Exams / SAP-C03 / Study guide / AI workload costs (Bedrock tiers, batch inference, caching, cross-Region inference)
AI workload costs (Bedrock tiers, batch inference, caching, cross-Region inference): SAP-C03 study notes and practice questions
What to know about AI workload costs (Bedrock tiers, batch inference, caching, cross-Region inference) for the SAP-C03 exam, from the Cost-Optimized Architecture Design domain (15% of the exam guide), then 2 free questions to test it. Part of the SAP-C03 study guide.
AI cost questions describe a growing Amazon Bedrock or SageMaker AI bill and a workload profile: an interactive assistant, a nightly document job, a chatbot that resends the same long context, or a team in a regulated Region. You pick the levers that fit.
Amazon Bedrock levers
Lever
What it does
Fits
Flex tier
Discounted on-demand tier for work that tolerates longer processing times
Evaluations, summarization, agent steps that are not user-facing
Standard tier
Default on-demand pricing
Everyday interactive traffic
Priority tier
Premium price, served ahead of Standard and Flex
Latency-critical customer-facing calls
Reserved tier
Fixed monthly price for reserved tokens-per-minute (1 or 3 month terms); overflow goes to Standard
Submit prompts as files in S3, results written back to S3; about 50% lower price than on-demand for supported models
Large offline jobs with no real-time need
Prompt caching
Caches a repeated prompt prefix; cache reads cost much less than normal input tokens and cut latency
Long system prompts, documents or tool definitions reused across calls
Model choice
Smaller or distilled models cost less per token
Simple classification or extraction steps
Intelligent prompt routing
Routes each request between models in a family by predicted quality and cost
Mixed easy and hard prompts
The tier is chosen per request with the service_tier parameter, and the tier that served a request is visible in CloudWatch metrics and CloudTrail. The on-demand quota is shared by Priority, Standard and Flex. Prompt caching is not available for batch inference.
Cross-Region inference and transfer
Geographic cross-Region inference profiles (for example US or EU) spread requests across Regions inside one geography at standard pricing; use them when data must stay in that geography.
Global cross-Region inference profiles route to any supported commercial Region for more throughput and about 10% lower token prices than geographic profiles.
There is no extra routing charge; you pay the price of the source Region you call from. Logs stay in the source Region.
An SCP that denies Bedrock calls with aws:RequestedRegion of unspecified blocks global routing where residency rules forbid it. See data residency.
Keep knowledge base data, vector stores and the application in the same Region as inference to avoid inter-Region transfer and latency.
SageMaker AI endpoints
Traffic pattern
Endpoint option
Steady, latency-sensitive
Real-time endpoint with auto scaling, covered by SageMaker AI Savings Plans
Intermittent, can accept cold starts
Serverless inference
Large payloads or long processing, queued
Asynchronous inference (can scale to zero)
Offline scoring of a dataset
Batch transform
Many small models
Multi-model endpoints or inference components sharing instances
Graviton or AWS Inferentia-based instances can lower inference cost where the model supports them.
How to choose
Scenario wording
Pick
“Nightly job summarizes millions of documents”
Bedrock batch inference
“Same 20-page policy sent with every question”
Prompt caching
“Agent background steps, latency not critical”
Flex tier
“Customer chat must stay fast at peak”
Priority tier, or Reserved tier for guaranteed capacity
“Cut token price, no residency limits”
Global cross-Region inference
“EU data must be processed in the EU”
EU geographic inference profile
Common traps
Provisioned Throughput or Reserved capacity for spiky, low-volume traffic.
Real-time calls for work that could run as batch.
Global inference for data that must stay in one geography.
Using the largest model for every step of an agent.
A news publisher uses an Amazon Bedrock model to write summaries for about two million articles each night. The summaries are read the next morning, so no request needs an answer in seconds.
The article texts for each night are written to Amazon S3 by 11 PM, and every summary must be ready by 6 AM. The publisher's finance team wants the lowest cost per summary, and it does not want a fixed-capacity commitment.
Which solution meets these requirements?
Choose one answer.
AA. Call the model on demand for each article during the night with the default settings, so that each request is answered as fast as possible at the standard on-demand rate.
