Cost Optimization and Governance for AI Workloads

Cost is where the AIF-C01 exam stops being abstract. Every other domain can be answered with a definition; this one asks you to compute, and then to pick the control that prevents the overspend rather than the one that merely reports it. The good news is that the arithmetic is shallow, multiplication and percentages, and the traps are few, sharp and repeatedly tested.

By the end of this lesson you will be able to:

• read the three meter families AWS uses for AI (instance-hours, tokens, request units) and say which one a question is billing;
• compute the monthly cost of an endpoint, a token workload or a document-processing pipeline with dated prices;
• choose among on-demand, Spot, Savings Plans and Serverless Inference using the published ceilings (≤90 %, ≤72 %, ≤66 %, ≤64 %);
• apply the Bedrock-specific levers: Batch (50 %), prompt caching (0.1× reads), Provisioned Throughput, service tiers and model routing;
• convert free-tier allowances into dollars so you can size a proof of concept;
• distinguish preventive controls (SCP, RCP, opt-out, Bedrock policy, tag policy, Service Catalog, Budget actions) from detective ones (Cost Explorer, AWS Config, cost allocation tags);
• avoid every dated-price and commitment trap this domain is known for.

---

1. The three meter families: what you are actually billed

1.1 Meter family 1, SageMaker AI: instance-hours plus storage

SageMaker AI bills per instance-hour (per-second in most regions) for training, processing, notebooks and hosted endpoints, plus EBS gigabyte-months. On-demand carries no minimum and no upfront commitment (aws.amazon.com/sagemaker/ai/pricing, verified 06 Oct 2026).

1.2 Meter family 2, Amazon Bedrock: input tokens and output tokens

Bedrock on-demand charges per input token and per output token separately (aws.amazon.com/bedrock/pricing, verified 06 Oct 2026). Because output is generated one token at a time, output rates run roughly 4 to 5× input rates on open-weight models, and far more on premium models.

Non-token Bedrock levers (all verified 06 Oct 2026 unless noted): Batch = 50 % of the on-demand token price for supported foundation models · cache read = 0.1×, cache write = 1.25× (5-minute TTL) or 2× (1-hour TTL) · Bedrock Guardrails on the AWS European pricing pages at €0.2587283 per 1,000 text units (content filters, denied topics) and €0.1724855 per 1,000 (sensitive information, contextual grounding), with regex and word filters free (aws.eu list, 07 Aug 2026: the USD equivalent was not retrieved, so treat it as (G)).

1.3 Meter family 3, prebuilt AI services: one unit per request

Rekognition, Textract, Comprehend, Transcribe, Polly and Lex bill per request unit, with volume tiering after the first 1 to 10 million units (aws.amazon.com/*/pricing, verified 06 Oct 2026).

1.4 Worked example 1: two idle endpoints, 1,488 hours

Two ml.c5.xlarge real-time endpoints run 24×7 for a 31-day month at $0.204/hour (06 Oct 2026):

$$2 \times 24 \times 31 = 1{,}488 \text{ hours}, \qquad 1{,}488 \times \$0.204 = \$303.55/\text{month}$$

$303.55/month regardless of traffic. Three ways to change it:

The lesson is not the percentages; it is that an idle endpoint bills the same as a busy one. Nothing on the meter knows whether anyone sent a request.

• 📚 Did you know? The AWS pricing page for SageMaker bills per second in most regions with a one-minute minimum, so a 90-second inference spike costs one minute, not an hour, but a notebook or endpoint is billed for the whole time it exists, from creation to deletion. That asymmetry is why the exam's cheapest-option questions almost always have "stop it on a schedule" or "use serverless" as the answer.

---

2. The cost-lever decision framework

2.1 The lever impact matrix

Not all levers are equal, and some are frequently oversold. AWS publishes the percentages below; where AWS publishes nothing, this lesson says so.

2.2 The cost-lever decision flow

This is the diagram the exam is really asking you to walk. Start from what is burning money, then from what the workload tolerates.

