Vercel vs AWS for AI agents in 2026: when Vercel is enough and when to move to your own account
Vercel vs AWS for AI agents: when Vercel's AI SDK, Fluid compute and Sandbox are enough, when agents belong in your own AWS account, and what each costs from public price pages.
Vercel is enough for AI agents that live inside a web app, call hosted models, finish within Vercel's function limits (800 seconds on Pro, 4 GB of memory) and keep state in a Marketplace database. The AI SDK, AI Gateway, Fluid compute and Sandbox make that path faster than anything you would assemble on AWS.
Move agents to your own AWS account when one of five things happens: they run all day as workers and need previews or autoscaling, they must reach private data inside your VPC, a customer requires data residency in your account, they need GPUs or custom instance shapes, or the bill is big enough that Savings Plans matter.
Costs depend on how busy the agent is. From public price pages: an always-on worker with 1 vCPU and 2 GB costs about $124 a month on Vercel Daemons at full CPU and about $40 at 10% busy, versus about $36 on AWS Fargate (x86, us-east-1) at any utilization, before AWS's fixed networking costs such as a NAT gateway at about $33 a month.
Vercel BYOC does not move the whole agent. It runs Vercel Functions as Lambda in your AWS account on Enterprise plans, while traffic still passes through Vercel and workers, databases and sandboxes stay outside.
You do not need to rebuild a platform to leave. Keep the frontend on Vercel, move agents, workers and databases into your AWS account, and use a platform like Qovery for git-push deploys, preview environments and scoped agent permissions.
"Should we just run this on AWS?" is the question I hear most from teams whose agent prototype on Vercel started getting real traffic. It usually comes from the person who saw the first invoice, or from the person who filled in the first enterprise security questionnaire.
There is no single right answer. Vercel and AWS solve different problems for agents, and plenty of teams end up using both. In this article I compare them on what an agent needs in production, show when Vercel is enough and when moving to your own AWS account pays off, and do the cost math using only public price pages, so you can redo it with your own numbers.
Vercel vs AWS for AI agents: what is the short answer?
Use Vercel when the agent is part of a web product and fits inside serverless limits. Use your own AWS account when the agent is a long-running service that touches private data, needs to scale on its own, or has to sit inside your compliance boundary.
Vercel is a managed platform. In November 2026 its homepage headline reads "Agentic Infrastructure", and it ships the pieces an agent builder wants on day one: an SDK, a model gateway, compute billed on active CPU, durable workflows and microVM sandboxes. You get speed and you accept Vercel's limits and Vercel's bill.
AWS is a cloud provider. It gives you every building block (Lambda, Fargate, EKS, RDS, SQS, Bedrock) and none of the assembly. You get control, your own network, your own discounts and your own audit scope, and you take on the work of wiring it together and running it.
Vercel
Your own AWS account
What it is
Managed platform for web apps, backends and agents
Cloud provider with raw building blocks
Agent runtime
Fluid compute functions (up to 800s on Pro), Workflows, Daemons (public beta)
Lambda (up to 15 minutes), Fargate or EKS (no time limit), Bedrock AgentCore
Untrusted code
Vercel Sandbox (Firecracker microVMs)
You assemble it (e.g. Firecracker, gVisor on EKS)
Model access
AI SDK and AI Gateway
Amazon Bedrock, or direct provider APIs
State and databases
Marketplace partners (Neon, Upstash, Supabase)
RDS, Aurora, DynamoDB, ElastiCache in your account
Network
Vercel's network; Secure Compute adds a dedicated VPC that Vercel runs (Enterprise)
Your VPC, private subnets, PrivateLink
Previews
Every push gets a preview URL (Daemons excluded during beta)
You build them, or use a platform
Billing
Vercel invoice, usage-based
AWS invoice; Savings Plans and credits apply
Who operates it
Vercel
Your team, or a platform on top of your account
What does Vercel give an AI agent out of the box?
