Understanding Azure AI Foundry Rate Limits and Concurrent User Capacity

Question

Understanding Azure AI Foundry Rate Limits and Concurrent User Capacity

Mohamed, Rihan 60

I'm investigating rate limits and capacity constraints for an AI assistant built using Azure AI Foundry.

During a training session, multiple users accessed the assistant simultaneously and some users experienced errors.

I would like to understand:

How can I check the applicable Requests Per Minute (RPM) and Tokens Per Minute (TPM) limits?

How does concurrent user activity affect these limits?

How can I identify whether errors are caused by rate limits or capacity constraints?

What are the recommended approaches to handle high concurrent usage, such as retries, request queuing, or increasing capacity?

Are there any best practices for preventing rate-limit errors during peak usage?

Any guidance or documentation would be appreciated.

Sai Kiran Mudavath 90 Reputation points Microsoft External Staff Moderator

2026-09-29T06:42:00.41+00:00
Hi @Mohamed, Rihan ,

When multiple users access an Azure AI Foundry assistant at the same time, the resulting traffic can be affected by the deployment's RPM/TPM limits, burst traffic, and in some cases temporary service-side capacity constraints. The exact behavior depends on the model and deployment type.

1. Checking RPM/TPM limits

For Azure OpenAI models, the TPM assigned to a deployment maps to its token-per-minute rate limit, and an RPM limit is also enforced based on the model's quota allocation. You can review the quota assigned to the specific model/deployment in the Foundry portal under Management → Quota. It is important to check the quota allocated to the deployment receiving the traffic, not only the overall subscription quota.

You can also inspect the rate-limit headers returned by the API, including:

x-ratelimit-limit-requests

x-ratelimit-remaining-requests

x-ratelimit-limit-tokens

x-ratelimit-remaining-tokens

retry-after-ms

These headers provide useful information about the effective request/token limits and how long to wait after throttling.

2. Effect of concurrent users

Concurrent users generate simultaneous requests against the same deployment. A sudden burst can trigger throttling even when the total number of requests measured over a full minute appears to be below the documented RPM.

Azure OpenAI evaluates request rates over shorter intervals as well, so distributing requests evenly is important. In addition, the rate-limit token calculation takes request characteristics such as the prompt and max_tokens into account. Setting max_tokens unnecessarily high can therefore consume more rate-limit budget than expected.

3. Distinguishing rate-limit throttling from capacity-related throttling

A 429 Too Many Requests response does not by itself prove that the configured TPM/RPM quota has been exhausted.

Microsoft documents different 429 scenarios:

Rate limit exceeded: the request exceeded the deployment's allocated RPM/TPM. In this case, reduce/smooth traffic, rebalance quota, or request additional quota.

System capacity throttling: the service is temporarily unable to process the request because of backend capacity constraints. In this case, retry according to the retry-after-ms value. If the condition persists for a workload requiring predictable throughput, Provisioned Throughput can be considered.

Temporary rate-limit adjustment: the effective rate limit can temporarily be lower than the configured deployment quota during periods of high demand. Comparing x-ratelimit-limit-tokens in the response with the deployment's configured TPM can help identify this situation.

4. Handling high-concurrency workloads

We recommend:

Implementing retries with exponential backoff and honoring retry-after-ms when returned.

Avoiding large request bursts and gradually ramping up traffic.

Reducing unnecessary prompt size and setting max_tokens close to the expected response size.

Monitoring the x-ratelimit-remaining-* headers and Azure Monitor metrics.

Rebalancing or increasing TPM allocation when the deployment's allocated quota is insufficient.

For requests that do not require an immediate response, using a queue and processing requests at a controlled rate.

Where a single deployment cannot provide the required throughput, considering multiple deployments/regions where supported.

5. To determine the exact cause in your environment

Since the question does not include the actual error returned during the training session, could you share:

The complete HTTP status code and error message/body from a failed request

The retry-after-ms and x-ratelimit-* response headers, if available

The model and model version

The deployment type, such as Standard, Global Standard, or Data Zone Standard

The Azure region

The approximate number of simultaneous users/requests

The TPM allocated to the affected deployment

With these details, we can determine whether the failures are due to RPM/TPM throttling, burst/concurrency behavior, or service-side capacity constraints and identify the appropriate remediation.

References:
Manage Azure OpenAI in Microsoft Foundry Models quota
Microsoft Foundry Models quotas and limits
Sai Kiran Mudavath 90 Reputation points Microsoft External Staff Moderator

2026-09-30T09:05:29.65+00:00

Hi @Mohamed, Rihan

Did you get any chance to review the above response. Do let me know if you have any further queries.

