Edit

Quota reservations for Microsoft Discovery

Microsoft Discovery requires specific quotas across multiple Azure services to function effectively. These quotas must be secured before deploying Microsoft Discovery infrastructure components. Proper quota planning ensures optimal performance and prevents deployment failures during infrastructure setup.

The primary quota categories include:

  • Virtual Machine SKUs: For supercomputer node pools and computational workloads
  • Azure Cosmos DB throughput (RU/s): For Discovery workspace and Discovery project resources
  • Chat completion and text embedding models: For Azure OpenAI and Azure AI Foundry services
  • Bookshelf service infrastructure: For Azure AI Search, Azure Container Apps, Azure SQL DB, and indexing nodepools

Prerequisites

Before requesting quota increases, ensure you have:

  • An active Azure subscription with the Microsoft Discovery resource provider registered
  • Contributor or Owner role on the Azure subscription
  • Understanding of your planned Microsoft Discovery deployment scale and requirements
  • Access to the Azure portal and Azure CLI (if using programmatic quota requests)
  • Knowledge of your target Azure regions for deployment

Virtual Machine SKU quotas

Standard VM SKUs are required for Microsoft Discovery infrastructure components, including supercomputer node pools, storage systems, and management services.

Required VM SKU families

Microsoft Discovery supports various VM SKU families for different computational workloads. For more information, see Azure VM SKU Families.

The following table lists sample VM SKU families supported in preview:

VM SKU Family Recommended SKUs Use Case
D-series v5/v6 Standard_D4s_v5, Standard_D4s_v6 Enterprise-grade applications, relational databases, in-memory caching, data analytics
NC-family (GPU) Standard_NC4as_T4_v3, Standard_NC8as_T4_v3, Standard_NC16as_T4_v3, Standard_NC64as_T4_v3, Standard_NC24ads_A100_v4, Standard_NC48ads_A100_v4, Standard_NC96ads_A100_v4 Compute-intensive AI/ML workloads, graphics-intensive applications, visualization, deep learning training
NV-family (GPU) Standard_NV6ads_A10_v5, Standard_NV12ads_A10_v5, Standard_NV24ads_A10_v5, Standard_NV36ads_A10_v5, Standard_NV36adms_A10_v5, Standard_NV72ads_A10_v5 Virtual desktop (VDI), single-precision compute, video encoding, and rendering, remote visualization
ND-family (GPU) Standard_ND40rs_v2 Large memory compute-intensive workloads, large memory graphics-intensive applications, distributed deep learning

VM vCPU quota is reserved per subscription. You can check vCPU quota by following the guidance at Check vCPU quotas.

Depending on the resources you plan to create in your subscription, allocate vCPU quotas accordingly. If you need GPU support for your tools, follow the same process with VM SKUs that include GPU support. All supported VM SKUs are listed in the preceding table.

For more information, see Increase VM-family vCPU quotas.

Azure Cosmos DB throughput quota

Microsoft Discovery uses Azure Cosmos DB. Cosmos DB throughput is measured in RU/s (Request Units per second) and should be planned to ensure both successful resource creation and steady runtime performance.

To learn more, see Request units in Azure Cosmos DB.

Cosmos DB account quota behavior

  • There's no per-subscription quota limit on RU/s.
  • Throughput availability is managed per Cosmos DB account.
  • Discovery Platform manages the Cosmos DB, which uses throughput within the default assignment range.
  • If there's a quota issue due to region-level restrictions (for example, a high-demand region), raise a support ticket to request the appropriate extension.

For more information, see Azure Cosmos DB service quotas.

Required throughput

Both the Discovery workspace resource and each Discovery project resource require autoscale throughput to be available.

Resource Minimum RU/s required Maximum RU/s (autoscale) Notes
Discovery workspace 2,400 RU/s 4,000 RU/s Cosmos DB service support autoscale
Discovery project 400 RU/s 4,000 RU/s Cosmos DB service support autoscale

Operational guidance

  • If the minimum RU/s isn't available, you might see resource creation failures.
  • If the maximum RU/s can't be fulfilled, the platform might experience performance degradation under load.

Example sizing

For a workspace with 10 projects:

  • Minimum: 2,400 + (400 × 10) = 6,400 RU/s
  • Maximum: 4,000 + (4,000 × 10) = 44,000 RU/s

Chat completion and text embedding model quotas

Azure OpenAI and Azure AI Foundry quotas are essential for Microsoft Discovery's AI-powered features, including Copilot, Bookshelf, Discovery Engine, and natural language processing.