2.3 What each lever cannot do

Three "cannot" facts are exam gold:

1. Savings Plans do not stack on Spot, and Spot spend does not consume a Compute Savings Plan commitment (docs.aws.amazon.com, 2026). A question offering to "cover Spot usage with a Savings Plan" is wrong on both halves of the sentence.
2. The prebuilt AI services have no published reservation or Savings Plan option; you optimize them by reducing unit volume (choose OCR before Forms, plain transcription before redaction) and by using the free tier, not by committing.
3. A budget action is not a real-time control. Alerts can be forecast-based or actual-based, but execution lags, so overspend is still possible between the threshold and the action.

---

3. SageMaker AI: on-demand vs Spot vs Savings Plans vs serverless

3.1 On-demand: the default with no commitment

On-demand is per instance-hour with no minimum and no upfront (aws.amazon.com/sagemaker/ai/pricing, verified 06 Oct 2026). It is the right answer when the exam says prototype, short-lived, unpredictable, or first two months (the SageMaker free tier gives 250 hours/month of ml.t3.medium notebooks for the first two months, verified 06 Oct 2026).

3.2 Spot: up to 90 % off, two minutes' notice

EC2 Spot, and therefore SageMaker Spot and SageMaker HyperPod Spot, is advertised at up to 90 % off on-demand, with two minutes' interruption notice and a historical average interruption frequency below 5 % (aws.amazon.com/ec2/spot, 2025 to 26). SageMaker HyperPod Spot is typically up to 90 % off HyperPod on-demand, priced by EC2 with no upfront (aws.amazon.com/sagemaker/ai/pricing, 06 Oct 2026).

Spot is only correct when the job can be interrupted: training with checkpoints, batch transform, offline processing. The exam's answer pattern is always the same pair: Spot plus checkpoint and retry.

3.3 Savings Plans: three different discounts

SageMaker Savings Plans are the flexible one: because they span every ml.* instance family in every region, you can move from ml.c5.xlarge to ml.g5.24xlarge and keep the discount. Cost Explorer surfaces the recommendation; the commitment should cover only your baseline, 70 to 80 % of your 30-day average hourly spend, not your peak.

3.4 Serverless Inference: the meter that reaches zero

SageMaker Serverless Inference scales to zero when idle, so you pay only for the compute you use, billed by the millisecond, plus data processed. Optional Provisioned Concurrency is billed on memory, duration and concurrency (docs.aws.amazon.com/sagemaker/latest/dg/serverless-endpoints.html, 2026). It is the correct answer whenever traffic is unpredictable and low: a prototype that receives a few requests a day, or none for a week.

3.5 Worked examples: SageMaker arithmetic

Worked example 2: an always-on GPU fleet (AWS worked example). Four ml.g5.24xlarge instances at $10.18/hour (06 Oct 2026) run 720 hours in a month, with 100 GB of EBS each at $0.14/GB-month (06 Oct 2026):

$$4 \times 720 \times \$10.18 = \$29{,}318.40, \qquad 4 \times 100 \times \$0.14 = \$56.00$$

$29,318.40 + $56.00 = $29,374.40 per month. At 30 % utilization the fleet can drop to roughly 1.2 instances ≈ $8,812/month, saving about $20,560/month, and then a SageMaker Savings Plan (≤64 %, 06 Oct 2026) applies on top of the smaller footprint. Right-sizing first, commitment second; reversed, you commit to a bloated baseline for three years.

Worked example 3: the idle custom endpoint burn. A custom NLP endpoint at $0.0005 per inference-unit-second (third-party rate, cloudzero.com, 04 May 2026: (G)) burns:

$$\$0.0005 \times 60 = \$0.03/\text{min} \;\Rightarrow\; \$1.80/\text{hour} \;\Rightarrow\; \$43.20/\text{day} \;\Rightarrow\; \$1{,}296/30\text{ days}$$

Billed from endpoint creation to deletion even at zero traffic. Fix: schedule it on and off with EventBridge, or use async/serverless so the meter stops.