Thank you!
Sai Kiran Mudavath 90 Reputation points Microsoft External Staff Moderator

2026-10-01T10:03:43.2666667+00:00

Hi @Mohamed, Rihan

We haven’t heard from you on the last response and was just checking back to see if you have a resolution yet. In case if you have any resolution please do share that same with the community as it can be helpful to others. Otherwise, will respond with more details and we will try to help.

Thank You!

1 answer

Your answer

Sai Kiran Mudavath 90 Reputation points Microsoft External Staff Moderator

2026-09-30T09:05:29.65+00:00

Hi @Mohamed, Rihan

Did you get any chance to review the above response. Do let me know if you have any further queries.

Thank you!
Sai Kiran Mudavath 90 Reputation points Microsoft External Staff Moderator

2026-10-01T10:03:43.2666667+00:00

Hi @Mohamed, Rihan

We haven’t heard from you on the last response and was just checking back to see if you have a resolution yet. In case if you have any resolution please do share that same with the community as it can be helpful to others. Otherwise, will respond with more details and we will try to help.

Thank You!

Answer 1

To check the applicable limits and diagnose the errors, use this model:

Check the deployment rate limits from responses and quota configuration
- Azure OpenAI includes rate-limit headers on every API response:
  - x-ratelimit-limit-requests
  - x-ratelimit-limit-tokens
  - x-ratelimit-remaining-requests
  - x-ratelimit-remaining-tokens
  - x-ratelimit-reset-requests
  - x-ratelimit-reset-tokens
  - retry-after-ms on 429 responses
- These headers show the effective request and token limits for the deployment and how much budget remains.
- TPM can also be adjusted through quota management in the Foundry portal for high-traffic deployments, and rebalanced across deployments based on observed usage.
Understand how concurrent users affect RPM and TPM
- Limits apply at the deployment level, not per user.
- Every incoming request contributes to the running token count for the current minute.
- The TPM calculation is based on an estimated maximum processed-token count at request time, including:
  1. prompt text
  2. max_tokens
  3. best_of
- This estimate is not the same as billed tokens, so 429s can occur even when Azure Monitor token usage appears below quota.
- RPM is enforced over short windows, typically 1 or 10 seconds. A burst of simultaneous users can trigger 429s even if the total per-minute request count stays below the published RPM.
Distinguish rate-limit errors from capacity constraints
- Not all 429s mean the same thing.
- Rate limit exceeded: messages such as "Requests to … have been limited" or "Rate limit is exceeded" indicate the deployment exceeded TPM or RPM.
- System capacity throttling: messages such as "The service is temporarily unable to process your request" or "System is experiencing high demand" indicate backend capacity pressure.
- Temporary rate limit adjustment: if configured quota has not changed but x-ratelimit-limit-tokens is lower than the configured TPM, the service has temporarily reduced the effective rate limit to protect shared capacity. This is typically temporary.
- Also check whether token usage metrics are misleading: billed token metrics only show successfully processed requests, while rate limiting is evaluated when requests arrive and can include rejected requests.
Recommended handling for high concurrent usage
- Implement retry logic with exponential backoff.
- Honor retry-after-ms when present.
- Queue requests and process them at a controlled rate if immediate responses are not required.
- Ramp traffic up gradually instead of sending sharp bursts.
- Rebalance quota across deployments and increase TPM on heavily used deployments.
- Spread requests across multiple deployments or regions when one deployment cannot support the required throughput.
- For workloads that require more predictable throughput and lower latency, consider Provisioned Throughput (PTU).
Best practices to reduce rate-limit errors
- Set max_tokens to the minimum value needed.
- Set best_of to 1 unless multiple completions are required.
- Reduce prompt size where possible.
- Monitor x-ratelimit-remaining-requests and x-ratelimit-remaining-tokens in the application and proactively throttle before 429s occur.
- Use Azure Monitor metrics to review 24-hour and 7-day usage trends and identify bursty patterns.
- Test different load-increase patterns before training sessions or peak events.
Concurrent session note for GPT-Live
- If the assistant uses the GPT-Live API, there is also a concurrent session limit per subscription by tier. For example, the default non-tiered limit is 10 concurrent sessions.

A practical troubleshooting path for the training-session scenario is:

Capture the HTTP response headers for successful and failed requests.
Compare x-ratelimit-limit-* and x-ratelimit-remaining-* values during the failures.
Check the 429 error text:
- rate-limit wording -> quota/RPM/TPM issue
- high-demand or temporarily unable to process wording -> capacity issue
Review usage trends in Azure Monitor to see whether the session created short bursts.
If bursts are the cause, add queuing and exponential backoff, reduce max_tokens, and distribute traffic more evenly.
If the workload regularly exceeds a single deployment’s capacity, rebalance quota, use multiple deployments or regions, or move to PTU for more predictable capacity.