20VC: The AI Bubble Is Wrong | AI Margins Need to Improve | Revenue Concentration Should be a Concern | Why People Over-Estimate Open Models But Enterprises Still Fear Frontier Models with Aaron Katz, ClickHouse

What It Covers

Aaron Katz, CEO of ClickHouse, discusses building a $15B database company from zero to $350M ARR, why AI gross margins will improve, why revenue concentration should concern investors, how agentic query patterns are reshaping infrastructure decisions, and why enterprises remain more cautious about open-weight models than frontier labs.

Key Questions Answered

• •Revenue durability over growth rate: When evaluating AI companies, prioritize switching cost depth over headline growth. ClickHouse maintains 99%+ gross retention and 200%+ net dollar retention because database migration costs are prohibitive. Agentic application layers, by contrast, carry low switching costs — model providers leapfrog each other weekly. Investors should stress-test whether revenue survives a single competitor breakthrough before underwriting any growth multiple.
• •AI gross margin trajectory: Companies showing 35% gross margins today can still attract capital if they demonstrate a credible path to expansion over two to three years while maintaining hypergrowth and healthy balance sheets. The old SaaS margin benchmarks no longer apply as a gating criterion at early scale — token costs are falling, not rising, which structurally improves unit economics without requiring pricing changes.
• •Revenue concentration threshold: Operators should treat any single customer, category, or industry exceeding 10% of ARR as a material risk requiring active management. ClickHouse deliberately limits AI-native company revenue to under 12% of total ARR, ensuring that even if half that cohort churns, expansion from surviving winners offsets losses. Concentration above 10% undermines the predictability and durability that justify premium valuations.
• •PLG-to-enterprise layering sequence: Build product-led growth distribution first to generate organic adoption and compress sales cycles, then layer enterprise sales capacity on top — not simultaneously. ClickHouse followed the Datadog playbook before adding enterprise motion, but underinvested in sales headcount too long. With only 100 quota-carrying reps against competitors fielding 2,000–3,000 sellers, earlier capacity investment would have accelerated enterprise penetration significantly.
• •Agent infrastructure requirements: Agentic query patterns are fundamentally unpredictable, exploratory, and simultaneous — unlike human-driven dashboards or scheduled reports. Systems must prioritize sub-millisecond latency and extreme ingestion throughput. Tesla ingests one billion events per second into ClickHouse. Builders designing agent-facing infrastructure should optimize for query unpredictability and resource efficiency first, then feature completeness, reversing the traditional enterprise software prioritization hierarchy.
• •Open-weight vs. frontier model enterprise adoption: Enterprises are not converging on open-weight models as quickly as predicted. Legal indemnification, output inference liability, and data privacy compliance requirements keep large regulated industries anchored to frontier lab providers. ClickHouse uses open-weight models internally for code review but not production deployments. Investors projecting rapid open-weight enterprise displacement should discount adoption timelines by at least two years for regulated verticals.