Portfolio 02 · production serving

Managed ML endpoint cost calculator

Translate a compatible load-test result into minimum, peak, rollout, and failover replicas across four managed endpoint surfaces.

Decision this answers: What does the endpoint cost before traffic arrives, and can its configured replica ceiling meet the measured peak?

Scenario

Decision summary

Required peak replicas

8exact-derived

Integer capacity after headroom.

[1]

Zero-traffic floor

$3066.00exact-derived

Minimum replicas for the whole month.

[1]

Rollout surcharge

$134.40exact-derived

Temporary duplicate peak capacity.

[2]

Monthly topology cost

$6951.00exact-derived

Idle, active extra, rollout, and failover terms.

[1][2][3][4]

Direct-labeled result profile

Required peak replicas Zero-traffic floor Rollout surcharge Monthly topology cost

The tables below are the accessible source of truth; bar lengths never carry identity alone.

Provider capability and measurement matrix

Provider capability and measurement matrix
Candidate	Mode	Region	Capacity evidence	Price	Scaling boundary	Evidence
Vertex AI	Dedicated endpoint	us-central1	User-measured 32 req/s	Editable rate	Scale-to-zero depends on mode	measured[1]
Azure Machine Learning	Managed online endpoint	eastus	User-measured 32 req/s	Editable rate	Deployment minimum applies	measured[2]
Amazon SageMaker AI	Real-time endpoint	us-east-1	User-measured 32 req/s	Editable rate	Serverless is a distinct mode	measured[3]
Databricks Model Serving	Serverless serving	us-east-1	User-measured 32 req/s	Editable DBU rate	Infrastructure included by documented meter	measured[4]

Official surface versus this site

Vertex AI pricing

Use the official surface to confirm native prices, counting rules, limits, and product eligibility.

This offline workbench

Official pages price individual meters. This workbench turns one user-measured sustainable rate into topology-aware capacity and keeps idle, peak, rollout, and failover costs separate.

Sources and methodology

Every result-affecting reference is visible here without JavaScript and is retained in the JSON export.

[1] Vertex AI pricing

Authority: Google Cloud
Native identifier: Vertex AI online prediction
Unit: replica-hour
Evidence: published list
As of: 2026-08-01
Retrieved: 2026-08-01
Freshness: reference
Normalization: Native endpoint meters and topology are preserved; throughput remains user measured.

Open the authoritative source

[2] Azure ML online endpoints

Authority: Microsoft
Native identifier: managed online endpoint
Unit: deployment replica
Evidence: exact-derived
As of: 2026-08-01
Retrieved: 2026-08-01
Freshness: reference
Normalization: Topology and deployment behavior are used without inferring capacity.

Open the authoritative source

[3] Amazon SageMaker AI pricing

Authority: Amazon Web Services
Native identifier: real-time inference
Unit: instance-hour
Evidence: negotiated
As of: 2026-08-01
Retrieved: 2026-08-01
Freshness: reference
Normalization: Serving modes remain distinct and use an editable effective rate.

Open the authoritative source

[4] Databricks serverless DBU consumption

Authority: Databricks
Native identifier: Model Serving
Unit: DBU
Evidence: negotiated
As of: 2026-08-01
Retrieved: 2026-08-01
Freshness: reference
Normalization: Serverless DBU billing is not given a second infrastructure bill.

Open the authoritative source

Limits and non-claims

The opening plan requires 8 replicas and is viable against the configured maximum: true.
Capacity requires a compatible measured workload point; instance specifications do not substitute for a load test.
Quota, availability, latency guarantees, and negotiated contracts are outside this static plan.