Microsoft Azure is a cloud computing platform provided by Microsoft. Its AI capabilities are provided externally in the form of cloud services, covering large model inference, model training and fine-tuning, retrieval enhanced generation (RAG), content security, and related engineering management capabilities. Azure's AI service is not a single product, but consists of multiple independent services, which are responsible for different responsibilities such as model invocation, data retrieval, training deployment, and compliance control.

In terms of billing methods, Azure's AI services mainly adopt two modes: pay-as-you-go and reserved capacity. The price is closely related to the selected model, deployment area, call scale and service form.

Azure OpenAI Service: the main entrance to large model capabilities

Azure OpenAI Service is the most direct service for large model applications on the Azure platform, and is used to call the OpenAI series model capabilities in enterprise cloud environments. Through this service, developers can use capabilities such as text generation, dialogue, multi-modal input, vector generation, and image generation and integrate them into their own systems.

In actual use, this service usually serves as the "generation core" of the application, responsible for reasoning about user input and returning results. The billing method is based on Token, distinguishing input Token and output Token. Different models and context lengths correspond to different unit prices. For scenarios that require stable throughput and concurrency capabilities, Azure also provides a Provisioned Throughput mode, which is billed by the hour and is used to lock model computing resources.

Azure AI Foundry: model hosting and unified management portal

As the number of models and usage scenarios increase, Azure centralizes model hosting and management capabilities into the Azure AI Foundry system. The system provides model catalogs, managed inference interfaces, and integration capabilities with the Azure permissions and auditing system for unified management of basic models from different sources.

In actual applications, AI Foundry usually plays the role of "model layer" to standardize the model calling method and facilitate integration with Agent, application orchestration or other Azure services. The billing methods of different models depend on the model itself. Common forms include billing by token or billing by instance or throughput.

Azure AI Search: Search service supporting RAG scenarios

In large model applications, Azure AI Search is often used to build retrieval augmented generation (RAG) capabilities. The service provides document indexing, full-text search, vector retrieval, and semantic relevance sorting to locate problem-relevant content in enterprise data.

The billing method of Azure AI Search is different from the model service. It uses Search Unit (SU) as the capacity unit and is billed continuously according to the selected service level. Capacity configuration is usually related to document size, query concurrency and retrieval complexity, so in RAG scenarios, this service is often a long-running basic component.

Azure Machine Learning: training, fine-tuning and deployment capabilities

For scenarios where you need to train or fine-tune the model yourself, Azure provides the Azure Machine Learning service to manage training tasks, model versions, and deployment processes. The service supports stand-alone and distributed training, and can directly deploy the trained model as an online or batch inference service.

At the billing level, Azure Machine Learning does not define a fixed price separately, but charges based on the underlying computing resources, including the usage time and scale of CPU, GPU, storage and network resources. Therefore, the cost is mainly determined by the selected aircraft model, operating time and the size of the training mission.

Content security and compliance: Azure AI Content Safety

In the process of launching large-model applications, content security is a common basic requirement. Azure provides the Azure AI Content Safety service for instrumentation of model inputs and outputs, covering text and image content. This service is typically used as a standalone API, embedded before and after the build process.

The billing method is based on the call scale and is billed based on the number of text records or images. Free tiers and standard paid tiers are provided for application scenarios of different scales.

Overview of common services and billing dimensions

In actual deployment, Azure's AI services are often used in combination, and there are obvious differences in the billing dimensions of different services:

Main billing dimensions of services

Azure OpenAI Service input/output token, or reserved throughput by hour

Azure AI Foundry depends on model type, by Token or instance

Azure AI SearchSearch Unit (SU), ongoing billing

Azure Machine Learning computing resource usage time

Azure AI Content Safety Number of text records or number of images