Model TPM requirements summary

The following table shows the total TPM required per model for a workspace with a single Bookshelf instance. For each Bookshelf, add the corresponding per-instance Bookshelf TPM values. For detailed per-service breakdowns, see Chat completion and text embedding model quotas.

Model Total minimum TPM Total recommended TPM Contributing services
GPT-5.4 450,000 2,000,000 Discovery Engine + Agents (Copilot Service)
GPT-5.2 200,000 2,000,000 Bookshelf
GPT-5 Mini 2,000,000 10,000,000 Bookshelf
Text Embedding 3 (Small) 2,000,000 2,000,000 Bookshelf

Tip

For deployments with multiple Bookshelves, multiply the Bookshelf per-instance TPM by the number of Bookshelf instances and add it to the workspace-scoped totals. For example, a workspace with three Bookshelves at recommended TPM requires: GPT-5.2 = 2,000,000 × 3 = 6,000,000 TPM.

Consolidated model TPM requirements

The following table provides a consolidated view of model TPM requirements across all Microsoft Discovery services. Use this table to plan your total Azure OpenAI quota needs for a single workspace deployment.

Per-service model TPM breakdown

Model Service Minimum TPM Recommended TPM Scope
GPT-5.4 Discovery Engine 250,000 1,000,000 Per workspace
GPT-5.4 Agents (Copilot Service) 200,000 1,000,000 Per workspace
GPT-5.2 Bookshelf 200,000 2,000,000 Per Bookshelf instance
GPT-5 Mini Bookshelf 2,000,000 10,000,000 Per Bookshelf instance
Text Embedding 3 (Small) Bookshelf 2,000,000 2,000,000 Per Bookshelf instance

Discovery Engine models

Discovery Engine requires two GPT model deployments. The first is created automatically during workspace provisioning for cognition (reasoning and task planning). The second must be created manually for task validation.

Model Deployment name Minimum TPM Recommended TPM Purpose
GPT-5.4 (auto-provisioned) 250,000 1,000,000 Cognition reasoning, task planning, and answer generation
GPT-5.4 gpt-5-4 200,000 250,000 Task validation. evaluates agent results against validation requirements
  • Cognition model: Deployed automatically during workspace creation with an initial quota of 250,000 TPM. This deployment is dedicated to the Discovery Engine and isn't shared with other services. After workspace creation, increase the TPM to the recommended 1,000,000 TPM for optimal performance. For instructions, see How to update quota assigned to a model deployment.
  • Validation model: You must manually create a chat model deployment named gpt-5-4 using model gpt-5.4. Without this deployment, the Discovery Engine can't validate task results and won't start. See Create Chat Model Deployment for setup instructions.

Important

Both deployments consume GPT-5.4 quota in your subscription. Ensure you have at least 500,000 TPM of GPT-5.4 quota available (250,000 minimum per deployment).

Bookshelf service quotas

Bookshelf requires infrastructure resources and LLM model quotas across enrichment, indexing, search, and search tuning.

Enrichment infrastructure

Bookshelf uses Azure AI Search for enrichment (built-in skills) and incremental enrichment watermarks. An Azure AI Search instance is created when the knowledge base is deployed.

Resource SKU Lifecycle
Azure AI Search Standard S1 Always on

Enrichment processing is variable and depends on the number of pages, files, images, and tables in your documents.

Indexing quotas

Bookshelf supports three index size tiers based on text content volume:

Index size Content volume Supercomputer nodepool SKU
Small 200 MB TXT Standard_D48s_v6 (192 GB)
Medium 500 MB TXT Standard_D128s_v6 (512 GB)
Large 1 GB TXT Standard_D192s_v6 (768 GB)

Supercomputer nodepools for indexing are on-demand and only consumed during indexing operations.

LLM model quotas

Bookshelf uses three models for indexing and search operations. Ensure that each Bookshelf has at least the minimum quota allocated.

Model Minimum TPM Preferred TPM Purpose
Text Embedding 3 (Small) 2,000,000 2,000,000 Document indexing and embedding generation
GPT-5.2 100,000 2,000,000 Query decomposition and answer generation
GPT-5 Mini 2,000,000 10,000,000 Primary search model

GPT-5 Mini has the highest preferred quota because a single search query can consume approximately 800K–1M tokens.

Search infrastructure

Bookshelf deploys always-on infrastructure for domain specific knowledge search when a Bookshelf is created.

