The best cloud platform for AI and data analytics

BigQuery trains its built-in model types, among them linear and logistic regression, k-means, PCA and ARIMA time-series models, inside the warehouse from one CREATE MODEL statement. Google's guide to managing those models in Model Registry says they register there "without needing to export them," and that built-in path from SQL to a registered model is why we name Google Cloud the best cloud platform for AI and data analytics. It covers the built-in types: deep neural networks, boosted trees, random forests and AutoML start from the same SQL and train in Gemini Enterprise Agent Platform, Google's separate machine learning service, and a query that calls Gemini or Claude is billed by that platform too.

AWS is the better pick when your data already sits in Amazon S3 and Redshift, or when you want OpenAI's GPT-6 Astra and Anthropic's Claude from one model service, Amazon Bedrock, whose model catalog lists both. Microsoft Azure suits a company whose reporting runs on Microsoft Fabric and Power BI and whose data scientists want one Azure Machine Learning workspace for training jobs, pipelines, registered models and endpoints.

What should an AI and data analytics team weigh in a cloud platform?

The FinOps Foundation's guidance on data cloud platforms counts integration and data movement among the things that decide where a workload should run, and it names cross-region transfer as a cost. The same CREATE MODEL statement trains in the warehouse engine on one platform and hands the job to a separate training service on another, so the statement alone does not tell you whether training data leaves the warehouse or which service sends the bill. Ask each vendor to show, for the model type you plan to use, where the training runs.

Amazon Bedrock's model catalog keeps OpenAI's GPT-6 models and its open-weight gpt-oss models in separate groups, and they are different models under one provider's name. A team that has tuned its prompts against one model needs that exact model on the cloud where its data sits, at a release status it can put into production. The FinOps Foundation's guidance on model selection frames the choice as cost per unit of business outcome and tells engineers to weigh "performance and other quality markers" beside cost, so compare the named model and its release status on each cloud.

Data cloud platforms "commonly bill based on activity, such as queries executed, data scanned," the same FinOps guidance on data platforms says, and it recommends tracking cost per query, pipeline or model run. The three warehouses compared here meter different things: data scanned, capacity-hours, or a choice of the two. A unit based on data scanned charges for how much each query reads, and a unit based on capacity charges for how long the warehouse runs, so set the unit and the free allowance against your own query pattern. Model calls made from SQL are billed by the model service on top.

NIST's AI Risk Management Framework says "AI systems should be tested before their deployment and regularly while in operation." For a model that starts in the warehouse, that puts four jobs after the SQL: training, a registry of model versions, pipelines that rerun the work, and monitoring of the deployed model. Each handoff between products is a place where a model version needs an owner, so map which service does each job and where one workspace ends.

Zero trust, in NIST SP 800-207, "assumes there is no implicit trust granted to assets or user accounts based solely on their physical or network location." For an AI workload, that separates the private connection from the perimeter: the connection decides how a request reaches a model endpoint, and the perimeter decides where data is allowed to go afterwards. Location needs the same care, because a global endpoint tells you where the request was sent and leaves open where it was processed. Ask each vendor for the processing location of every service in the workflow.

Google Cloud, AWS and Azure on the five factors

Google Cloud: built-in models train inside BigQuery

Linear and logistic regression, k-means, PCA and ARIMA time-series models train inside BigQuery itself, according to Google's introduction to BigQuery ML. The same page puts deep neural networks, wide-and-deep models, autoencoders, boosted trees, random forests and AutoML in a second group that are "external to BigQuery ML and trained in" Agent Platform, the service Google previously called Vertex AI, though each still starts from a CREATE MODEL statement. Models built elsewhere in ONNX, TensorFlow or XGBoost format can be imported and run for prediction inside BigQuery.

A BigQuery ML model registers to Model Registry from the CREATE MODEL statement, with no export step, and it appears in the registry once training finishes. Remote models are the exception: Google's guide says they cannot be registered. Eligible model types deploy from the registry to an endpoint for online prediction, and the guide lists imported XGBoost models and the two ARIMA_PLUS model types as ones that register and cannot be deployed in Agent Platform.

A remote model created in SQL points a query at Gemini or Claude, and the AI.GENERATE_TEXT function calls it. Claude has to be enabled in Model Garden, with its terms accepted, before that model is created. BigQuery vector search runs in SQL as well, with IVF indexes and the ScaNN algorithm.