Worked example 4: the 64 % Savings Plan ceiling. An ml.m5.xlarge-class instance at $0.23/hour (rate used for this example, 06 Oct 2026) for 730 hours:

• 📚 Did you know? AWS's SageMaker Savings Plan discount of up to 64 % (06 Oct 2026) is lower than the EC2 Instance Savings Plan ceiling of up to 72 % (06 Oct 2026), yet the SageMaker plan is usually the better buy for AI teams, because the EC2 plan is locked to one instance family in one region, while the SageMaker plan follows you across every ml.* family, size, region and component. A 8-point discount difference is often worth less than the freedom to switch from CPU to GPU without re-buying a commitment.

3.6 Pricing-model decision for an endpoint

3.7 Comparative verdict

---

4. Amazon Bedrock: tokens, Batch, caching, throughput and tiers

4.1 Worked example 5: the token bill, and why model choice beats volume

A workload sends 300M input tokens and produces 60M output tokens per month.

The same token volume costs 15.5× more on GLM 5. Model selection, not usage reduction, is the dominant lever for generative workloads, and it is why every "reduce cost" question that offers "switch to a smaller model" is usually right provided quality evaluations are mentioned.

Note also that output is ~4× input per token on Nemotron Nano 2 ($0.23 vs$0.06, 06 Oct 2026): capping max_tokens, asking for terse answers and returning structured output instead of prose all cut the expensive half of the meter.

4.2 Worked example 6, Batch: 50 % off for async work

A nightly document-classification job costs $800/month on Bedrock on-demand; the model is batch-enabled and nobody reads the results before morning.

$$\$800 \times 50\% = \$400/\text{month} \quad\Rightarrow\quad \$400/\text{month saved} = \$4{,}800/\text{year}$$

Bedrock Batch = 50 % of the on-demand token price for supported models (aws.amazon.com/bedrock/pricing, 06 Oct 2026). The cost of the discount is async latency, and the model must support batch at all.

4.3 Worked example 7, prompt caching: ~90 % off, with a write premium

This is the highest-yield arithmetic in the domain. Using the rates published in an AWS Builder Center post for Sonnet 4.6 on Bedrock, $3.00 per 1M input tokens, $3.75 cache write, $0.30 cache read (2026; not confirmed on the official pricing page, see section 10), a stable 33,000-token system prompt sent 1,000 times/day for 30 days:

Uncached:

$$33{,}000 \times 1{,}000 = 33\text{ MTok/day} \times \$3.00 = \$99/\text{day} \;\Rightarrow\; \$2{,}970/\text{month}$$

Cached (5-minute TTL: write 1.25×, read 0.1×):

Saving ≈ $2,668/month (≈ 89.9 %). The mechanism (all Bedrock caching docs, 2026): cache reads cost 0.1× the input price (90 % off), cache writes cost 1.25× at the default 5-minute TTL and 2× at the opt-in 1-hour TTL, the default TTL is 5 minutes, the 1-hour TTL is opt-in per cache checkpoint, minimum 512 to 4,096 tokens per model and maximum 4 checkpoints.

4.4 Provisioned Throughput and service tiers

Concrete Provisioned Throughput numbers for Llama 2 13B (06 Oct 2026): $21.18/hour/MU on a 1-month commitment versus $13.08/hour/MU on a 6-month commitment, a 38.3 % reduction, while inference without a commitment is $23.50/hour/MU. The trap: you pay for the MU whether you use it or not, so Provisioned Throughput is only right above roughly 60 % utilization (a third-party break-even figure, (G), not AWS-published).

4.5 Quota burndown is not token count

Bedrock quota consumption is weighted, and the weights are model-specific (docs.aws.amazon.com, 2026):

So "why am I throttled when my token count looks fine?" is answered by long outputs on a 15× model, not by input volume.

4.6 The Bedrock cost-lever map

• 📚 Did you know? Bedrock has no AWS-published free-tier table (06 Oct 2026): a claim you will find on vendor blogs saying "no free tier" or quoting a "30 to 50 % Provisioned Throughput discount" is (G), because AWS publishes neither. What AWS does publish is a 50 % Batch discount, a 90 % cache-read discount and the 64 % SageMaker SP ceiling; those are the numbers you may quote on the exam.