Resource Size tier Configuration
Azure Container Apps (dedicated profile) Small E4 vCPU, 32-GiB memory
Azure Container Apps (dedicated profile) Medium E8 vCPU, 64-GiB memory
Azure Container Apps (dedicated profile) Large E16 with 16 vCPU, 128-GiB memory

Scaling behavior

  • Indexing (Text Embedding)—Scales with dataset size. Large datasets across multiple Bookshelves might require millions of TPM. High embedding quota is easy to obtain.
  • Querying (GPT models)—Independent of dataset size. Driven by concurrent users, search frequency, and relevance budget. Quota is shared at the subscription and region level across all Bookshelves.
  • Fixed infrastructure—Azure AI Search, Azure Container Apps dedicated profile, and Azure SQL DB are always-on resources created at Bookshelf deployment.
  • Variable components—Enrichment processing, embedding generation, and model inference scale with usage.

Copilot Service models

The Copilot Service requires the following GPT model for agents. Ensure that each workspace has at least the minimum TPM allocated. Quota is reserved per workspace.

Model Minimum TPM Recommended TPM Purpose
GPT-5.4 200,000 1,000,000 Conversation and reasoning for Copilot agents

Note

GPT-5.4 is the recommended model. Other models available in the Azure AI Foundry Model Catalog can also be supported. If you choose a different model, ensure the corresponding model quota is reserved in your subscription.

By default, Global Standard is the deployment mode for model deployments. If data residency is required, Data Zone Standard is supported as an alternative deployment mode. For more information, see Data Zone Standard deployment type.

Supported model variants

Microsoft Discovery supports the following model families for agent reasoning, orchestration, and knowledge retrieval. You can use these models as alternatives or upgrades depending on your use case and regional availability.

GPT model family

Model Use case Region availability
GPT-5.2 Bookshelf query decomposition and answer generation All Azure OpenAI regions
GPT-5.4 Discovery Engine, validation agent, BYO agents (recommended) All Azure OpenAI regions
GPT-5.5 Discovery Engine, validation agent, BYO agents (latest) Limited: eastus2, northcentralus, southcentralus, westus3, polandcentral, swedencentral

For the full GPT model region availability matrix, see Region availability for Foundry Models sold by Azure.

Grok model family

Grok models are available through Azure AI Foundry and offer less restrictive content filtering for biology-vertical queries compared to GPT models, making them suitable for life sciences research workflows that require broader reasoning over biological and biomedical content.

Model Use case Notes
Grok-4.20-reasoning Complex scientific reasoning tasks requiring chain-of-thought in biology and life sciences domains Extended reasoning capabilities with fewer content restrictions on biology-related queries
Grok-4.20-non-reasoning Fast inference for BYO agents in biology-focused workflows Standard inference without chain-of-thought; fewer content restrictions on biology-related queries

Note

Grok models are supported as BYO (bring your own) model deployments in Discovery. If you choose a Grok model, ensure the corresponding model quota is reserved in your subscription and that the model is available in your target region. Check the Azure AI Foundry Model Catalog for current availability.

Requesting Azure OpenAI quota

To request or increase model quota:

  • Sign in to the Azure AI Foundry Portal.
  • Navigate to Management Center > Quota.
  • Select the correct subscription and the region where you plan to deploy Microsoft Discovery.
  • Select Request quota for the desired model and fill in the request form with the model name, deployment type (Standard), and the TPM values from the tables in this article.

For more information, see Request quota for Azure OpenAI models.

Regional quota considerations

Choose regions based on quota availability and proximity to your users, and the locations where the platform is available.

Before requesting quotas, verify regional availability:

# Check VM quota availability by region
az vm list-usage --location "swedencentral" --query "[?contains(name.value, 'cores')]"

# Check Azure OpenAI model availability
az cognitiveservices model list --location "swedencentral" --kind "OpenAI"

Quota request best practices

  • Request quotas 2–4 weeks before deployment to allow for processing time. Standard requests typically take 1–3 business days; large quota requests might take 5–10 business days.
  • Plan for multiple regions in case primary region quotas are unavailable.
  • Set up Azure Monitor alerts at 80% quota utilization and configure notifications to platform administrators.
  • Integrate with cost management by linking quota monitoring with cost management, setting up spending alerts for Azure OpenAI usage, and implementing budget controls for quota-intensive resources.

Acronyms

Acronym Full term Definition
VM Virtual Machine An on-demand, scalable compute resource in Azure that emulates a physical computer.
SKU Stock Keeping Unit A specific configuration or pricing tier of an Azure resource, such as a VM size or storage level.
VDI Virtual Desktop Infrastructure A technology that hosts desktop environments on a remote server, delivered to users through services like Azure Virtual Desktop.
GPT Generative Pretrained Transformer A family of large language models used in Azure OpenAI Service for text generation, reasoning, and conversation.
TPM Tokens Per Minute A unit measuring the rate of token consumption allocated to an Azure OpenAI model deployment.