2nd/10_Wiki/Topics/AI_and_ML/Cost-Benefit Analysis in AI.md

---
id: wiki-2026-0508-cost-benefit-ai
title: Cost-Benefit Analysis in AI
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [AI ROI, AI cost analysis, total cost of ownership, FinOps for ML, model economics, LLM cost]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [ai-economics, roi, finops, mlops, llm-cost, total-cost-ownership, business-strategy]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: business analysis
  applicable_to: [AI Strategy, MLOps, Cost Engineering, Product Decision]
---

# Cost-Benefit Analysis in AI

## 매 한 줄
> **"매 model parameter 의 cost 의 value 의 정당화?"**. 매 GPU bill + 매 data + 매 MLOps 의 cost vs 매 revenue / time saved / risk reduce. 매 modern: LLM token economics, 매 build vs buy vs API, 매 sustainability.

## 매 핵심

### 매 cost component

#### Direct
- **Compute**: GPU/TPU hour.
- **Storage**: model weight + data + log.
- **Inference**: 매 token / image cost.
- **Training**: 매 fine-tune.
- **Network**: 매 egress.

#### Indirect
- **Data**: 매 collect, label, clean.
- **MLOps**: 매 platform team.
- **Engineering**: 매 develop, integrate.
- **Maintenance**: 매 retrain, monitor.
- **Compliance**: 매 audit, security.

#### Opportunity
- **Slow latency**: 매 user lose.
- **Failure modes**: 매 incident.
- **Vendor lock-in**.

### 매 benefit component

#### Quantitative
- **Revenue**: 매 conversion ↑.
- **Cost saving**: 매 labor automate.
- **Time**: 매 turnaround ↓.
- **Quality**: 매 error ↓.

#### Qualitative
- **UX improvement**.
- **Brand / competitive moat**.
- **Compliance protection**.
- **Strategic optionality**.

### 매 LLM economics (modern)
- **API**: 매 OpenAI / Anthropic — 매 per-token.
- **Self-host**: 매 Llama / Mistral — 매 hardware + ops.
- **Managed**: 매 Bedrock / Azure OpenAI — 매 enterprise contract.
- **Hybrid**: 매 critical = self-host, 매 burst = API.

### 매 cost / 1M token (2026 typical)
| Tier | Input $/1M | Output $/1M |
|---|---|---|
| Frontier (GPT-4o, Claude Sonnet) | 3-5 | 15-20 |
| Mid (Haiku, GPT-4o-mini) | 0.3-1 | 1-5 |
| Open (Llama-3.3 70B via API) | 0.5-1 | 0.7-1 |
| Self-host (estimated, amortized) | 0.05-0.5 | 0.1-1 |

→ 매 self-host 의 break-even = 매 use volume.

### 매 ROI calculation
$$\text{ROI} = \frac{\text{Benefit} - \text{Cost}}{\text{Cost}}$$

매 simple 가, 매 multi-period DCF / NPV 가 더 정확.

### 매 build vs buy vs API
| 옵션 | When |
|---|---|
| API | Low/medium volume + frontier capability |
| Self-host | High volume + cost-sensitive + privacy |
| Build (custom) | Differentiator + IP |
| Buy (vendor) | Generic + fast |

### 매 cost optimization technique
1. **Caching**: 매 same prompt → 매 cached.
2. **Routing** (Mix-of-models): 매 simple → cheap, 매 complex → big.
3. **Batching**: 매 throughput ↑.
4. **Quantization**: 매 INT8 / INT4.
5. **Distillation**: 매 small fine-tune.
6. **Spot / preemptible**: 매 batch only.
7. **Right-sizing**: 매 over-provision X.
8. **RAG vs fine-tune**: 매 cheaper.
9. **Token compression** (prompt engineering).

### 매 sustainability
- 매 CO₂ per inference.
- 매 ML CO2 calculator.
- 매 green compute (Google, Microsoft).

## 💻 패턴

### TCO calculator
```python
def total_cost_of_ownership(
    monthly_inference_count: int,
    cost_per_inference: float,
    monthly_storage_gb: float,
    storage_cost_per_gb: float,
    monthly_engineering_hours: float,
    eng_hourly_rate: float,
    one_time_setup_cost: float,
    months: int = 12,
):
    inference_cost = monthly_inference_count * cost_per_inference * months
    storage_cost = monthly_storage_gb * storage_cost_per_gb * months
    eng_cost = monthly_engineering_hours * eng_hourly_rate * months
    return one_time_setup_cost + inference_cost + storage_cost + eng_cost

tco = total_cost_of_ownership(
    monthly_inference_count=1_000_000,
    cost_per_inference=0.001,  # $0.001 / inference
    monthly_storage_gb=500,
    storage_cost_per_gb=0.025,
    monthly_engineering_hours=80,
    eng_hourly_rate=150,
    one_time_setup_cost=50_000,
)
print(f'12-month TCO: ${tco:,.0f}')
```

### Build vs API break-even
```python
def break_even_volume(
    api_cost_per_call: float,
    self_host_fixed_monthly: float,  # 매 GPU + ops
    self_host_marginal_per_call: float,  # 매 electricity
):
    """매 break-even calls / month."""
    if api_cost_per_call <= self_host_marginal_per_call:
        return float('inf')  # 매 API 의 cheaper always
    return self_host_fixed_monthly / (api_cost_per_call - self_host_marginal_per_call)

# 매 example
break_even = break_even_volume(
    api_cost_per_call=0.005,
    self_host_fixed_monthly=4000,  # 매 1× A100 + ops
    self_host_marginal_per_call=0.0001,
)
print(f'Break-even: {break_even:,.0f} calls/month')
```

