Pricing Misconception — AWS AI Practitioner (AIF-C01)
The answer assumed a pricing model that doesn't apply. Reserved vs. on-demand, data transfer costs, or tier thresholds tripped you up.
No Commitment Required Doesn't Mean Lowest Cost
The scenario describes a Bedrock prototype running at low volume. On-demand token pricing gets selected as the cost-efficient option—no upfront commitment, pay as you go. The trap: on-demand pricing scales linearly with every token. Once throughput crosses a predictable threshold, Provisioned Throughput yields a lower per-token cost in exchange for a capacity commitment. The exam tests whether candidates know that "flexible" and "cheapest" diverge once usage patterns stabilize.
The Scenario
A startup wants to reduce costs for a development environment running 8 hours per day, 5 days per week. You recommend Reserved Instances for the 40% savings. But RIs commit you to paying 24/7 whether the instance runs or not. For a workload running 29% of the time, you are paying for 71% idle hours. On-Demand with EC2 Instance Scheduler (stop/start automation) or even Spot Instances for fault-tolerant dev workloads cost less. The trap is applying the "RIs always save money" rule without calculating the break-even utilization point — which for 1-year No Upfront RIs is roughly 40-50% utilization.
How to Spot It
- •When you see "reduce costs," do not default to Reserved Instances. Calculate the actual utilization: hours running / hours in period. RIs only save money when utilization exceeds the break-even point for that commitment type. Dev environments, batch jobs, and seasonal workloads rarely qualify.
- •NAT Gateway data processing charges are $0.045/GB — a 1TB/month workload costs $45/month just in NAT processing, on top of the $0.045/hour fixed cost. Questions that describe high-throughput private subnets accessing the internet are testing whether you know NAT Gateway is not "free" infrastructure.
- •Data transfer across AZs is $0.01/GB each way. A chatty microservices architecture with 100GB/month cross-AZ traffic adds $2/month per direction — invisible until you multiply by dozens of services. The exam tests whether you spot cross-AZ transfer costs in architectures that replicate data between availability zones.
Decision Rules
When a supervised NLP inference task maps exactly to a purpose-built service API, should the team select that API on total-cost grounds — absorbing zero training or endpoint cost — rather than choosing a custom ML platform based on a lower-sounding per-invocation rate?
When a stated business problem maps directly to a purpose-built AWS AI service domain (fraud detection), choose that managed service over a custom SageMaker model because total cost of ownership—including model development labor, training compute, and persistent endpoint hosting—exceeds the purpose-built service's per-prediction charge.
Whether to apply an in-prompt technique (chain-of-thought prompting) that incurs only additional input-token cost, or a model customization approach (fine-tuning) that appears to be a one-time fixed cost but introduces a separate training-infrastructure billing dimension that violates the no-training-budget constraint.
Whether to apply an in-prompt output-format directive within Amazon Bedrock to constrain FM response length, or to replace the FM invocation path with Amazon Lex to exploit its flat per-request pricing and avoid per-token generation costs.
Whether few-shot prompt examples embedded directly in the Amazon Bedrock invocation can enforce tone and format consistency, versus adopting Amazon Kendra for retrieval-augmented generation—which adds per-query retrieval and index-sync costs that are unjustified when the defect is format adherence rather than factual knowledge gaps.
Whether to suppress hallucinated citations through an in-prompt grounding instruction applied at inference time, or through a retrieval-augmented or model-level service that introduces persistent billing beyond inference tokens.
Domain Coverage
Difficulty Breakdown
Related Patterns