Databricks CEO on AI Pacing, Cyber Risk, and the Enterprise

What It Covers

Databricks CEO Ali Ghodsi joins a16z's Martin Casado and Sarah Wang to examine why AI adoption stalls in enterprises, how cybersecurity risk is accelerating faster than human response capacity, and what four specific criteria would signal genuine recursive self-improvement risk — while arguing current existential AI fears are overstated and misdirected.

Key Questions Answered

• •Enterprise AI Bottleneck: Models are already smart enough for most enterprise use cases — the real barrier is context, not intelligence. Organizations need an "ontology": a structured graph mapping people, projects, goals, and processes. Databricks built this internally and now runs millions of nodes. The result is that employees default to querying the system rather than each other, compressing decision cycles from days to minutes.
• •Cybersecurity Acceleration: The window between a published CVE vulnerability and active weaponization collapsed from two to three years in 2018-2019, to eight to nine months in 2022, to hours today. Human security operations centers cannot process detections at this speed. Organizations must shift to fully automated, agent-driven threat detection and threat hunting — most have not yet made this transition, creating compounding exposure.
• •RSI Four-Criteria Test: Recursive self-improvement only becomes genuinely dangerous if four conditions occur simultaneously: each successive model requires fewer GPUs, less training time, produces higher accuracy, and the cycle repeats compoundingly. Current frontier model training does the opposite — it demands more compute, more engineers, and longer timelines each generation, making runaway self-improvement implausible under present conditions.
• •Token Cost Management: Databricks held AI spending flat despite rising token volume by deploying Unity Gateway with per-person budget caps, smart model routing that substitutes cheaper models for simple tasks, and harness switching — the same model run through a different harness produces up to a 2x cost difference. Organizations should audit harness selection before switching models when optimizing AI spend.
• •Model Routing Strategy: Enterprises and startups are converging on a tiered model architecture: a frontier model like Claude or Gemini handles architecture design and audit, a cheaper or open-source model handles implementation, then frontier re-enters for review. By token count, open-source models now represent over 60% of usage on Databricks; by dollar spend, they represent only around 5%, reflecting cost-per-token differences.
• •Organizational Ontology Implementation: Building an enterprise ontology requires first digitizing all organizational activity — recording meetings, capturing decisions, logging agent actions. Legal teams frequently resist full recording policies, creating gaps. Once data is collected, it must be distilled into a permission-aware graph (unlike public web indexing, access controls vary by user). Databricks built this infrastructure internally before offering it to customers through the Genie product.