Databricks and NVIDIA Are Building a Platform for the Agentic Era

Databricks and NVIDIA are significantly deepening their partnership to build an end-to-end platform purpose-built for the "agentic era," integrating NVIDIA's latest silicon (Vera CPU, Blackwell/Rubin GPUs), serverless GPU runtimes, and agent-specific tooling directly into the Databricks Data Intelligence Platform .

This collaboration targets the full agent lifecycle — from training and fine-tuning on serverless NVIDIA GPUs to inference and orchestration on NVIDIA Vera CPUs — all anchored to the Databricks lakehouse for data governance. Here are the specific technologies, integrations, and industry trends this collaboration encompasses.

Key Technologies & Integrations

NVIDIA Vera CPU Integration

Databricks is integrating NVIDIA's new Vera CPU — branded as "the CPU for agents" — into its platform. Vera is a custom Arm-based CPU with a 256 MB L3 cache and a dedicated hardware unit for JSON parsing, designed specifically to handle the orchestration and reinforcement-learning loops that are bottlenecks for agent workloads . This marks a fundamental shift in infrastructure: NVIDIA and Databricks are betting that AI agent performance is now limited by the host CPU (which runs planning, tool-calling, and memory loops) rather than GPU throughput alone .

AI Runtime: Serverless NVIDIA GPUs

Databricks launched AI Runtime, a serverless compute layer that provides on-demand access to NVIDIA A10 and H100 GPUs for training and fine-tuning deep learning models without infrastructure overhead . Multi-node distributed training and multi-GPU support are in beta, and the service is designed for computer vision, LLMs, and deep learning-based recommendation systems .

GPU Support in Databricks Free Edition

NVIDIA GPU acceleration is now available in Databricks' free tier, lowering the barrier for developers to experiment with AI agents .

Model Serving Enhancements & Agentic AI Stack

The expanded collaboration includes improved model serving for agentic workloads and full support for NVIDIA's agent-oriented software stack (e.g., NVIDIA Agent Toolkit, NIM microservices) running on governed enterprise data within the Databricks lakehouse .

NVIDIA CUDA-Accelerated Photon Engine

NVIDIA's accelerated computing is integrated into Databricks Photon, the high-performance SQL engine, to speed up data preprocessing and ETL for AI pipelines . NVIDIA CEO Jensen Huang noted it took five years to build the libraries that make this acceleration possible .

Industry Trends Driving the Collaboration

The Rise of Agentic AI

The partnership is explicitly framed around the shift from static AI models to autonomous, multi-step AI agents that require tight coupling of data, compute, and orchestration. A Databricks report reveals a 327% surge in autonomous AI systems adoption heading into 2026 .

CPU Becomes the New Bottleneck

NVIDIA and Databricks are betting that AI agent performance is now limited by the host CPU (which runs planning, tool-calling, and memory loops) rather than GPU throughput alone. Vera CPU is designed to close this gap .

Governed Data as the Foundation

A core trend is that agents must act on governed, high-quality enterprise data. The partnership emphasizes bringing NVIDIA acceleration to Databricks' Unity Catalog-governed lakehouse, so agents reason over trusted data rather than siloed or ungoverned sources .

Ecosystem Expansion

Databricks is also investing in open protocols (OpenSharing for Agent Skills) and agent-building tooling (Agent Bricks, Lakebase for agent memory), while NVIDIA is simultaneously expanding its AI Factory partnerships with HPE and others — positioning this collaboration within a broader industry push toward standardized, production-grade agent infrastructure .