The Full Stack of AI, Accelerated

NVIDIA accelerated computing powers some of the most demanding AI workloads on Databricks, from large-scale training, fine-tuning, and inference to industry-specific AI solutions. Today at Data + AI Summit, we're highlighting how NVIDIA AI infrastructure lies at the center of new announcements from Databricks AI Runtime, Model Serving, and Industry AI solutions, including a look at how the new NVIDIA Vera CPU will power the next generation of agentic infrastructure.

"Our partnership with NVIDIA spans the full AI lifecycle. From NVIDIA accelerated infrastructure powering distributed training in AI Runtime to software running inside our serving and developer platforms. We're excited to combine NVIDIA technology with the data and governance capabilities of Databricks to unlock incredible value for our customers: enterprise AI that's fast, scalable, and built on a foundation they can trust."— Adam Conway, SVP, Product, Databricks

“Databricks enables enterprises to build, deploy, scale and govern AI agents that are informed by their most valuable resource: business data. Through our expanded partnership, NVIDIA and Databricks are supercharging the next wave of enterprise AI by embedding full-stack NVIDIA accelerated computing with Vera CPUs, Rubin GPUs, NVIDIA Quantum InfiniBand networking and NVIDIA Agent Toolkit software into the Databricks platform.”— Pat Lee, Vice President, Enterprise Strategic Partnerships, NVIDIA

Here's how Databricks and NVIDIA are building an AI platform together, from GPUs for training and inference, to purpose-built CPUs for the agentic era.

1. Training and Fine-Tuning

Databricks AI Runtime (AIR) brings NVIDIA GPU acceleration directly to data and AI teams, so they can train and fine-tune models on governed enterprise data without managing separate GPU infrastructure. With AIR, customers obtain the advanced NVIDIA hardware and networking, directly where their governed data is on Databricks:

• NVIDIA Hopper GPUs with NVIDIA Quantum InfiniBand: purpose-built for multi-node distributed training. Whether you're pre-training a foundation model or running large-scale fine-tuning, AIR provides built-in support for NVIDIA’s high-bandwidth, low-latency GPU interconnects (RDMA-capable networking) that eliminate communication bottlenecks across nodes. AIR is also being prepared for the NVIDIA Blackwell architecture, ensuring customers are always on the leading edge of accelerated computing.
• NVIDIA GPUs in Free Edition: at DAIS, we’re excited to announce the support of GPUs within Databricks Free Edition, supporting developers, students, and startups worldwide to build and deploy their AI workloads on GPUs.
• Support for NVIDIA containers: Soon, Databricks will support NGC containers and custom NVIDIA CUDA environments, enabling them to run natively on data within the platform.

AI Runtime enables seamless access to NVIDIA GPUs within Databricks.

2. Inference: NVIDIA Acceleration in Databricks Model Serving

Databricks Model Serving powers production inference for thousands of Databricks customers. At the core of Model Serving, NVIDIA hardware and software deliver the low-latency, high-throughput inference at scale our customers need, across frontier models like Qwen, GPT-OSS and custom neural networks our customers build. Additional serving capabilities include NVIDIA hardware and Triton Inference Server. Model Serving supports leading inference-optimized GPUs with Triton's advanced dynamic batching and optimized performance coming soon. With Model Serving, customers can serve the models they train on NVIDIA hardware directly on managed Databricks infrastructure.

3. Agentic infrastructure: exploring NVIDIA Vera for the next compute bottleneck

The rise of autonomous agents introduces a new infrastructure challenge. While GPUs excel at model inference, the agent harness, tool calls, CPU-powered analytics and managing multi-step reasoning, all run on CPUs. Today's CPUs are often the bottleneck: latency in tool calling, communication overhead between agent steps, and inconsistent performance under load all degrade the agentic experience.

