AWS Credits Are Free Until They Are a $200K Exit Cost

AWS Activate gives qualifying startups up to $100,000 in credits. The application takes 20 minutes. The credits are real — they cover compute, storage, databases, and most managed services. For a pre-revenue startup, this is genuinely useful. It also starts a dependency clock that most developers do not notice until the credits are gone and the first real bill arrives.

This is not a conspiracy. It is a business model. Understanding it makes you a better architect and a less expensive customer.

How the Credits Work

AWS Activate has two tiers. The Founders tier gives up to $1,000 in credits to anyone with a startup idea. The Portfolio tier, available through accelerators, VCs, and incubators, gives up to $100,000. Azure for Startups gives up to $150,000. Google for Startups Cloud Program gives up to $200,000 for AI-focused startups. All three programs renew annually for qualifying companies and provide technical support alongside the credits.

The credits are designed to cover your infrastructure for 12–24 months during the product-building phase. During that time, you architect your product on their platform. You use their managed databases, their serverless functions, their AI APIs, their queuing systems. You hire engineers who know their tools. Your team builds operational knowledge around their console, their CLI, their monitoring stack.

None of this is bad. The services are genuinely good. The problem is that each decision that makes architectural sense during the credits phase adds to the exit cost you will pay if you ever want to leave.

The Exit Cost Breakdown

Here is the actual math for a mid-scale SaaS running on AWS — not a hyperscaler, a real $1M–$10M ARR company:

Data egress: $0.09/GB for the first 10TB/month out of AWS to the internet (pricing as of early 2026, with small discounts at higher volumes). If your application transfers 50TB/month to users, that is $4,500/month just in outbound transfer fees. Moving that data to another cloud requires paying egress on the migration plus rebuilding the destination infrastructure. A one-time migration of 100TB of user data costs approximately $9,000 in egress alone, before engineering time.

DynamoDB migration: AWS DynamoDB is a managed NoSQL database with a specific API. If you are running DynamoDB at scale and want to migrate to Cloud Bigtable on GCP or Cosmos DB on Azure, you are rewriting your data access layer. Estimate 4–8 weeks of engineering time for a production system. At $200/hour fully loaded developer cost, that is $32,000–$64,000.

Lambda function rewrites: AWS Lambda functions use AWS-specific event sources, IAM roles, and SDK calls. Migrating to Google Cloud Functions or Azure Functions requires rewriting event handlers, replacing IAM with equivalent permission models, and retesting all integrations. For a system with 30–50 Lambda functions, estimate 3–6 weeks of work.

Operational retraining: If your team has two years of AWS expertise — knows how to debug CloudWatch logs, understands VPC networking, can read IAM policies — switching platforms means a 3–6 month period of reduced velocity while the team learns equivalent tools on the new platform.

Add it up for a modest mid-scale startup: egress ($9K one-time), database migration ($48K), Lambda rewrites ($40K), operational velocity loss at 30% team productivity for 4 months (~$80K in opportunity cost). You are at roughly $177,000 before you have moved a single user. That is the lock-in.

The Proprietary Services Escalator

The deeper trap is not the credits themselves but the proprietary managed services that AWS, Azure, and Google offer that have no clean equivalent elsewhere.

AWS Bedrock: If you build your AI features on Bedrock — using Claude, Llama, or Titan via the Bedrock API — your application makes AWS-specific API calls with AWS-specific authentication. Migrating to direct Anthropic API or Azure OpenAI requires updating every AI call in your codebase plus changing your auth model.

Amazon Cognito: AWS's managed auth service. Deep integration with API Gateway and Lambda. Moving off Cognito to Auth0, Clerk, or a self-hosted solution requires rewriting your entire auth layer.

AWS Step Functions: Managed workflow orchestration. If your business logic lives in Step Functions state machines, migrating to Temporal, Prefect, or Azure Durable Functions is a significant rewrite.

Each of these services saves real engineering time compared to building equivalent functionality yourself. That is why developers use them. The cost is that each one deepens the exit cost.

How US Tech Ecosystems Are Designed

The credit-to-dependency-to-pricing pattern is not exclusive to cloud. It is the standard US tech ecosystem monetization playbook:

Phase 1 — Low friction onboarding: Credits, free tiers, generous APIs, developer evangelism, open source community contribution. The goal is adoption. Cost matters less than surface area.

Phase 2 — Deepening integration: Proprietary services that solve real problems better than alternatives. The developer chooses them because they are good, not because they are locked in. But each choice adds switching cost.

Phase 3 — Pricing optimization: Credits expire. Usage-based pricing scales with your revenue. Enterprise tiers appear. "Talk to sales" replaces the pricing page. AWS, Azure, and GCP all have a pattern of raising prices on services where they have established dominant market share in a category.

This is not predatory in the conventional sense. All three cloud providers deliver genuine value. The pattern simply means that the marginal cost of staying on the platform is always lower than the cost of leaving, once you are sufficiently integrated.

What Multi-Cloud Actually Means (and Costs)

"Multi-cloud" is frequently recommended as the solution to lock-in. In practice, multi-cloud has its own costs that are often underestimated.

Running the same workload across AWS and GCP simultaneously requires: maintaining two sets of infrastructure-as-code (Terraform modules for each provider), operational expertise on both platforms, data replication with the associated egress costs, and consistent security posture across two different IAM systems.

The companies that successfully run multi-cloud are generally large enough to have dedicated platform engineering teams whose entire job is making multi-cloud work. For a 10-person startup, multi-cloud is usually a complexity cost that exceeds the lock-in risk it is trying to mitigate.

A more practical strategy for most teams: portable application layer, cloud-specific infrastructure. Your application code runs in containers and does not contain AWS-specific SDK calls. Your infrastructure is cloud-specific (managed databases, cloud-native queuing), but your application is cloud-agnostic. This does not eliminate lock-in, but it reduces the rewrite cost significantly if you ever need to move.

The Negotiation Leverage You Do Not Know You Have

If your company is spending $50,000+/month on AWS, you have more negotiating power than your account manager will volunteer. AWS Enterprise Discount Program (EDP) agreements provide committed spend discounts of 10–30% in exchange for 1–3 year commitments. The discount percentage is negotiable. Most companies accept the first offer.

Getting competing quotes from Azure and GCP — even if you have no real intention of migrating — is a legitimate negotiating tactic. Both Azure and GCP have teams whose job is to win AWS customers and they will offer credits, migration assistance, and pricing incentives to facilitate a switch. Use those quotes in the negotiation.

The conversation AWS account managers do not want to have: "Here is a GCP offer for equivalent workload at 25% lower cost. What can you do?" That conversation, when real credits run out and real bills arrive, is worth having.

Key Takeaways

• AWS Activate credits run to $100K — they are real and useful, and they start a dependency clock most developers do not track
• Exit cost math for a mid-scale startup: data egress ($9K one-time), database migration ($48K), function rewrites ($40K), retraining velocity loss ($80K) = approximately $177K before moving a single user
• Proprietary services (Bedrock, Cognito, Step Functions) save engineering time but increase switching cost — every managed service you adopt is a deliberate tradeoff, not just a free convenience
• Multi-cloud is often more expensive than lock-in for teams under 50 engineers — the operational complexity cost exceeds the risk it mitigates
• Portable application layer + cloud-specific infrastructure is the practical middle ground: containers without cloud-specific SDK calls in application code
• $50K+/month spend unlocks EDP negotiation — get competing quotes from Azure/GCP before accepting AWS pricing

Track API and infrastructure costs with LLM API Pricing. Use Will AI Replace Me to assess how AI infrastructure shifts may affect your role.