Why AWS Bedrock Pricing Destroys Margin at Scale

The Margin Trap Inside AWS Bedrock’s Per-Token Billing

Your SaaS product launched an AI summarisation feature six months ago, powered by Claude 3 Sonnet through AWS Bedrock. Early adoption was modest — 500 users generating maybe 10,000 summaries per month. The Bedrock bill barely registered at $340. Then the feature went viral within your user base. Six months later, 15,000 users generate 800,000 summaries monthly. The Bedrock invoice reads $27,200. Your AI feature charges users $15 per month, generating $225,000 in revenue against $27,200 in inference costs — a healthy 88% margin. But your growth projections show 100,000 users within a year, and at that scale the Bedrock bill hits $181,000 monthly against $1.5 million in revenue. Your margin has dropped to 88% — wait, that still sounds fine. Except your infrastructure, engineering, and support costs also scale, and that $181,000 per month on a single API is now your largest variable cost line item, one you cannot negotiate, optimise, or control.

The problem with per-token pricing is that it scales linearly with usage while your revenue growth eventually plateaus. Dedicated GPU infrastructure flips this equation — costs grow in steps while throughput scales with hardware, not billing meters.

Bedrock Pricing vs. Dedicated at Scale

The Three Margin Killers

1. Compounding token costs. As your AI features get smarter, they use more tokens. Adding context windows, chain-of-thought reasoning, or retrieval-augmented generation multiplies your token consumption per request. A simple summarisation that used 2,000 tokens per request becomes a RAG-enhanced pipeline consuming 8,000 tokens. Your per-request cost quadruples without any increase in user value or pricing power.

2. No volume discounts that matter. AWS offers reserved capacity through Provisioned Throughput, but the discounts are modest and require upfront commitment. You’re still fundamentally paying per token — just at a slightly lower rate. There’s no marginal cost reduction at scale comparable to what dedicated hardware provides.

3. Model lock-in prevents optimisation. Bedrock offers specific models at specific prices. If a new open-source model achieves 90% of Claude’s quality at 10% of the compute cost, you can’t deploy it on Bedrock. On dedicated GPUs, you swap models freely, running open-source alternatives that match your quality requirements at a fraction of the cost.

The Dedicated GPU Margin Advantage

Dedicated GPU costs grow in discrete steps — you add servers as needed — while processing capacity within each server is essentially unlimited by billing. A single RTX 6000 Pro 96 GB running a 70B model through vLLM handles roughly 1,500-2,500 requests per hour for typical summarisation workloads. That’s over a million requests per month from a single $1,800 server. Your cost per request drops asymptotically toward zero as utilisation increases.

Use the LLM cost calculator to model your specific workload, or compare directly with the GPU vs API cost comparison tool.

Protecting Margin as AI Features Scale

Per-token API pricing is a bet that your AI usage will stay small. For startups building AI-native products, that bet loses every time. Moving inference to dedicated GPU infrastructure converts your largest variable cost into a predictable fixed expense, preserving the margin that funds your growth.

Explore the alternatives section for provider-specific comparisons, read the cost analysis guides for deeper financial modelling, or check private AI hosting for data-sensitive enterprise deployments. Browse tutorials for migration walkthroughs.

Filed under: Alternatives