PurchaseRequisition:AWS

lastmod 2026-04-20created 2026-04-204 edits

Sample AWS purchase requisition

Basic information

• Title: [??? task] Submitting an approval request for leasing an AWS GPU instance for model model development.
• Project name: Development of an AI-based high-efficiency autonomous production system (RS-2025-XXXXXX)
• Project budget: Research activity expenses from the 2026 project budget (cloud usage fees)
• Operating policy: The instance stays in the running state only on weekdays (Mon–Fri) 09:00–18:00; at other times and on weekends it is manually stopped. S3 is not used; datasets and checkpoints are kept within 500GB of EBS.

Lease items and costs (monthly)

• Exchange rate basis: 1 USD = 1,400 KRW → about ₩625,226 / month (excluding VAT)
• Annualized: ₩625,226 × 12 = about ₩7,502,712 / year

Calculation formulas

[① EC2 사용료]
시간당 단가 × 월 가동시간 = $2.2454 × (8 hr × 22일)
                        = $2.2454 × 176 hr
                        = $395.19

[② EBS 스토리지]
GB 단가 × 용량 = $0.0928 × 500 GB = $46.40

[③ S3 스토리지]
미사용 = $0.00

[④ 아웃바운드 데이터 전송]
GB 단가 × max(사용량 - 100GB 무료, 0) = $0.126 × max(50 - 100, 0)
                                    = $0.00

[⑤ 부대비용 (탄력적 IP, CloudWatch)]
정액 ≈ $5.00

[월 합계]
TOTAL_USD = ① + ② + ③ + ④ + ⑤
          = 395.19 + 46.40 + 0.00 + 0.00 + 5.00
          = $446.59

TOTAL_KRW = TOTAL_USD × 환율(1,400)
          = ₩625,226 / 월

Billing model by instance state

• Monthly bill if not manually stopped (running 24/7): $2.2454 × 720 + $46.40 + \(5 = <strong markdown="1">\)1,668.09 (₩2,335,326)
• Operating with stop secures about 73% cost savings

Reason for lease

• A model training environment is needed to develop the deliverable "Sensor-based shape recognition system" for stage N, year N of this task
• Compared with the in-house workstation GPU (RTX 5090 32GB), securing 48GB VRAM enables large config training and multi-scenario experiments
• The L40S (sm_89) is compatible without modification with the prebuilt wheels of Pointcept's officially recommended environment (CUDA 12.4 + PyTorch 2.5)
• Given the irregular nature of the development workload, a lease billed only for hours used is more efficient than purchasing an asset

Lease details

Please approve this request to lease an AWS GPU instance for building the model development environment.

Notes

• Unit prices are AWS Seoul region list prices as of 2026-04-20; a conservative exchange rate (1,400 KRW/USD) is applied
• EBS costs are incurred even in the stopped state, so snapshotting and then deleting the EBS volume during long periods of non-use is recommended
• To monitor the monthly bill, also set up an AWS Budgets alarm (notify when over $500)

Recommended phased transition

1. Months 1–2: Stabilize the Pointcept environment on EC2 g6e.2xlarge + validate training scripts
2. Month 3 onward: Package the same environment as a Dockerfile → upload to ECR → move to SageMaker Training Jobs
3. Full-scale training phase: Scale up with SageMaker + Spot Training + multi-GPU (ml.p4d.24xlarge)

Another point that strengthens the approval request is that SageMaker meshes naturally with this task's MLOps pipeline deliverables (Model Registry, AutoML, Drift detection, etc.), so it can be justified as part of the task deliverables, beyond simple infrastructure savings.

Include the AWS Pricing Calculator

It is a good idea to do the final calculation on the AWS Pricing Calculator page, print it to PDF, and attach it.

See also

• AWS
• Korean: 구매 요청서