Use FIPS Endpoints

Amazon SageMaker AI publishes FIPS 140-3 endpoints alongside its standard ones. Switching to them changes only the hostname your client connects to; the endpoint, the model, and the request payload stay the same. For what these endpoints do and do not cover, see Security and Compliance.

The AWS SDKs give you three ways to select FIPS endpoints, from broadest to narrowest scope:

Apply use_fips_endpoint to each client you build. Both the control plane (sagemaker) and the inference client (sagemaker-runtime) need it:

import boto3
from botocore.config import Config

REGION = "us-west-2"
fips = Config(use_fips_endpoint=True)

sagemaker = boto3.client("sagemaker", region_name=REGION, config=fips)
runtime = boto3.client("sagemaker-runtime", region_name=REGION, config=fips)

print(sagemaker.meta.endpoint_url)
# https://api-fips.sagemaker.us-west-2.amazonaws.com
print(runtime.meta.endpoint_url)
# https://runtime-fips.sagemaker.us-west-2.amazonaws.com

Invoke the endpoint exactly as you would otherwise:

response = runtime.invoke_endpoint(
    EndpointName="<your-endpoint-name>",
    ContentType="application/json",
    Accept="*/*",
    Body=audio_bytes,
    CustomAttributes="model=nova-3&language=en&smart_format=true",
)

Bidirectional streaming reaches the runtime host on port 8443, and the FIPS runtime host serves that port as well. The HTTP/2 bidirectional streaming client takes an explicit endpoint, so point it at the FIPS hostname and keep the port:

import {
  SageMakerRuntimeHTTP2Client,
} from "@aws-sdk/client-sagemaker-runtime-http2";

const region = "us-west-2";
const client = new SageMakerRuntimeHTTP2Client({
  region,
  endpoint: `https://runtime-fips.sagemaker.${region}.amazonaws.com:8443`,
});

The equivalent in Python, using aws_sdk_sagemaker_runtime_http2:

endpoint_uri = f"https://runtime-fips.sagemaker.{region}.amazonaws.com:8443"

For the full streaming request shape — payload parts, control messages, and result handling — see Invoke a Deepgram SageMaker Endpoint.

Asynchronous endpoints read their input and write their output to Amazon S3, so configure the S3 client for FIPS as well. Otherwise the invocation travels over FIPS while the payload does not:

s3 = boto3.client("s3", region_name=REGION, config=fips)
runtime = boto3.client("sagemaker-runtime", region_name=REGION, config=fips)

print(s3.meta.endpoint_url)
# https://s3-fips.us-west-2.amazonaws.com

The AWS CLI honors AWS_USE_FIPS_ENDPOINT and use_fips_endpoint, and also accepts --endpoint-url:

aws sagemaker describe-endpoint \
  --endpoint-name <your-endpoint-name> \
  --region us-west-2 \
  --endpoint-url https://api-fips.sagemaker.us-west-2.amazonaws.com

Print the resolved endpoint URL rather than assuming the setting took effect. meta.endpoint_url reports what the client will actually call, after every configuration source has been applied:

for name, client in {"control plane": sagemaker, "inference": runtime}.items():
    url = client.meta.endpoint_url
    print(f"{name:14} {url}  FIPS={'-fips.' in url}")

This check matters most in the IAM Identity Center case above, where a misconfigured run can reach SageMaker over standard endpoints while your logs claim FIPS.

• Security and Compliance — what FIPS endpoints cover, FedRAMP coverage, network isolation, and VPC options
• Invoke a Deepgram SageMaker Endpoint — full request examples for each transport