---

5. Prebuilt AI services: units, tiers and the free tier

5.1 How to optimize a per-unit meter

There is no reservation for Rekognition, Textract, Comprehend, Transcribe, Polly or Lex. The three levers are:

1. Cheapest feature set that returns what you read: Textract DetectDocumentText at $0.0015/page (06 Oct 2026) versus AnalyzeDocument Forms at $0.05/page (06 Oct 2026) is a 33× premium for structure. Read OCR first; invoke Forms only when you need the key-value pairs.
2. Volume tiering: Comprehend drops from $0.0001 to $0.00005 (10 to 50M units) to $0.000025 (above 50M) per 100-char unit (06 Oct 2026); Textract DetectDocumentText drops from $0.0015 to $0.0006 after the first million pages in a month (06 Oct 2026).
3. Free tier: 3 to 12 months of allowances, converted to dollars in section 5.4.

5.2 Worked example 8: Textract tier math

Same API, three bills. The third row is the one that surprises teams: feature choice outweighs volume.

5.3 Worked example 9, Transcribe: batch, not streaming, and redaction as an add-on

Two million minutes of audio per month (rates verified 06 Oct 2026):

And if the same job adds content redaction on 2M minutes (tiered: 250K at $0.0024, 750K at$0.0015, 1M at $0.00102):

$$(250{,}000 \times 0.0024) + (750{,}000 \times 0.0015) + (1{,}000{,}000 \times 0.00102) = 600 + 1{,}125 + 1{,}020 = +\$2{,}745/\text{month}$$

Redaction is an opt-in meter, billed on top of the transcription minute: the exam's "cheapest accurate pipeline" questions hinge on noticing that.

5.4 The AI free tier, converted to dollars

Worked example 10: the free-tier ceiling for a pilot. Polly Standard 5M characters = $20.00 plus Neural 1M characters = $16.00 (06 Oct 2026 rates) → $36/month for 12 months of voice output; Comprehend 50K units = $5.00 per API per month (06 Oct 2026); Textract 1,000 pages = $1.50 (06 Oct 2026); Transcribe 60 minutes = $0.60 streaming (06 Oct 2026). A six-month voice + entity-extraction pilot can therefore run on $0 of compute: provided you stay inside the allowance, because the free tier does not warn you before it stops.

• 📚 Did you know? The AWS Free plan introduced on 15 Jul 2025 gives up to $200 in credits ($100 immediately plus up to $100 more) over 6 months (aws.amazon.com/free, 06 Oct 2026); which is separate from the service free tiers that run 3, 12 or 24 months. Exam options frequently conflate "free credits" with "free tier"; the credits expire on a calendar, the free tiers expire on usage.

---

6. Cost visibility: Cost Explorer, cost allocation tags and budgets

Visibility is detective, not preventive, but without it you cannot set a threshold, and without a threshold a budget action never fires.

6.1 Cost Explorer

6.2 Cost allocation tags must be activated

Cost allocation tags come in two flavours and are activated separately in the Billing console before Cost Explorer can group by them (docs.aws.amazon.com, 2026):

The classic symptom: an engineer applies a CostCenter tag to every resource yesterday and it still does not appear in Cost Explorer. The cause is almost never "tags do not work with AI services" or "the key must start with aws:"; it is that the user-defined cost allocation tag has not been activated (and can take up to 24 hours to appear).

6.3 AWS Budgets: free alerts, then $0.10/day per extra action

Worked example 11: the budget gate. A team sets a $5,000/month budget with alerts at 80 % and 100 % plus automatic actions (attach an SCP, stop non-production instances):

• monitoring and alerts: $0;
• actions on the first two action-enabled budgets: $0;
• a third action-enabled budget: $0.10/day ≈ $3.00/month (06 Oct 2026).