Vercel gives a TypeScript or Python agent a complete loop: build with the AI SDK, route model calls through AI Gateway, run on Fluid compute, pause long tasks in Workflows, and execute generated code in Sandbox. Every piece is on one platform with one dashboard.
The AI SDK handles model calls, streaming and tool definitions. AI Gateway puts many model providers behind one endpoint, with fallbacks and spend tracking. Fluid compute bills functions on Active CPU, so the seconds your agent spends waiting for a model to answer do not count as CPU time. You still pay for provisioned memory while the function runs.
Limits are published. With Fluid compute, functions run up to 300 seconds on Hobby and 800 seconds on Pro and Enterprise, with a 30-minute extended maximum in beta, and use up to 2 GB on Hobby or 4 GB and 2 vCPUs on Pro (Vercel Functions limits). For work that runs longer, Vercel points you to Workflows, which pause and resume durable steps for "minutes to months".
For agents that run code, Vercel Sandbox starts isolated Firecracker microVMs, with snapshots, persistent sandboxes and Drives for persistent storage. For workers that poll a queue all day, Daemons run container images continuously. Daemons are in public beta since October 15, 2026, and the beta comes with real limits: production only, no preview deployments, no autoscaling, no persistent storage, and instance counts set by hand per region.
That is a lot of agent infrastructure for a team of three. If I were prototyping an AI feature in a Next.js app tomorrow, I would start there.
What does AWS give an AI agent?
AWS gives you every primitive an agent could need and expects you to compose them. That composition is the cost of entry, and it is also the reason teams move.
The usual building blocks for an agent on AWS are Lambda for short event-driven steps (15 minutes maximum per invocation), Fargate on ECS or EKS for long-running containers, EKS when you want Kubernetes, SQS or EventBridge for triggers, RDS or Aurora for state, and S3 for artifacts. For the model layer, Amazon Bedrock gives access to several model families, and Bedrock AgentCore adds a managed agent runtime, memory and identity. We looked at where AgentCore stops in Bedrock AgentCore and the agent governance gap.
What you get in exchange for the assembly work:
Your own VPC. The agent reaches your private databases, internal APIs and data warehouse without exposing them to the internet.
Your own audit scope. CloudTrail, IAM and your region choice, which is what an enterprise customer's security team asks about.
Your own discounts. Compute Savings Plans cut Fargate prices by up to 50%, and Fargate Spot by up to 70% for interruption-tolerant tasks (AWS Fargate pricing).
No platform ceiling. Any instance shape, GPUs, any runtime, any duration.
What AWS does not give you is the developer workflow. Git-push deploys, a preview URL per pull request, rollbacks and environment variables per branch are all things you build or buy.
When is Vercel enough for AI agents?
Vercel is enough when all five of these are true. I would not move an agent that passes them.
The agent lives inside a web product. It answers a user, streams a response, maybe calls a few tools, and its UI is a Next.js or similar frontend already on Vercel.
Each run fits the limits. Tool loops finish within 800 seconds, or you can split them into Workflow steps.
State is light or fine in a Marketplace database. Conversation history in Neon or Upstash raises no compliance question for your customers.
The agent calls hosted models. No GPU, no local inference.
Nothing has to stay inside your network. The agent does not query a private database or an internal API that you cannot expose.
If you also run untrusted code, Vercel Sandbox is one of the better options on the market, and it stays a good choice even after other parts of the agent move.
When should you move AI agents to your own AWS account?
Move when the agent stops being a feature of a web app and becomes a backend service with its own lifecycle. In practice, one of these five signals shows up first.
The agent runs all day. A queue consumer, a scheduled research agent, a support triage worker. Daemons can run it, but during the beta you cannot test it in a preview deployment, it does not autoscale, and it has no persistent disk. On Fargate or EKS in your account, it is an ordinary service.
