3.2.5. AI Cost Planning and Optimization
💡 First Principle: AI cost follows the pricing model, and the pricing model follows the layer of the stack. Consumption pricing scales with every request, instance pricing scales with capacity you reserve whether used or not, and seat pricing scales with the number of people. Matching the pricing model to the usage pattern, estimating before committing, and tracking after launch are the three controls that keep AI spend predictable.
The exam guide names three AWS AI pricing structures:
| Pricing structure | How you pay | AWS examples | Fits when… | Watch out for… |
|---|---|---|---|---|
| Consumption-based | Per unit used: input and output tokens, requests, images, pages | Amazon Bedrock on-demand inference; batch inference at a lower rate | Usage is variable, uncertain, or growing; pilots and new products | Costs rising with long prompts, verbose outputs, and adoption; no natural ceiling |
| Instance-based | Per hour (or other time unit) of compute capacity provisioned | Amazon SageMaker AI training and hosting instances; Amazon Bedrock Provisioned Throughput (hourly, optional 1- or 6-month commitment) | Usage is steady and high, needs guaranteed capacity or performance, or uses custom models | Paying for idle capacity; over-provisioning |
| Seat-based | Per user per month | Amazon Quick subscriptions; many SaaS AI products | Broad employee use with predictable headcount | Paying for inactive seats; low adoption inflating cost per active user |
Planning and tracking tools:
- AWS Pricing Calculator: estimates the cost of AWS services before you deploy, based on assumptions about volume (requests per day, tokens per request, instance hours). Use it to build the cost side of a business case and to compare scenarios (on-demand vs. provisioned, model A vs. model B).
- AWS Cost Explorer: analyzes actual spend after deployment, with trends, filters, and forecasts. Use it to compare reality against the business case, attribute cost to initiatives (with cost allocation tags), and spot anomalies.
- AWS Marketplace: procures third-party AI products and models with consolidated billing on the AWS invoice, which simplifies cost tracking for "buy" decisions.
- Budgets and alerts (for example, AWS Budgets) notify owners when spending approaches a threshold, which turns cost planning into cost control.
Optimization strategies a strategist should expect teams to apply:
- Right-size the model: use smaller, cheaper models for simple tasks (classification, extraction, short answers) and reserve larger models for complex reasoning. Routing requests to the cheapest adequate model can reduce costs substantially.
- Control tokens: concise prompts, retrieval of only relevant passages (2.3.2), output length limits, and prompt caching for repeated instructions.
- Batch non-urgent work: batch inference on Bedrock costs less than on-demand for jobs that do not need real-time answers.
- Match capacity to load: use Provisioned Throughput only when usage is steady enough to keep it busy; shut down idle endpoints.
- Commit for steady use: Savings Plans trade a commitment to a consistent amount of usage, measured in dollars per hour for a one- or three-year term, for lower rates. Machine Learning Savings Plans for Amazon SageMaker AI cover its ML instances (notebooks, processing, training, real-time inference, and batch transform) and can reduce those costs by up to 64%. They do not apply to Amazon Bedrock token charges; on Bedrock, the commitment option is Provisioned Throughput. Commit only after usage has stabilized, because the hourly commitment is billed whether or not it is used.
- Manage seats: reclaim inactive licenses; align seat tiers with real usage.
- Attribute and review: tag resources by initiative and review spend against value monthly.
⚠️ Exam Trap: Use the right tool for the right moment: AWS Pricing Calculator to estimate costs before deployment, AWS Cost Explorer to analyze actual costs after deployment. Also match pricing models to usage: variable or uncertain usage favors consumption pricing, steady high-volume usage may justify provisioned or instance-based capacity (and, on SageMaker AI, a Savings Plan), and broad employee tools are typically seat-based.
Reflection Question: A document-processing workload runs 2 million pages overnight every night, with no real-time requirement, and volume is stable. Which pricing and optimization options would you ask the team to evaluate, and which AWS tool would you use to compare them before committing?