BB. Route each request through a cross-Region inference profile, so that requests are spread across Regions and each one finishes sooner.
CC. Buy a one-month provisioned throughput commitment for the model and send the nightly summaries through it, so every request gets dedicated capacity.
DD. Write the night's requests as JSONL files in S3, run them as an Amazon Bedrock batch inference job, and load the summaries from the job's S3 output before 6 AM.
Show the answer and explanation
Answer: D
A. Wrong: On-demand calls work, but they are billed at the standard rate even though no request needs a fast answer. Batch inference is priced lower for this kind of offline job.
B. Wrong: Cross-Region inference helps with throughput and availability, but it does not lower the per-token cost of the summaries. The publisher's cost goal is not met by this change.
C. Wrong: Provisioned throughput charges for reserved capacity whether or not the job uses it. The nightly pattern does not need dedicated capacity, and the finance team asked to avoid a fixed commitment.
D. Right: Correct. Batch inference processes large offline sets of prompts from S3 at a lower price than on-demand calls, and it needs no capacity commitment. The overnight window gives the job time to finish before the summaries are needed.
Explanation
Batch inference suits an offline nightly job whose input is already in S3, and it costs less than on-demand calls without any commitment. On-demand calls cost more, provisioned throughput is a commitment, and cross-Region routing does not reduce the per-token cost.
A retailer runs a product-advice assistant on Amazon Bedrock in us-east-1. Traffic is high and steady across the day and night, and shoppers expect answers within a couple of seconds, so the team needs guaranteed capacity for that steady baseline.
During promotions, traffic bursts well above the baseline, and on-demand requests in us-east-1 are throttled. Burst requests may be processed in any US Region but never outside the US.
The team wants no extra capacity commitment for the bursts and no Region-selection logic of its own.
Which TWO designs meet these requirements at the LOWEST cost? (Choose two.)
Choose two answers.
AA. Purchase provisioned throughput for the model sized to the steady baseline, and send the baseline traffic to the provisioned model.
BB. Send all requests, including the shopper-facing ones, through batch inference jobs that run every hour, and read the results when the job finishes.
CC. Send requests that exceed the provisioned throughput through the US geographic cross-Region inference profile for the model, which routes them among US Regions with available capacity.
DD. Purchase provisioned throughput sized to the highest promotion peak, so that every request, including the bursts, runs on dedicated capacity.
EE. Enable model access in us-west-2 as well, and have the application send each burst request to whichever Region returned fewer throttling errors in the previous minute.
Show the answer and explanation
Answers: A, C
A. Right: Correct. Provisioned throughput gives guaranteed capacity for the predictable, round-the-clock baseline. Sizing it to the baseline rather than the peak avoids paying for capacity that sits idle outside promotions.
B. Wrong: Batch inference is for offline work with no interactive response time. Shoppers would wait for the next job, which breaks the couple-of-seconds expectation.
C. Right: Correct. A geographic cross-Region inference profile spreads on-demand requests across US Regions, so bursts are less likely to be throttled and stay within the US. It needs no commitment, and Bedrock chooses the Region, so the team writes no routing logic.
D. Wrong: Peak-sized provisioned throughput covers the bursts, but the retailer pays for that capacity every hour while promotions are occasional. It is also the extra commitment the team wants to avoid.
E. Wrong: This spreads bursts across two US Regions, but the team would write and maintain the Region-selection logic. A cross-Region inference profile does this routing for them.
Explanation
Provisioned throughput sized to the baseline guarantees capacity for steady traffic, and a US cross-Region inference profile absorbs bursts across US Regions with no commitment or custom routing. Hourly batch jobs break the response time, peak-sized provisioned throughput wastes money, and home-made Region selection adds code.
Want more on AI workload costs (Bedrock tiers, batch inference, caching, cross-Region inference)? The 5 full-length practice sets have 6 more AI workload costs (Bedrock tiers, batch inference, caching, cross-Region inference) questions, 375 in all, none of them in the free exam. One payment of $14.99, no subscription. See the practice sets →
Study AI workload costs (Bedrock tiers, batch inference, caching, cross-Region inference)