GenAI on AWS for startups: a low-cost stack, and who pays for it.

IWhat GenAI on AWS costs a startup

You do not need a GPU budget or a platform team to ship GenAI on AWS. You need to know where the money goes, and set a few defaults before traffic arrives.

Most startups build on Amazon Bedrock. It is a managed service that gives you foundation models from Anthropic (Claude), Meta (Llama), Mistral, Amazon (Nova), Cohere, DeepSeek and others through one API. AWS runs the servers. You pay per token, with no minimum. AWS says your data is not used to improve the base models and is not shared with model providers. The full platform guide is at Amazon Bedrock.

A typical startup GenAI app has only a few cost lines: model inference (tokens in and out, usually the largest line), embeddings for the documents you index, the vector store behind retrieval, and reserved capacity if you choose it (Bedrock Provisioned Throughput or SageMaker endpoints). Most runaway bills come from one of these used carelessly: a large model where a small one would do, real-time calls where batch would do, or reserved capacity nobody turned off.

IIA low-cost GenAI stack for startups

This is a plain reference stack a small team can stand up in days: a grounded assistant over your own data on Amazon Bedrock, with cost controls in from the first commit. At early traffic it often costs a few hundred dollars a month or less. Your bill depends on model, Region and traffic, so check the Bedrock pricing page.

Documents live in Amazon S3. A Bedrock Knowledge Base chunks them, creates embeddings, stores them in a vector index and, at query time, retrieves the relevant passages so the model answers from them with citations. The Converse API gives one request format across models, so switching models is a one-line change. A Guardrail filters harmful content and can redact PII.

The most important choice is the default model. For classification, routing, extraction, short answers and drafts, a small model such as Amazon Nova Micro or Nova Lite or Claude Haiku costs far less per token than a large model and is usually good enough. Send only the hard steps to a larger model such as Claude Sonnet or Nova Pro. See Amazon Nova and Claude on Bedrock.

IIIFive choices that keep GenAI cost low

Four of these cost nothing to adopt. They are defaults, not purchases. Only the last one is a commitment, and the advice is to delay it.

• Default to a small model: Send classification, routing, extraction and short answers to a small model. Send only the hard reasoning steps to a larger one. With the Converse API this is a branch in your code, not a second integration.
• Cache repeated context: Chat apps and agents resend the same system prompt, tool schema or document on every turn. Prompt caching bills those repeated tokens at a discount.
• Batch what can wait: Embedding your corpus, nightly summaries and bulk classification do not need an instant answer. Run them as batch inference.
• Retrieve, do not stuff: Pasting whole documents into every prompt costs more as your data grows. A Knowledge Base retrieves only the relevant chunks, so each call stays small.
• Reserve capacity last: Provisioned Throughput and SageMaker endpoints bill by the hour whether you use them or not. Use on-demand until volume is high and steady. Some custom models need Provisioned Throughput to run.

IVCommon GenAI cost traps, and the fix for each

Runaway bills usually come from a few repeat patterns. Each one maps to a choice above.

VBedrock or SageMaker?

For most startup GenAI features, start on Bedrock. Add SageMaker only for a specific need.

Amazon Bedrock fits when you want existing foundation models behind a managed API: chat, RAG, agents, content generation and extraction. There is no cluster to size and no endpoint to keep warm.

Amazon SageMaker fits when you need to own the ML lifecycle: your own model or architecture, custom training, or classical ML such as forecasting, recommendations or a custom vision model. Watch real-time endpoints: they bill by the hour while they run. Both services work in the same AWS account. The full comparison is at Bedrock vs SageMaker.

VIWhat changes by industry

The stack above is the same in every vertical. What changes is the data you handle, the rules you follow and the controls you add.

