Databricks' Ali Ghodsi on AI risk, cyber and enterprise ontology

Key points

Key takeaways from an a16z interview with Databricks CEO Ali Ghodsi (September 2026):

Existential risk near zero. Ghodsi argues leaders should not needlessly alarm the public, estimating current existential AI risk as close to zero while acknowledging real cyber risk from autonomous agents.

Pacing debate. He is sceptical of the "pacing the frontier" framing as a PR strategy, saying the underlying Dario Amodei document is sensible but the framing muddies the focus on security and safety controls.

RSI litmus test. Ghodsi proposes four conditions for genuine recursive self-improvement — a next model trained with less compute and time, higher intelligence, repeated across generations — and sees no evidence they hold today, noting frontier training runs are slower, more brittle and more expensive.

Cyber as the real risk. Time from a vulnerability to weaponised exploit has collapsed from years to hours; Databricks' LakeWatch targets automated detection because human security teams cannot keep pace.

Ontology as the unlock. Ghodsi frames enterprise AI value as building an ontology — a graph of an organisation's people, projects and processes — that gives models the context most companies still lack.

Dogfooding at scale. Databricks built the largest such ontology for itself, with millions of nodes, and employees query it via Genie for answers that once took meetings; roughly 90% of Databricks software is now AI-written.

Cost control. Through UniGateway, model routing and harness multiplexing, Databricks has kept AI costs roughly flat even as token usage climbs, and Ghodsi notes open-source models now make up over 60% of token volume.

Post-training demand. Startups are increasingly post-training open-source models for specific tasks, an approach Databricks supports, though most enterprises still default to frontier models for convenience.

Read more: a16z on YouTube