Published November 13, 2026
MLOps software runs machine learning models after they are trained. It records experiments, keeps a registry of model versions and who approved them, deploys models to endpoints and watches them in production for drift and falling quality. This edition covers the platforms that operate classical models, fine-tuned language models and the applications built on them, not the tools for preparing data and training models.
Where the market stands
CoreWeave completed its acquisition of Weights & Biases on May 5, 2025. On September 30, 2026, CoreWeave launched CoreWeave Forge, which unifies Weights & Biases Models, OpenPipe's post-training work and the open-source marimo notebook project with CoreWeave's own services. Since that date, Weights & Biases SaaS accounts sign in through Forge. Self-hosted and dedicated-instance accounts are unchanged.
AWS renamed Amazon SageMaker to Amazon SageMaker AI on December 3, 2024. On December 2, 2025, it added a serverless MLflow capability to SageMaker AI.
Google announced Gemini Enterprise Agent Platform on April 22, 2026. Google says it brings together the model selection, model building and tuning services of Vertex AI with new agent features. Google's documentation now calls the pipeline service Agent Platform Pipelines.
At Microsoft, support for version 1 of the Azure Machine Learning CLI extension ended on September 30, 2025, and support for version 1 of the Python SDK ended on June 30, 2026. Microsoft says new features launch only in version 2.
Kubeflow moved to the Graduated level of the Cloud Native Computing Foundation on July 24, 2026, and the foundation announced the graduation on August 17, 2026. Kubeflow had entered the foundation as an incubating project on July 25, 2023.
The standard underneath
MLflow is open source under the Apache License 2.0. MLflow 3.0, announced on June 11, 2025, added tracing for more than 20 generative AI libraries, evaluation with LLM judges and prompt versioning. Those features are in open-source MLflow and in Databricks' managed service. The current release is 3.17.1, published on October 13, 2026.
The cloud suites accept MLflow. Microsoft says Azure Machine Learning workspaces are MLflow-compatible and can act as the tracking server for MLflow code running anywhere. Its version 2 SDK has no logging functions of its own, and Microsoft recommends MLflow logging instead. SageMaker AI offers managed MLflow, and its serverless MLflow works with SageMaker Model Registry and SageMaker Pipelines.
If you want models and run history you can move between platforms, log them through the MLflow APIs rather than a vendor SDK.
Five kinds of product
This edition groups the products into five kinds.
- Cloud suites. SageMaker AI includes Pipelines, Model Registry and Model Monitor. Azure Machine Learning offers managed online and batch endpoints and a Responsible AI dashboard. Google's Agent Platform Pipelines runs pipelines written with the Kubeflow Pipelines or TensorFlow Extended (TFX) frameworks, alongside a Model Registry and Model Monitoring.
- Part of a data platform. Databricks runs managed MLflow and Databricks Model Serving next to its data platform.
- Independent enterprise platforms. Domino sells Domino Cloud, a single-tenant managed service, and self-managed editions for a VPC or on premises. DataRobot lists on-premises, virtual private cloud and SaaS deployment.
- Tracking first. Weights & Biases, now part of CoreWeave Forge, and Comet record and compare experiments. Comet also publishes Opik, its LLM tracing and evaluation tool, under the Apache License 2.0.
- Assembled from open source. MLflow for tracking and the registry, Kubeflow for pipelines, training and serving on Kubernetes, Feast as a feature store and Evidently for evaluation and monitoring. MLflow, Kubeflow Pipelines, Feast and Evidently are under the Apache License 2.0. Evidently lists more than 100 built-in metrics, from data drift detection to LLM judges.
What each costs where prices are public
Figures are list prices as published on October 17, 2026.
- SageMaker AI. Pay as you go, with no minimum fee. AWS offers serverless MLflow at no additional charge. MLflow tracking servers are billed per hour by server size, plus the metadata they store. Model Monitor is billed on the instances it runs on, and built-in rules include up to 30 hours of monitoring at no charge.
- Azure Machine Learning. No additional charge for the service. You pay for compute and for the other Azure services the workspace uses, including Blob Storage, Key Vault, Container Registry and Application Insights.
- Gemini Enterprise Agent Platform. Models in the Model Registry cost nothing until deployed to an endpoint or used for batch prediction. Pipelines charge $0.03 per run, plus the compute the pipeline uses. Model Monitoring costs $3.50 per GB of data analyzed. Google prints the TensorBoard price as "$10 GiB/month" for storage of logs and metrics. Google lists these prices in USD.
- Databricks Model Serving. Metered in Databricks Units (DBUs) per hour. One T4 GPU uses 10.48 DBUs an hour and one H100 GPU uses 100. Databricks says its prices vary by region and cloud provider, and Microsoft sets the price of Azure Databricks.
- CoreWeave Forge (Weights & Biases). Free covers up to 5 seats and 5 GB of storage a month. Pro starts at $60 a month for up to 10 seats and 100 GB, and is limited to teams with fewer than 50 employees. Enterprise is custom.
- Comet. Free covers 1 user and 100 GB. Pro costs $19 per user per month for up to 10 users, with 1,500 training hours and 500 GB included. Extra training hours cost $1 each. Enterprise is custom.
- Domino and DataRobot. Domino's pricing page lists no prices. Domino sells by quote, directly or through the AWS and Azure marketplaces. DataRobot's platform page lists no prices.
Training, serving and monitoring run on compute billed separately from any platform fee. Model the cost from your number of production models, retraining schedule and endpoint hours, and count idle endpoints.
What self-hosting requires
- MLflow needs a tracking server, a database for run metadata (PostgreSQL, MySQL, SQLite or MSSQL) and object storage for artifacts. Access control is switched on with a server flag. MLflow's documentation says it ships no default admin password, and the server refuses to start until one is set.
- Weights & Biases Self-Managed runs on Kubernetes v1.34 or newer with an ingress controller. It needs a supported MySQL version, with 8.4 recommended, a single-node Redis 7.x and object storage that supports pre-signed URLs and CORS.
Budget the people who will patch, back up and upgrade these services.
What a deployment requires
Plan the rollout by what happens to a model after it is trained.
- Put experiment tracking and a versioned registry in place first. Make one real model reproducible from data to endpoint, with its approval recorded.
- Build the training, evaluation and deployment pipeline in the CI system your engineers already use. Include a canary or shadow rollout and a rollback you have tested.
- Turn on drift and quality monitoring. Join predictions to ground truth when labels arrive late, and set the alert that triggers retraining.
- Rehearse an incident: feed a deployed model shifted data and time how long the platform takes to flag it and roll back.
- Bring language models and agents onto the same registry and pipeline, with prompt versions and evaluation sets.
Questions to settle first
Settle these before choosing a platform.
- Which cloud or data platform holds your training data and features today, and how much portability will you give up to stay close to it?
- Can you trace a production prediction back to the endpoint, model version, training run and data that produced it?
- If you use Weights & Biases or Vertex AI, what does the contract say about price, roadmap and sign-in now that they sit inside CoreWeave Forge and Gemini Enterprise Agent Platform?
- Will language models and agents share the registry and monitoring with classical models, or run on separate tools?
- Who will run the database, object storage and upgrades if the platform is self-hosted?
Ask to contribute
Vendors in this market can ask to answer the six questions for State of Play 2026: MLOps. Use an email address on your company’s own domain: answers come from the company itself, not from an agency. We email a link to confirm the request.