The exam's "minimum feature cost" answer is therefore: alerts free, two action-enabled budgets free, then $0.10/day each, and remember that a budget action is applied after the threshold, so it is a brake, not a wall.

• 📚 Did you know? AWS Budgets supports budget actions that attach an SCP or IAM policy, or stop resources, which makes it the one cost tool that does something preventive, but only after the threshold is crossed and the action executes. That is why the exam pairs it with SCPs: the budget detects, the SCP enforces.

---

7. Governance: preventing the spend instead of reporting it

7.1 The preventive stack in AWS Organizations

AWS Config rule growth: AWS added 75 new rules in Nov 2025 and 191 more in Jul 2026 (AWS news, 2025 to 26), and Conformance Packs deploy rule groups across an entire organization: the answer whenever a question asks how to detect drift at scale.

7.2 Worked example 12: the right control for each requirement

7.3 Service Catalog vs SageMaker Catalog: do not confuse them

7.4 Governance layer diagram

---

8. Savings Plans, Spot and TCO: the numbers you may quote

8.1 Published savings percentages (all official, verified 06 Oct 2026)

8.2 SageMaker TCO: the whole-organisation number

AWS's published TCO comparison (AWS ML blog and TCO executive summary, 2020, still cited 2026) claims SageMaker's 3-year TCO is ≥ 54 % lower than self-managed EC2 or EKS, and breaks down by team size: 5 data scientists −90 %, 15 data scientists −87 % / −85 %, 50 data scientists −79 % / −65 %, 250 data scientists −77 % / −54 % (EC2 / EKS respectively). The point of the table is that labour, security and operational overhead dominate for small teams, while infrastructure dominates as the team grows.

8.3 Commitment safety rules

1. Commit the baseline, not the peak: 70 to 80 % of the 30-day average hourly spend.
2. Never commit to Spot: Savings Plans do not stack on Spot, and Spot spend does not consume a Compute Savings Plan commitment (docs.aws.amazon.com, 2026).
3. Prefer the flexible plan when your estate mixes CPU and GPU: the SageMaker AI Savings Plan spans all ml.* families, sizes, regions and components (06 Oct 2026).
4. Re-evaluate before renewal: a 1-year Compute Savings Plan on a workload that moves to serverless in month 6 is money burned.
5. Provisioned Throughput is a volume bet: you pay per Model Unit per hour whether used or not: go above roughly 60 % utilization (third-party figure, (G)) before committing, and remember custom models require Provisioned Throughput at all.

---

9. Governance and cost in one operating rhythm

Putting it together into the loop the exam expects you to describe:

1. Tag first: apply Project, Environment, CostCenter; activate user-defined and AWS-generated cost allocation tags separately (≤24 h).
2. Measure: Cost Explorer (console free, API $0.01/request, hourly ≈ $0.01/1,000 records/month, 06 Oct 2026).
3. Threshold: AWS Budgets alerts (free) at 80 %/100 %, forecast or actual.
4. Act: Budget actions (first two action-enabled budgets free, then $0.10/day, 06 Oct 2026): attach an SCP, or stop non-production resources.
5. Prevent: SCPs/RCPs for Region and API allow-lists, Service Catalog template constraints + TagOptions for launch-time enforcement, Bedrock policies for pinned guardrails, tag policies for key standardization.
6. Detect: AWS Config rules and Conformance Packs (sagemaker-feature-group-tagged, encryption, logging) with SSM remediation.
7. Optimize continuously: right-size, stop idle, add Spot for fault-tolerant work, add a Savings Plan only for the measured baseline, and apply the Bedrock-specific levers (Batch, caching, routing, tiers).

• 📚 Did you know? sagemaker-feature-group-tagged is a real AWS Config managed rule, and tag keys in Config are case-sensitive and must not use the aws: prefix. A rule written for Project will not fire on project, and an attempt to register a user tag beginning with aws: is rejected. That single detail explains a whole genre of "the rule shows compliant but spend is unattributed" incidents.

