Databricks CEO: Stop Scaring People About AI

Summary

Databricks CEO Ali Ghodsi joins a16z GPs Martin Casado and Sarah Wang to discuss AI pacing, recursive self-improvement, cyber risk, and enterprise adoption. Ali argues existential risk is overblown while cyberattacks are an immediate problem, and that model capability is ahead of enterprise adoption because AI lacks organizational context. He explains Databricks' work on ontologies, cost controls, multiple models/harnesses, and agent-optimized infrastructure, and why open-source models are gaining token share. The conversation also covers regulation, lab incentives, and Ali's low p(doom).

• Ali Ghodsi argues leaders should not stoke existential AI fear without evidence; he puts near-term existential risk close to zero.
• Cyberattacks are framed as the immediate AI risk, with exploit weaponization collapsing from years to hours and human SOC teams unable to keep pace.
• Frontier model training is described as increasingly resource-intensive, requiring more GPUs, data centers, networking, and robustness.
• Enterprise AI adoption is held back by missing organizational context; ontologies and context graphs are presented as the key unlock.
• Databricks' internal use of Genie/ontology, AI cost controls, multi-model routing, and harness multiplexing is discussed as evidence.
• The market is shifting toward multiple models and open-source post-training, especially for cost-sensitive or specialized product tasks.
• Agents are becoming primary database users, favoring fast, elastic, branchable infrastructure such as Neon/Lakebase.
• The discussion touches regulation, lab incentives/IPO tensions, and Ali's low p(doom).