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

In this episode recorded at Fulham Football Club in London, Harry Stebbings interviews Aaron Katz, founder and CEO of ClickHouse, the leading online analytical processing database. ClickHouse recently crossed $350M ARR with customers including Microsoft, Anthropic, OpenAI, Tesla, and Netflix, valued at over $15 billion.

Katz emphasizes that the current AI acceleration is unprecedented compared to previous tech cycles (internet, mobile, social), noting revenue growth has moved from $0 to $12M to $50M to $200M to over $500M annually. He predicts ClickHouse will reach $1B ARR before December 2027, taking the under on that date.

On gross margins, Katz argues that while AI companies currently operate at lower margins (30-35%), this is acceptable as long as they demonstrate a path to margin expansion while maintaining healthy balance sheets and growth rates. The real investor concern should be revenue durability, particularly for agentic applications where switching costs are low, unlike infrastructure software where switching costs are high.

Katz describes his 12-year tenure at Salesforce under Mark Benioff as formative, learning that "you can overestimate what you can achieve in one year and underestimate what you can achieve in five." At Elastic, he studied two infrastructure giants—Datadog (PLG-driven) and Snowflake (enterprise sales-driven)—and initially chose the Datadog playbook. However, he now recognizes the need to layer enterprise sales on top of PLG distribution, a lesson he wishes he'd learned earlier.

Regarding AI tooling adoption, Katz details his "AI awakening" email to employees, noting a 100x increase in Anthropic spend internally. The company is now exploring open-weight models for specific use cases (code review) but maintains concerns about using them for production code due to output inference and indemnification issues. He emphasizes that while open-weight model adoption is increasing, enterprise customers still strongly prefer frontier models (Anthropic, OpenAI) for security, compliance, and legal protection reasons.

On the agent-centric future, Katz articulates a fundamental shift: agents don't have personas, traverse multiple applications, and execute dozens of SQL queries simultaneously. This creates new requirements: low latency, efficiency, and unpredictability of query patterns. He argues agents will eventually have identities, budgets, and authorization mechanisms, and will select infrastructure stacks themselves rather than humans making those decisions.

Katz addresses the open-source vs. proprietary debate, predicting a 50-50 split in enterprise software adoption, similar to today's landscape. He notes that open-source (like Databricks) and proprietary (like Datadog and Snowflake) companies coexist at the top, making it difficult to declare a winner.

On competitive threats, Katz states he fears "the technology coming from the rear view mirror"—not visible competitors but future disruptors, just as ClickHouse disrupted traditional data warehousing companies. The moat lies in maintaining competitive advantage in cloud offerings and proprietary features that are difficult to replicate.

Regarding sales capacity, Katz reflects that he should have invested more aggressively in sales earlier. ClickHouse employs approximately 100 quota-carrying salespeople for revenue significantly exceeding $100M, compared to competitors with 2,000-3,000 salespeople focused on single use cases.

On M&A strategy, ClickHouse has completed six acquisitions over four years, with the most recent being Langfuse for agent observability. Katz decides to build organically when his product and engineering teams can innovate, but acquires when founders are building on ClickHouse and addressing future platform components.

Katz discusses the 50-50 frontier vs. open model token split prediction, arguing enterprise customers have legitimate security concerns about open-weight models (particularly Chinese origins) that will persist for 1-2 years before being addressed. He cites examples of enterprises using Anthropic for sensitive code but open-source models only for code review.

On compensation and remote work, Katz notes single-digit attrition at ClickHouse with employees across 27 countries. While he spent 12 years at Salesforce in the office learning from mentors, ClickHouse was distributed by design during COVID. The company is now introducing optional in-person options across 16-20 global offices within 12-18 months, driven by customer support and partnership needs across time zones.

Katz reflects that the most uncomfortable phase was immediately following product launch when ChatGPT also launched—a six-month period where revenue grew slowly despite a $300M Series B fundraising at a then-unjustifiable $2B valuation with no revenue or product.

On college education, Katz believes it remains valuable for life experience but questions whether software engineering curricula can update fast enough to remain relevant. He would advise his teenage daughters to attend for general development, not specifically for technical training.

Finally, Katz argues that a widely held belief that AI is "overblown" and we're in a bubble is wrong. He maintains "we're just getting started" and that winners in this cycle will far offset losers. The company is positioned to go public next year if desired but will likely remain private, avoiding stock market volatility and short sellers while maintaining employee morale and focus on a 20-year horizon.