Databricks Makes AI Runtime Generally Available for GPU Training

Databricks on October 21, 2026, announced the general availability of AI Runtime (AIR), a fully managed, serverless GPU platform built into the Databricks Lakehouse for training and fine-tuning deep learning models on enterprise data.

Databricks said hundreds of companies, including Robinhood, PicPay, Merck, Flo Health, TinyFish, and FactSet, already use AIR to train and ship deep learning models to production, with use cases spanning healthcare, manufacturing, and consumer-facing agents. The company said its own Databricks AI Research team uses the platform to train the models behind KARL and AI Parse, and it described AIR as battle-tested by more than 100,000 GPU-hour training runs.

The company positioned the release as an answer to AI teams spending more time securing GPU capacity, managing failures, and waiting for experiments to launch than improving their models. Running GPU infrastructure separately from the Lakehouse adds governance gaps and puts sensitive training data at risk, the announcement said.

Managed GPU Infrastructure

Databricks takes responsibility for procuring GPUs, configuring clusters, and managing networking, so customers can begin training without operating a full GPU platform of their own. The company said it runs dozens of GPU and network health checks before and during training so workloads can scale to hundreds of GPUs reliably; the platform validates GPUs before scheduling them, monitors active workloads continuously, and periodically tests idle GPUs to detect and remove unhealthy nodes.

AIR moves training data through built-in loaders that stream and cache it, reducing the time jobs spend waiting on I/O. A companion tool, UCVolumeDataset, is meant to raise data throughput without a custom loading pipeline. Users can reach GPUs from a command-line interface or work interactively inside notebooks and IDEs. Databricks said a first workload requires a single CLI command, and that its AIR Skills feature migrates workloads from Kubernetes or SLURM clusters with one natural language prompt.

AIR integrates natively with Lakeflow, the company’s pipeline orchestration product, so GPU workloads stay in sync with upstream and downstream pipelines. Moving a workload to production requires a few extra lines of YAML, according to the announcement, which also described built-in priority and pre-emption controls that minimize idle GPU time and monitoring through tools such as Genie ZeroOps. The AI Runtime product page lists on-demand NVIDIA GPUs with multi-region capacity, distributed training across multiple GPUs and nodes with managed RDMA networking, optimized checkpointing for long-running jobs, and support for custom Docker containers, including NVIDIA NGC containers.

Kostia Omelianchuk, an applied research scientist and area tech lead at Superhuman, said in the announcement that the platform let his team shift large-scale training from one node with eight A100 GPUs to 16 H100 GPUs across multiple nodes. Training time fell from roughly seven days to 48 hours, which he described as a 3.5x speedup.

Seconds-Level Launches and Debugging

Databricks said AIR delivers seconds-level launch latency even on large RDMA-enabled clusters with dozens of GPUs. The service uses warm serverless GPU pools and lazy image loading, starting workloads in seconds while the remaining image data loads in the background. Users can run pre-built environments or bring their own containers, and environment caching means environments are built only once. The company described AIR as compatible with all major frameworks, and the product page names PyTorch and Ray among the supported options.

Databricks said a new observability experience in the MLflow experiment-tracking tool streams logs seconds after a run is submitted when users add the watch flag to an AIR command, and a new logs viewer is built specifically for multinode training. Genie Code is embedded across the experience with pre-packaged agent skills for debugging issues and optimizing GPU training performance. The company said those skills carry insights from its research and performance teams, including how it reduced Qwen training time by 6x on AIR.

Joe Tyler, lead AI researcher at Sonar, said research engineers there used AIR to assemble LLM training pipelines and iterate on hypotheses quickly, with visibility into each stage of the pipeline. Nikhil Sunderraj, principal machine learning engineer at FactSet, said the time needed to set up, train, and deploy a custom Text To Formula model fell from days to hours.

Security, Governance, and Getting Started

Databricks said AIR operates within the governance perimeter of Databricks and Unity Catalog, which enforces consistent read and write permissions on data access, checkpoints, and logs for GPU workloads. The company said AIR provides VM-level isolation rather than isolation mechanisms that expose the host kernel, and that it supports PrivateLink and serverless egress controls, backed by its serverless security investments. The announcement cited the OpenAI Hugging Face incident as underscoring the risks of managing GPU infrastructure separately from governed enterprise data.

AIR usage and costs appear in Databricks billing system tables alongside other Databricks workloads, with workload-level usage records showing which training runs consume GPU budget.

Robert Macy III, a machine learning engineer at Robinhood, said AIR helps his teams get from experimentation into training faster, with governance and data controls built in. Eran Barash, vice president of artificial intelligence and machine learning at QuantHealth, said his team scales distributed training on serverless GPUs across millions of records within existing data workflows. Zach Richardson, head of AI at TinyFish, said his team can train on GPUs, evaluate results with MLflow, and serve the company’s web-native model without standing up separate infrastructure for each stage.

Databricks pointed new users to a quickstart guide, template notebooks, and starter guides, along with a clonable ai-runtime-cookbook repository on GitHub that it said can launch a first GPU workload with one CLI command. The company directed customers to their account teams for more information.