Summary
AI tool costs scale with consumption rather than revenue, so neither unlimited tokens nor hard spend caps hold up, and Databricks has shifted to spend visibility plus progressive friction. Four levers (efficient default models, task-based routing to the cheapest capable model, developer spend visibility, and lower token overhead) can cut unit costs by up to 90%, with an AI gateway providing the access, routing, budget, and tracing control layer.
Summary generated by brickster.ai. For the full article, follow the source link above.
More from Databricks Blog
Built for Agents: The Case for Spark Declarative Pipelines
Procedural pipelines used about 34x more compute time than declarative ones for the same work, mainly because Spark Declarative Pipelines can catch structural errors with a cheap dry run instead of fully executing. Declarative pipelines also need 51% less code, leaving less for agents to generate and humans to review, which makes SDP the stronger architectural choice for autonomous agents building production pipelines.
Real-time energy trading and analytics on Databricks
Energy trading teams can run live trading and analytics on one Databricks platform instead of separate systems that leave risk data stale and governance duplicated. Spark Structured Streaming Real-Time Mode streams market data into Lakebase (serverless Postgres), which feeds the trading app, ML models, dashboards and Genie One, while Unity Catalog handles permissions, lineage and audit once for all of it.
The new economics of AI cost management
Enterprises are moving from "tokenmaxxing" to "valuemaxxing," judging AI by the impact generated per dollar rather than by raw usage. Getting there takes attribution granular enough to trace spend to a user, team, project, or even a single pull request, built on spend guardrails, end-to-end observability, and a centralized point through which all AI traffic flows.
The Fraud Analyst AI Coworker: Detecting Telecom Fraud in Real-Time on One Platform
Databricks Lakebase, Databricks Apps, and Genie combine into a single fraud detection stack for telecom, with Lakebase removing the analytical/transactional split for faster data sync, Apps powering a custom case-management workbench, and Genie acting as an AI coworker that answers natural-language questions over governed fraud data. The post argues that real-time network telemetry strengthens fraud models as attackers use agentic AI at machine speed, and the Telecom Fraud Prevention Solution Accelerator deploys the risk engine, workbench, and investigation agent via Databricks Asset Bundles, adaptable to other industries.
Databricks AI Runtime and NVIDIA make it easier to customize open models with enterprise data
Databricks AI Runtime (AIR) is now generally available, letting teams train and fine-tune models on serverless NVIDIA GPUs with NeMo and containers, alongside governed enterprise data and without managing GPU infrastructure. It pairs NVIDIA Kumo models for structured-data predictions and Nemotron Retriever models for finding evidence in unstructured data, with Genie Code, NVIDIA-specific Agent Skills and joint Nemotron cookbooks to help you get started.
AI Runtime is Generally Available
AI Runtime is now generally available, bringing high-performance GPU infrastructure directly into Databricks with seconds-level launches, flexible environments, and integrated observability, so you can scale training without managing GPU infrastructure. Unity Catalog controls and workload isolation are built in to keep every training workload secure and governed as training becomes more agent-led.