When does your own GPU beat Amazon Bedrock?
Per-token pricing against a GPU that bills every hour. Put in your volume and prices, see the monthly cost on both sides and the token volume where self-hosting starts to pay off.
- Llama 3.1 8B and 3.3 70B presets
- AWS list prices
- Sources linked
- Free
Your volume, your break-even.
Self-hosting vs Amazon Bedrock: break-even calculator
Runs in your browser. Presets use AWS list prices in us-east-1 checked in August 2026[1][2][3]. Change any value to your own.
Check my numbers in a free Idea Call Get a written break-even sheet
The break-even is a utilisation requirement.
A self-hosted model costs the same whether it answers one request or a million: the GPU bills every hour[2]. Bedrock bills per token[1]. So the question is not whether a GPU can serve your model, but whether your traffic keeps it busy all month.
For Llama 3.1 8B at USD 0.22 per million tokens against one g6e.xlarge at USD 1.861 per hour, the break-even is about 6.18 billion tokens a month, or roughly 2,350 tokens per second without a break. One L40S GPU can serve an 8B model faster than that[5], so capacity is not the limit. Traffic is.
The calculator leaves out what pushes the break-even higher: a second instance for redundancy, storage, data transfer[9], monitoring, upgrades and the people who run it. Reserved capacity or Savings Plans[3][8] push it lower.
Three presets, checked against the price lists.
us-east-1, on-demand unless stated, 730 hours per month. From my lecture on AI architectures; re-check the AWS pages before you decide.
| Case | Bedrock | EC2 per month | Break-even |
|---|---|---|---|
| Llama 3.1 8B, 1 × g6e.xlarge | USD 0.22 / 1M tokens | USD 1,358.53 | 6.18 bn tokens |
| Llama 3.3 70B, 1 × g6e.12xlarge | USD 0.72 / 1M tokens | USD 7,657.70 | 10.6 bn tokens |
| Llama 3.3 70B, 1 × g6e.12xlarge, 1-year reserved | USD 0.72 / 1M tokens | USD 4,825.30 | 6.7 bn tokens |
A tuned open-weight model does not force self-hosting: Amazon Bedrock Custom Model Import serves supported models and bills for the time the model copies are active[7].
Often the answer is routing, not moving.
If only part of the traffic must stay on your hardware, route by data class: a managed model for everything that may leave, a self-hosted model for what may not.
Get the break-even for your case.
One workshop with your architects and whoever owns the budget. Your traffic, your models, your prices, redundancy and operations included.
AI cost and break-even review
Intake with your usage data, a 3 to 4-hour workshop, written results within 5 business days.
€2,000 each, 3 to 4 hours, remote or on-site
Delivered within 5 business days after the last session.
Cost review
The AWS services behind this work.
Plain-language notes on the services I use for this kind of project. Each name links to the official AWS documentation.
Amazon Bedrock
Managed access to foundation models from several providers through one API. You choose the model and write the prompt; AWS runs the model.
Amazon EC2 G6e instances
GPU instances with NVIDIA L40S cards for self-hosted inference, for example with vLLM.
Red Hat OpenShift Service on AWS (ROSA)
Managed OpenShift clusters in your AWS account. The base for running OpenShift AI next to other AWS services.
Amazon SageMaker AI
Managed platform for the whole machine learning lifecycle: notebooks, training jobs, model registry and endpoints.
AWS Pricing Calculator
Builds a cost estimate per service before anything is deployed.
Amazon CloudWatch
Logs, metrics and alarms. For AI systems: latency, errors and cost per call.
What people ask before booking.
Is self-hosting an LLM cheaper than Amazon Bedrock?
Only when the GPU stays busy. With list prices, Llama 3.1 8B on one g6e.xlarge breaks even at about 6.18 billion tokens a month before operations cost. Below that, Bedrock is cheaper.
Why does throughput not matter more?
Because a single modern GPU can serve a small model faster than the break-even rate. The limit is how much traffic you have, not how fast the GPU is.
What does the calculator leave out?
Storage, data transfer, monitoring, a second instance for redundancy, platform upgrades and staff time. Add operations cost in the last field to include people.
Can I run my own tuned model on Bedrock?
For supported architectures, yes, with Custom Model Import. It bills while model copies are active, which changes the comparison again.
Where do the prices come from?
AWS list prices for us-east-1, checked in August 2026 and linked in the sources. Prices differ by Region and change over time.
Linda Mohamed
I design and build AI and cloud systems on AWS and hybrid platforms, and I teach how they work. I have organised the AWS User Group Vienna for more than seven years, co-organise AWS Community Day DACH, teach AI architectures at Hochschule Burgenland, speak at conferences in Europe and the US, and maintain ai-solutions.wiki. Based in Vienna. Remote or on-site, in English or German.
Where the facts come from.
Prices, service names and limits change. The linked official pages are the reference; figures on this page were checked in September 2026.
- AWS: Amazon Bedrock pricing
- AWS: Amazon EC2 On-Demand pricing
- AWS: Amazon EC2 Reserved Instances pricing
- AWS: Amazon EC2 G6e instances
- NVIDIA NIM for LLMs: benchmarking and performance
- vLLM documentation
- AWS: Custom Model Import pricing in Amazon Bedrock
- AWS Savings Plans
- AWS Architecture Blog: Overview of data transfer costs
- ai-solutions.wiki: AI total cost of ownership
- ai-solutions.wiki: Capacity planning for AI
Start with one conversation.
30 minutes to look at your idea, your data and the right starting point. Or book 15 minutes if you only have a question.
- Your real token volume and traffic shape
- Which models you need
- Where the data may run
- Whether routing beats moving
Vienna, Austria · remote across Europe · [email protected]