NVIDIA Vera is a next-generation CPU designed specifically for this workload. Engineered for three core use cases, agentic workloads, reinforcement learning, and CPU-based data analytics, Vera delivers:

• High-performance NVIDIA-designed, Arm-compatible cores that deliver up to 3x faster SQL queries and 80% faster agentic performance, optimized for the latency-sensitive, bursty compute patterns such as tool calls and agent orchestration
• Massive memory bandwidth for the data-intensive operations that agents perform between model calls
• Fast core-to-core communication helping deliver predictable performance as agent complexity scales

The vision is an end-to-end NVIDIA-accelerated stack on Databricks: models run on NVIDIA GPUs for inference, while the agent harness and tool calls could run on Vera CPUs, each workload on silicon purpose-built for its characteristics. Developers customize models on Databricks using proprietary data, deploy them via Model Serving, and the surrounding agentic infrastructure runs on compute designed from the ground up for that exact pattern.

4. Developer experience: making accelerated AI easier to build

NVIDIA Agent Toolkit: Deploy on Databricks

Built on Databricks Apps, teams can host and run NVIDIA Agent Toolkit, NVIDIA's open source development platform for building, customizing, and deploying agentic AI workflows, directly within their Databricks environment. This means you get:

• NVIDIA Agent Toolkit capabilities: guardrails, tool use, retrieval-augmented generation, and multi-step reasoning, running in applications hosted on Databricks.
• Databricks Apps as the hosting layer: deploy any codebase, including agents or services built with NVIDIA Agent Toolkit, as managed applications with built-in authentication, networking, and governance through Unity Catalog.
• Seamless integration with Databricks' data, models, and serving infrastructure. Your agents can access governed data, call models via FMAPI, and leverage the full platform without leaving the environment.

Using Genie Code for GPU Workloads

GPUs are powerful, but getting great utilization, diagnosing performance issues, and debugging workloads has traditionally required deep systems expertise. We're changing that with an agent-first approach:

Genie Code supports skills designed around NVIDIA hardware and software. Developers can:

• Debug GPU workloads conversationally: describe the issue, get actionable guidance
• Performance-optimize: identify utilization bottlenecks, memory issues, and communication overhead
• Leverage NVIDIA-specific knowledge: skills that understand CUDA, cuDNN, NCCL, and NVIDIA profiling tools

Genie Code and NVIDIA debugging tools are also fully integrated with various Databricks product surfaces, including:

• Notebooks: first-class GPU monitoring, profiling, and debugging in the notebook environment
• MLflow: track GPU metrics and utilization alongside experiments
• Model Serving: diagnose endpoint health and GPU performance, identify opportunities to optimize endpoint mechanics like autoscaling

5. Industry AI: NVIDIA software on governed Databricks data

Every industry faces unique computational challenges shaped by the data it generates and the models it builds. These challenges span everything from analyzing genomes and accelerating drug discovery to optimizing supply chains, interpreting medical images, and simulating factories, robots, and digital twins.

To help solve these problems, NVIDIA has invested heavily in domain-specific accelerated computing libraries and frameworks. We're excited to bring these capabilities directly into the Databricks platform.

Customers can leverage NVIDIA's accelerated computing stack across the end-to-end Databricks experience – from data engineering and experimentation to model development and production workflows; and now domain-specific R&D teams can use NVIDIA's accelerated capabilities without leaving the Databricks platform.

The partnership extends across NVIDIA's accelerated computing libraries and domain frameworks that customers can use with Databricks for industry-specific AI workloads:

Looking Ahead: Building for the Agentic Era

NVIDIA AI infrastructure supports critical layers of AI on Databricks: the GPUs powering training and inference, the Vera CPUs that will power your agent orchestration and data analytics, the NVIDIA Agent Toolkit enabling your agentic applications, and the developer tools that help you get the most out of every compute cycle.

Whether you're a startup experimenting with your first GPU workload in Free Edition, a pharma company running BioNeMo for drug discovery, or an enterprise deploying frontier models at scale, Databricks and NVIDIA together deliver the performance, simplicity, and governance you need.

Get started today: try NVIDIA GPUs in Databricks Free Edition, deploy NVIDIA Agent Toolkit on Databricks Apps, or explore our Foundation Model API powered by NVIDIA accelerated computing.