• SaaS (multi-tenant): Keep each customer's data separate: filter retrieval by tenant ID on the server, never from the client. Use application inference profiles to see cost per tenant, and rate-limit each tenant.
• Fintech and insurance: Expect SOC 2 or PCI DSS reviews to cover the GenAI flow. Log every model call, keep data in the Region you promise customers, and use Guardrails to block advice you are not licensed to give.
• Healthcare: Use HIPAA-eligible services under a Business Associate Addendum (BAA) with AWS, and keep protected health information out of logs. See GenAI on AWS for healthcare.
• Legal: Ground every answer in retrieved documents with citations, and keep client documents separate by matter. Treat output as a draft for a lawyer to review.
• Customer support: Retrieve answers from your help center and past tickets, add Guardrails, and hand off to a person when the model is unsure. See RAG on AWS.
• Edtech: Set Guardrails for age-appropriate content and follow the student-data rules that apply to you. Keep student records out of prompts unless you need them.
• Media and gaming: Text, image and video models on Bedrock (for example Amazon Nova Canvas, Nova Reel or Stability AI) can draft content and assets. Add moderation, and check each model's license terms for commercial use.
• Real estate: Common uses are listing descriptions, Q&A over leases and disclosures, and search by meaning. Ground answers in the documents and keep a person in the loop for anything a client relies on.
• Enterprises: At enterprise scale the same stack adds central governance: shared Guardrails, model access policies, cost allocation by team and private network paths.

VIIWho pays for it: partner-filed funding and Activate

Two kinds of AWS money can cover a startup GenAI build. A partner files one. You apply for the other.

Partner-filed funding. AWS gives its consulting partners funding they can request for qualifying customer work. Your matched partner can file POC (proof of concept) funding to test a GenAI use case, whether or not you run on AWS today. If you already run on AWS and are launching new work, such as an AI feature, a new customer or a new contract, the partner can file incremental workload (IW) funding. If you are moving to AWS from another cloud, it can file MAP (Migration Acceleration Program) funding. For startups this is typically $50K to $150K. Depends on the workload; AWS approves. Only a partner can request it, and it is tied to real work the partner does with you. See AWS partner funding and MAP on AWS.

AWS Activate credits. You apply for these yourself at aws.amazon.com/startups/credits. Activate Founders gives $1,000 to start, up to $5,000 to self-funded startups. Activate Portfolio gives up to $200,000 when a VC firm, accelerator or angel group that is an Activate Provider gives you its Org ID. AWS Credits for AI Startups gives $200,000 or more, by invitation from your AWS account manager. A consulting partner cannot file any Activate tier for you. AWS says applicants hear back within 5 to 10 business days.

The AWS Generative AI Accelerator is a separate, competitive program. AWS picked 40 startups for its 2025 cohort and gave each up to $1 million in credits. AWS selects the startups; no partner files it. See the 2025 cohort.

Which route is larger depends on who backs you. If a VC, accelerator or angel investor backs you, Activate Portfolio can be large, and you should apply. If no investor backs you, Activate Founders tops out at $5,000, and partner-filed funding is usually the larger route. You can use both: Activate for general AWS spend, and partner-filed funding for the build itself. More detail: Bedrock credits and POC funding explained.

VIIIYour first GenAI build on AWS, step by step

Here is the order of work for a grounded, governed, low-cost assistant, with the cost controls in from the start.

The cost controls go in before traffic, not after. That is the difference between a feature that stays cheap and one you have to rebuild the month it gets popular. The model access rules are in the Bedrock user guide.

• Step 1: pick a Region and scope IAM: Pick the Region for data residency and latency. Access to many Bedrock models is on by default. The first call to a third-party model subscribes your account through AWS Marketplace, so the role needs Marketplace permissions once, and Anthropic models need a one-time use-case form. Scope your IAM policy to the model IDs you use.
• Step 2: make a first Converse call: Call a small model through the Converse API. Set maxTokens so output cannot run away. This is your baseline.
• Step 3: add a Knowledge Base: Point a Bedrock Knowledge Base at documents in S3. Run the first embedding pass as batch. You now have RAG without building retrieval yourself.
• Step 4: add a Guardrail: Configure PII redaction and denied topics once, and apply them across models.
• Step 5: add the cost controls: Add model routing, turn on prompt caching for the system prompt and retrieved context, and move offline jobs to batch.
• Step 6: set spend alerts: Tag GenAI resources, set an AWS Budgets alert, and turn on model invocation logging so you can see tokens per feature.
• In parallel: sort out funding: Apply for Activate credits yourself if you qualify. If a partner builds with you, ask it to check whether the work qualifies for POC or IW funding before the work starts. The partner files the request; AWS reviews it and decides.