### LLM routing (multi-model)
```python
class CostAwareRouter:
    def __init__(self):
        self.simple_model = 'gpt-4o-mini'  # 매 $0.15 / 1M input
        self.complex_model = 'gpt-4o'  # 매 $5 / 1M input
        self.judge_model = 'gpt-4o-mini'  # 매 cheap classifier
    
    async def route(self, query: str):
        # 매 1. classify complexity
        complexity = await self.classify_complexity(query)
        
        # 매 2. route
        if complexity == 'simple':
            return await llm.generate(query, model=self.simple_model)
        else:
            return await llm.generate(query, model=self.complex_model)
    
    async def classify_complexity(self, query):
        prompt = f"Classify '{query}' as 'simple' or 'complex'. Reply 1 word."
        return (await llm.generate(prompt, model=self.judge_model)).strip().lower()
```

### Prompt cache (Anthropic / OpenAI)
```python
# 매 Claude prompt caching
import anthropic
client = anthropic.Anthropic()

response = client.messages.create(
    model='claude-sonnet-4-6',
    max_tokens=1024,
    system=[
        {
            'type': 'text',
            'text': LARGE_SYSTEM_PROMPT_10K_TOKENS,  # 매 cached
            'cache_control': {'type': 'ephemeral'},
        },
    ],
    messages=[{'role': 'user', 'content': user_query}],
)
# 매 90% cost reduction on cached portion.
```

### Token cost estimation (Tiktoken)
```python
import tiktoken

def estimate_cost(text: str, model='gpt-4o', kind='input'):
    enc = tiktoken.encoding_for_model(model)
    n_tokens = len(enc.encode(text))
    rates = {
        'gpt-4o': {'input': 5.00, 'output': 15.00},
        'gpt-4o-mini': {'input': 0.15, 'output': 0.60},
    }[model]
    return n_tokens / 1_000_000 * rates[kind]
```

### A/B test ROI measurement
```python
def calculate_ai_lift(control_metrics, variant_metrics):
    """매 A/B test 의 lift 의 calculate."""
    revenue_lift = (variant_metrics['avg_revenue'] - control_metrics['avg_revenue']) / control_metrics['avg_revenue']
    
    monthly_users = variant_metrics['user_count']
    monthly_lift_revenue = monthly_users * (variant_metrics['avg_revenue'] - control_metrics['avg_revenue'])
    
    annual_lift = monthly_lift_revenue * 12
    
    return {
        'revenue_lift_pct': revenue_lift * 100,
        'monthly_lift_$': monthly_lift_revenue,
        'annual_lift_$': annual_lift,
    }
```

### Batch vs real-time decision
```python
def latency_cost_tradeoff(latency_sla_ms, traffic_qps):
    if latency_sla_ms > 5000 and traffic_qps < 10:
        return 'batch'  # 매 spot, async
    if latency_sla_ms < 100:
        return 'realtime + warm + autoscale'
    return 'standard online'
```

### Sustainability tracker
```python
def co2_per_inference(model_size_b, gpu_tdp_w, tokens_per_sec, grid_g_per_kwh=400):
    """매 estimate CO2 / token."""
    energy_per_token_wh = gpu_tdp_w / tokens_per_sec / 3600  # 매 Wh
    co2_g = energy_per_token_wh / 1000 * grid_g_per_kwh
    return co2_g  # 매 grams CO2 / token

# 매 GPT-4o: 매 ~0.0001 g CO2 / token (estimated).
```

## 🤔 결정 기준
| 상황 | Approach |
|---|---|
| Low volume | API |
| High volume + privacy | Self-host |
| Mixed | Routing (cheap + expensive) |
| Predictable batch | Spot + offline |
| Real-time | Cache + warm + ANN |
| Frontier capability | API (latest) |
| Cost-sensitive | Open model + quantization |

**기본값**: 매 routing + 매 caching + 매 batching + 매 right-size.

## 🔗 Graph
- 부모: [[Business-Strategy]] · [[FinOps]] · [[MLOps]]
- 응용: [[LLM_Optimization_and_Deployment_Strategies|Quantization]] · [[RAG]]
- Adjacent: [[Batch-Inference]] · [[Bottlenecks]] · [[Bayesian-Optimization]] (hyperparam ROI) · [[Bioenergetics]] (energy)

## 🤖 LLM 활용
**언제**: 매 AI strategy. 매 build vs buy decision. 매 cost optimization. 매 vendor selection.
**언제 X**: 매 research / experiment (different metric).

## ❌ 안티패턴
- **Vanity model**: 매 frontier 의 unnecessary use.
- **No caching** (repeat prompt): 매 huge waste.
- **Single model 의 모든 task**: 매 cost ↑.
- **No A/B**: 매 ROI 의 prove X.
- **Hidden cost** (egress, monitoring): 매 surprise.
- **No sustainability tracking**.

## 🧪 검증 / 중복
- Verified (FinOps Foundation, ML CO2 papers, OpenAI / Anthropic pricing).
- 신뢰도 A.
- Related: [[Batch-Inference]] · [[MLOps]] · [[Bottlenecks]] · [[Antifragility]] · [[Axify]].

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — TCO + LLM economics + 매 routing / caching / break-even / CO2 code |