AWS AI/ML Migration Guide: Technical Reference

Migration Reality Assessment

Failure Scenarios and Consequences

• OpenAI billing shock: Started at $300/month for prototype, escalated to $47k/month at scale
• Azure ML reliability failures: Training pipelines failed every few weeks with poor error messages
• Google product discontinuation anxiety: 200+ products killed, including AI Platform Notebooks and ML Engine
• Migration timeline failures: Teams promising 90-day migrations typically require 6-12 months

Critical Success Thresholds

• Minimum spend threshold: Don't migrate if spending <$5k/month on current platform
• Break-even timeline: 6-20 months depending on current spend and migration complexity
• Engineering bandwidth requirement: 2-3 months minimum dedicated engineering time

Technical Migration Specifications

Service Mapping and Costs

Source Platform	AWS Alternative	Migration Difficulty	Timeline	Cost Impact
OpenAI GPT-4	Bedrock Claude 3.5	Medium	6-8 weeks	Save ~$15k/month at scale
OpenAI GPT-3.5	Bedrock Nova Lite	Low	4-6 weeks	Save ~60%
Azure ML	SageMaker	Maximum	4-6 months	Variable, often 2x more initially
Google Vertex AI	SageMaker	Extreme	6-8 months	Usually higher costs
Anthropic Direct	Bedrock Claude	Minimal	2-3 weeks	Pay 20% more

Model Cost Specifications (per 1M tokens)

• Claude 3.5 Sonnet: ~$15 (vs OpenAI GPT-4 at $30)
• Nova Pro: ~$8 (85% quality of GPT-4)
• Nova Lite: ~$2.65 (basic tasks only)
• Llama 3: $2.65 (open source, quality limitations)

Implementation Requirements

Infrastructure Setup Critical Points

• IAM permissions: Minimum 2 weeks setup time, error messages are unhelpful
• Service quotas: Bedrock starts at 10 requests/minute, requires 3-5 day approval for increases
• Regional limitations: Claude 3.5 only in us-east-1 and us-west-2
• Billing alerts: Must set up at $500, $1000, $2500, $5000 thresholds BEFORE migration

Required IAM Policy (Minimal Working Version)

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": [
        "bedrock:InvokeModel",
        "bedrock:InvokeModelWithResponseStream",
        "bedrock:ListFoundationModels"
      ],
      "Resource": "*"
    },
    {
      "Effect": "Allow", 
      "Action": [
        "sagemaker:InvokeEndpoint",
        "sagemaker:DescribeEndpoint"
      ],
      "Resource": "*"
    }
  ]
}

Migration Proxy Implementation Pattern

class MigrationProxy:
    def __init__(self):
        self.use_aws = os.getenv('USE_AWS_PERCENT', 0)
        self.openai_client = openai.OpenAI()
        self.bedrock = boto3.client('bedrock-runtime', region_name='us-east-1')
    
    def chat_completion(self, messages, model="gpt-4"):
        if random.randint(1, 100) <= int(self.use_aws):
            return self._call_bedrock(messages)
        else:
            return self._call_openai(messages, model)

Operational Intelligence

Hidden Cost Factors

• Data transfer: $0.09/GB out of AWS (adds up with large models)
• SageMaker endpoints: $50-200/month each, even when idle
• CloudWatch logs: Can exceed compute costs without proper retention policies
• Model artifacts storage: Accumulates over months of training

Performance Optimization Requirements

• Cold start mitigation: Bedrock models go cold after 5 minutes, first request takes 10-15 seconds
• Keep-warm strategy: Ping every 4 minutes costs ~$5/month, saves user experience
• Batch processing savings: 70% cost reduction vs real-time inference
• Spot instance training: 70% cost savings when properly configured

Critical Failure Points

• Token counting differences: Same prompt costs 20% more/less between providers
• Prompt optimization requirement: Each model needs different prompt styles
• Regional service unavailability: Not all services available in all regions
• Automatic rollback triggers: Error rate >2% for >10 minutes, response time >5 seconds for >5 minutes

Timeline and Resource Requirements

Real Migration Phases

1. Assessment (2-4 weeks): Finding hidden dependencies, baseline metrics
2. Parallel system build (4-8 weeks): IAM setup, proxy layer, testing
3. Gradual rollout (2-6 weeks): 5% → 10% → 25% → 50% → 75% → 100%
4. Optimization (6+ months): Cost optimization, performance tuning

Engineering Resource Allocation

• Simple API migration: 6-8 weeks, 1-2 engineers
• ML platform migration: 4-6 months, 2-4 engineers
• Migration costs: $40k-60k per engineer for 2-3 months
• Parallel system operation: 2x current monthly costs for 3-6 months

Decision Support Framework

Don't Migrate If:

• Current spend <$5k/month
• Team has never used AWS
• No 3+ months engineering bandwidth
• Current system works well and under deadline pressure
• Simple OpenAI integration with just basic API calls

Mandatory Migration Triggers:

• Monthly AI costs >$10k and growing
• Vendor reliability issues causing business impact
• Need for cost predictability and control
• Already using other AWS services extensively

Success Metrics (6-month targets)

• 40-60% cost reduction from pre-migration
• <2 second response times for 95% of requests
• <0.1% error rate
• Team can deploy new AI features in days vs weeks

Risk Mitigation Strategies

Rollback Requirements

• Keep old system running in parallel for minimum 3 months post-cutover
• Automated rollback triggers for error rate spikes
• Feature flags for instant traffic routing
• Emergency contact procedures for 3am failures

Quality Assurance Process

• A/B testing framework for model comparison
• Production prompt optimization for each model
• Response caching for repeated queries (60-80% cost savings)
• Continuous monitoring of quality vs cost metrics

Vendor Lock-in Mitigation

• Multi-cloud strategy for critical workloads
• Standardized API layer for easy provider switching
• Local model deployment capabilities as backup
• Regular cost and performance benchmarking

Common Anti-Patterns

Technical Debt Creation

• Using Resource "*" permissions in production
• Not implementing proper error handling and retries
• Skipping cost monitoring and alerting setup
• Manual rollback processes instead of automated

Project Management Failures

• Promising heroic timelines (90 days instead of 6 months)
• Running single system instead of parallel migration
• Not accounting for learning curve and team training
• Ignoring compliance and security requirements

Cost Optimization Mistakes

• Not using spot instances for training (70% savings missed)
• Running idle SageMaker endpoints ($200+/month waste)
• Over-provisioning instances (teams typically over-provision by 50%)
• Not implementing request caching for repeated queries

This technical reference provides the operational intelligence needed for successful AWS AI/ML migration while avoiding the common pitfalls that cause projects to fail or exceed budgets.