Gemini 3.8 Flash is generally available on Agent Platform, and Claude Opus 5.5 and Sonnet 5.5 are generally available there as partner models. The Opus 5.5 page lists availability through United States and Europe multi-region endpoints and a global endpoint. Google's managed open models include OpenAI's gpt-oss 120B, an open-weight model under the Apache 2.0 license that is also generally available.

BigQuery bills on-demand queries per TiB scanned, with the first 1 TiB of query processing and 10 GiB of storage free each month, according to its pricing page. Capacity editions bill per slot-hour, and BigQuery ML runs on the Enterprise and Enterprise Plus editions, with the Standard edition excluded. Remote calls to Gemini or Claude add Agent Platform charges to the bytes BigQuery processes.

Unico, which runs identity verification, moved from Databricks to BigQuery, and Google's case study says its analysts query with standard SQL and that BigQuery policy tags and access controls protect personal data.

Training outside the warehouse runs on Agent Platform's managed training service, which takes applications built on any ML framework. Model Registry is "a central repository" for versions of custom, AutoML and BigQuery ML models, and Model Monitoring runs on demand or on a schedule against tabular models, with an alert when a metric crosses a threshold.

A VPC Service Controls perimeter around Agent Platform stops training data, models, online inference requests and batch results from leaving it. Google's page treats private connectivity as a separate control: Private Service Connect and Private Google Access "don't eliminate public internet accessibility," and the perimeter is what blocks callers outside the allowlist. Inside a perimeter, Agent Platform Pipelines cannot reach third-party internet services such as the public PyPI package registry.

BigQuery ML processes and stages data in the same location as the dataset for every model that is not remote. A global endpoint is available for remote Gemini calls, and Google tells customers with processing-location requirements not to use it, because the region that handles a request cannot be controlled or known. For partner and open models, Google's security controls matrix marks data residency at rest, customer-managed encryption keys, VPC Service Controls and Access Transparency as supported for online prediction, and it excludes preview models.

AWS: Redshift ML on S3, SageMaker and Bedrock

When a Redshift user runs CREATE MODEL, Redshift exports the training data to Amazon S3 and SageMaker AI Autopilot chooses the algorithm and trains the model. The finished model comes back as a SQL function in the cluster. For models that compile to run there, predictions are computed locally with no charge beyond the cluster, and the training is billed as a SageMaker AI line item with S3 storage for the training data.

The statement names the S3 bucket and IAM role to use, and by default it deletes the training data from S3 once training finishes; an optional KMS key encrypts the files while they are there. With enhanced VPC routing turned on, traffic between Redshift ML and the bucket goes through your VPC. A team that already runs its permissions and encryption keys through S3 and SageMaker gets Redshift ML on that same setup.

A model already trained in SageMaker AI can be imported for local inference when it is an XGBoost, MLP or Linear Learner model, and Redshift can call a SageMaker endpoint in the same Region and account for remote inference. CREATE EXTERNAL MODEL does the same for Amazon Bedrock: it creates a SQL function that sends a prompt to a Bedrock model, and Bedrock bills the call. Athena, the separate AWS query engine over S3, runs inference from SQL against an existing SageMaker endpoint.

Amazon announced GPT-6 Astra as generally available on Bedrock, and Bedrock's model catalog lists further GPT-6 models beside it, Anthropic's Claude Opus 5.5 and Sonnet 5.5, and Google's open Gemma models. The GPT-6 Astra model card and the Claude Opus 5.5 card show both models served through geographic or global cross-Region profiles on the standard Bedrock runtime endpoint, with an in-Region option on a second endpoint in a short list of Regions. A geographic profile keeps requests inside a geography such as the United States, and a global profile routes them to any commercial Region.

Redshift Serverless bills in RPU-hours, metered per second with a 60-second minimum, and its pricing page gives first-time Serverless users a $300 credit that expires after 90 days. Athena bills by the TB of data scanned.

SageMaker Unified Studio is a single AWS environment for data and AI development, and it brings together tools from Redshift, Athena, Bedrock and SageMaker AI among others. Its menu puts a query editor and notebooks beside training jobs, registered and foundation models, MLflow and inference endpoints, and an administrator decides which Bedrock models each user can reach. SageMaker AI adds a Model Registry with versions and approval status and Pipelines for orchestration. Model Monitor is "no longer open to new customers"; existing customers keep it, and AWS names open-source monitoring with CloudWatch and QuickSight as the replacement.

