Databricks: GLM 5.2 takes ~40% of open-weight tokens

Databricks' field engineers describe a strategy shift after two years of enterprises standardizing on single frontier model providers. The triggers are concrete: AI budgets have outpaced revenue and cost savings, and the top intelligence ranking has changed hands 20+ times since May 2023, making single-provider architectures fragile. Guardrails have also become a practical obstacle — during the Hugging Face incident, hosted frontier models refused to analyze real attack logs and exploit payloads, so the team ran the analysis on an open-weight model (GLM 5.2) instead.

The harness around a model turns out to matter as much as the model itself. In Databricks' own benchmark, the same model scored 61.7% in one harness, 48.1% in another, and 38.4% bare, which is why the company argues evaluations should test model and harness together. Databricks positions Unity Gateway as the layer that lets teams switch models by configuration or smart routing rather than code changes, with budgets, fallback chains, and centralized governance across open-weight, external, and frontier models.

The usage data is evidence that customers are acting on this. After GLM 5.2 became available in mid-July, its usage rose while growth among Unity Gateway customers stayed relatively consistent — suggesting existing customers shifted traffic rather than new customers arriving. Databricks' $3 billion raise in July, which CEO Ali Ghodsi tied partly to securing GPU capacity, is presented as support for that demand.