---

10. What is verified, what is third-party, and what conflicts

The exam rewards dated, sourced numbers. This is exactly what this lesson will not assert without a date:

---

11. 2025 to 2026 Updates

The prices and the tooling behind this domain did not stand still between mid-2025 and late 2026. Every change below carries a date, because an undated number is unusable on this exam, and because several rows retire tools that older study guides still recommend paying for.

Table A: pricing, cost-tooling and availability changes, each with its date (researched 06 Oct 2026):

Table B: the Bedrock cost-lever rates you may quote for 2025 to 2026 (aws.amazon.com/bedrock/pricing and AWS blogs, verified 06 Oct 2026):

• 📚 Did you know? The AIF-C01 exam guide v1.1, published 30 Apr 2026, added seven objectives, among them 2.1.4 (token-based pricing and its effect on cost and performance), 2.1.5 (context engineering) and 2.1.6 (agentic AI: MCP, memory management, tool usage, orchestration), while Amazon MemoryDB was removed from scope (docs.aws.amazon.com/aws-certification/latest/ai-practitioner-01/, 30 Apr 2026). So the cost arithmetic you have been practising in this lesson is not a legacy topic: AWS wrote it into the exam guide itself, roughly one month before those objectives started appearing on live questions.

---

Real-World Case Studies

A pricing page is an abstraction; these are the same levers with a customer's name attached. All figures below are customer- or AWS-claimed and unaudited, only Sun Finance (n = 585 images) and Adobe (own test set) disclose a sample basis, and "up to" is always a ceiling, never an average.

Forethought, the DIY tax. A three-person team was running models and Kubernetes; moving off its own Amazon EKS to multi-model endpoints consolidated GPU capacity (−66 % with better latency), and Serverless Inference took the idle-endpoint bill to roughly −80 % for the small classifiers. Those are exactly the two levers from sections 3.4 and 2.1, consolidation, then scale-to-zero. Note what the case does not publish: no Spot figure, no Savings Plan figure. Do not import this lesson's 90 %/64 % ceilings into a case that never claimed them.

Sun Finance: pay each meter for the job it is best at. Textract bills per page ($0.0015 to$0.070 per page, 06 Oct 2026) while Claude bills per token; splitting OCR from structuring cut cost per document by 91 % and lifted accuracy from 79.73 % to 90.80 % across a 585-image evaluation set. The documented first attempt: Claude Sonnet 4 alone: scored 61.8 % overall and 43 % on ID numbers and was rejected, because AWS's privacy protections block direct PII extraction. LLM-only extraction is not the cheap option; it is a failed prototype with a token bill attached.

HAYAT HOLDING, managed beats self-built. "Time-consuming and cumbersome" self-built ML environments gave way to Automatic Model Tuning over 194 OPC-UA sensor streams with SageMaker Edge Manager running models on-device, for a published $300,000 per year saving (2023). The lever here is not a discount percentage at all, it is removing the labour the TCO table in section 8.2 says dominates for small teams.

RareJob, right-size and managed-service first, discount second. The team went local PC → EC2/ECS (still too expensive) → SageMaker managed spot training fed by Glue and Athena: −25 % training time, >10× developer efficiency and 100 hours per month back (2020). Spot's published ceiling is up to 90 % off with two minutes' notice (06 Oct 2026), but the case's own numbers are throughput and labour, not price per hour, quote them that way.

Table: the AWS-published Bedrock lever figures behind these outcomes (AWS News Blog, 2024; AWS Cloud Financial Management, 2025; all re-verified 06 Oct 2026):

• 📚 Did you know? AWS publishes exactly one production-rate across its entire generative-AI case corpus: 65 % of Generative AI Innovation Center projects reached production in 2025, some in as little as 45 days, out of more than 1,000 implementations, using the Five V's framework (Value → Visualize → Validate → Verify → Venture) (aws.amazon.com/blogs/machine-learning/beyond-pilots-a-proven-framework-for-scaling-ai-to-production, 2025). The implicit other 35 % is the part customer-story pages leave out, and AWS itself never claims a higher figure.

---

Practice Questions

---