Bedrock traffic runs over PrivateLink interface endpoints when you set them up, each with an attachable endpoint policy. AWS defines a data perimeter as a set of permissions guardrails built from service control policies, resource control policies and VPC endpoint policies, which the customer configures. SageMaker AI training jobs and hosted models run in your VPC when configured to, and its containers are internet-enabled by default until network isolation is switched on.

Insurer Generali uses Redshift, SageMaker AI and Bedrock, according to an AWS case study.

Microsoft Azure: Fabric SQL and the Azure Machine Learning workspace

In Azure Synapse, model "training takes place on Apache Spark pools," Microsoft's overview says, and SQL pools run predictions against the data where it sits. One documented route into the SQL warehouse is a scoring wizard: a model is trained and registered in Azure Machine Learning, and the wizard copies it in ONNX format into a dedicated SQL pool table and generates the T-SQL that predicts with it.

Fabric Warehouse adds T-SQL functions for sentiment, classification, extraction, summarization and translation. Microsoft says they "call external AI APIs to process text," marks them Preview and requires a paid Fabric capacity, F2 or higher or a P edition. A Fabric reporting team reaches those functions from the warehouse SQL it already writes.

Microsoft's catalog of models sold directly by Azure lists gpt-6-astra, gpt-6-sol, gpt-6-luna and gpt-6.1-sol under Azure OpenAI. Claude Opus 5.5 and Sonnet 5.5 are generally available in Foundry, with Anthropic as seller and operator. Microsoft's hosting comparison describes two routes: Azure-hosted Claude processes on Azure infrastructure and stores data at rest in the selected Azure geography, and Anthropic-hosted Claude processes on Anthropic infrastructure and can do so outside Azure.

Synapse charges serverless SQL by the TB of data processed, dedicated SQL pools by DWU and hours running with storage billed separately, and Spark pools by vCore-hour. Fabric runs on purchased capacity, and its AI functions and Foundry Tools draw on that capacity.

For models sold by Azure, the deployment type sets where prompts are processed, for the types a given model supports: a Standard deployment processes in its own region, and the Data Zone and Global types widen that to a named zone or to any region where the model is deployed. Microsoft's data privacy page says prompts and completions for these models are not used to train base models. An Azure Machine Learning workspace takes a private endpoint, and Microsoft says that alone does not secure the workflow end to end: the storage account, Key Vault and container registry behind the workspace are secured separately.

The Azure Machine Learning workspace keeps training jobs, reusable pipelines, registered models and online endpoints in one place, with its own networking, identity, access-control and encryption settings. A registry shares models, environments and components across workspaces and promotes a model to an endpoint in another workspace, and compute and training data stay with the workspace that ran the job. Monitoring covers a managed online endpoint once production-data collection is enabled; for batch endpoints and models deployed elsewhere, the customer collects that data and registers it as a data asset. BRF, a consumer goods company, built its recommendation work on Azure Machine Learning, according to a Microsoft customer story.

Which cloud platform should an AI and data analytics team choose?

Google Cloud is our pick for AI and data analytics, because BigQuery's built-in model types train where the data already sits and the finished model registers in Model Registry with no export step. Redshift ML and the Synapse scoring wizard each document a route in which training happens in a second service and the model comes back to the warehouse for prediction.

That advantage stops at the built-in types. The model types that train in Agent Platform, and SQL calls to Gemini or Claude, carry that platform's charges. We weigh the built-in path more for this buyer because it lets the people who query the warehouse train a model on its tables and register it from the same statement. Work that outgrows those model types moves to a dedicated training service, and each of the three clouds has one.

AWS is the better pick when the data is already in S3 and Redshift. Redshift ML trains through the S3 bucket and SageMaker account the team already governs and returns predictions in the cluster. Bedrock adds GPT-6 Astra and Claude behind one model service that Redshift SQL can call.

Azure is the better pick for a company whose reporting runs on Fabric and Power BI and whose data scientists want one workspace. Fabric puts text AI functions in warehouse SQL, and the Azure Machine Learning workspace holds training jobs, pipelines, registered models and endpoints, with a registry to promote a model from one workspace to the next.

Other cloud and data platforms