, enterprise software lived or died on whether gross margins cleared 80 percent.
Katz takes the other side of that debate. As ClickHouse crossed $350 million in ARR, his focus shifted entirely away from traditional margin obsession. “I think as long as they can demonstrate 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 5 years ago in traditional enterprise software,” Katz explained.
Raw compute costs can be solved with time, scale, and hardware efficiencies. Ephemeral revenue cannot.
Why Agentic Applications Bleed Customers
The real threat to AI startups is the illusion of stickiness. In infrastructure software, changing a database requires rewriting schema, redirecting pipelines, and risking downtime across production workloads. The switching costs are massive. That friction creates natural revenue durability.
Agentic applications have the opposite profile. When a developer builds an automated workflow on top of an API wrapper or a thin agent layer, replacing that vendor with a cheaper, faster frontier model requires minimal code changes.
“If you were to come down to one, if you were to say this, what's the single biggest risk? Would be durability of revenue because the switching costs that you and I talked about are very high for infrastructure software,” Katz stated. “The switching costs can be very low for agentic applications.”
When a new model drops every six weeks, the top layer of the stack constantly faces displacement. Katz noted: “For any category that goes from zero to 100 million in a year I worry what's the competitive moat that they have to preserve that 100 million from that customer going to something else.”
Efficiency vs Velocity in the AI Era
Katz does not ignore costs inside ClickHouse, but he measures them against roadmap acceleration rather than pure margin optimization. When evaluating internal spend on coding agents, the calculation shifts toward engineering speed.
“The primary measure that we care about, as you know, is revenue growth,” Katz said. “And if we see the sustained revenue growth that we've experienced over the last 3 years and we're a very efficient company as you know, some would argue we're too efficient, then I worry less about the expense that we're incurring on coding agents for example because we're covering a very broad surface area and our road map is accelerating at a pace that we've never seen before.”
If software spend directly increases product velocity, the cost justifies itself. But if your product is merely passing calls to an underlying model without embedding deep into customer operations, your revenue will evaporate the moment an alternative appears.
What to Do With This
Audit your product's customer offboarding friction this week. Map every step a customer takes to replace your tool with a raw frontier model or an open-weights alternative: if the migration takes less than two hours and a single config update, your top-line revenue is fragile regardless of your growth rate. Begin building proprietary state, complex data pipelines, or deep operational workflows into the core product to raise switching barriers immediately.