The agent needs private data. If it reads from a database in a private subnet, or calls an internal API, it should run next to that data. Vercel's Secure Compute gives functions a dedicated network with VPC peering, but it is an Enterprise feature, the network is one Vercel runs, and it applies to functions.
A customer asks where the data lives. Enterprise security reviews ask for the account, the region, the encryption keys and the subprocessors. "In our AWS account in eu-west-1" is a short answer. "In a Vercel function, with history in a partner database" is a longer one.
The agent needs hardware Vercel does not sell. GPUs for local inference, memory-heavy embedding jobs, or an instance type your data science team already standardized on.
The bill crosses the line where commitments matter. Once agent compute is a real line item, the Savings Plans and credits you already negotiated with AWS cannot apply to a Vercel invoice.
Vercel's own answer to the first and third signals is Bring Your Own Cloud, now generally available on AWS for Enterprise plans. It runs your Vercel Functions as AWS Lambda functions in your AWS account. It supports Node.js and Python, excludes Edge functions, container functions and Services, and keeps Vercel in the data path: TLS terminates at the Vercel CDN, and requests and responses pass through Vercel. Your team runs the builds and owns IAM, quotas and cleanup. If only the agent's function code has to run in your account, that can work. If the agent comes with a worker and a database, those stay outside, and you are back to running them on AWS anyway. We went through the details in what Vercel BYOC keeps outside your cloud.
Move your agents to your own AWS account without building a platform.
Qovery deploys AI agents, workers and databases into your own AWS account, or GCP, Azure, Scaleway or an existing Kubernetes cluster, with git-push deploys and a preview environment per pull request.
How do Vercel and AWS compare on cost for AI agents?
On unit prices, AWS is cheaper for compute that is busy most of the time, and Vercel is competitive or cheaper for agents that mostly wait on models. AWS then adds fixed networking costs that Vercel includes. Every number below comes from a public price page, at list price, in us-east-1 (iad1 for Vercel), checked November 19, 2026.
First 100 GB free, then $0.09/GB for the first 10 TB (EC2 pricing)
NAT gateway
Not applicable
$0.045 per hour plus $0.045 per GB processed (VPC pricing)
Load balancer
Included
ALB: $0.0225 per hour plus $0.008 per LCU-hour (ELB pricing)
Public IPv4 address
Included
$0.005 per hour each
Commitment discounts
None
Compute Savings Plans up to 50% on Fargate; Fargate Spot up to 70%
Worked example: an always-on agent worker
Take a worker with 1 vCPU and 2 GB of memory, running 730 hours a month. These are my assumptions, so swap in your own.
Vercel Daemons, CPU busy 100% of the time: 730 × $0.128 = $93.44 of CPU, plus 2 × 730 × $0.0212 = $30.95 of memory. About $124 a month.
Vercel Daemons, CPU busy 10% of the time (an agent that mostly waits on model calls): $9.34 of CPU plus $30.95 of memory. About $40 a month.
AWS Fargate, x86: 730 × $0.04048 = $29.55 of CPU, plus 2 × 730 × $0.004445 = $6.49 of memory. About $36 a month, whatever the utilization, because Fargate bills the allocation. On Arm (Graviton) it drops to about $29.
Then add AWS's fixed costs, which are shared by every service in the VPC and do not scale with the number of agents. One NAT gateway is about $32.85 a month before data processing, one ALB about $16.43 before capacity units, and an EKS control plane $73 if you choose Kubernetes. For a single agent, those fixed costs can erase the unit-price advantage. For twenty services, they disappear into the noise, and Savings Plans start cutting the larger line.
Worked example: a request-driven agent
Request-driven agents flip the comparison. Lambda's classic pricing bills duration from start to finish, including the seconds your function waits for a model to answer. Vercel's Fluid compute bills Active CPU, and waiting does not count. For an I/O-bound agent that spends most of each run waiting on an LLM, that difference can make Vercel cheaper than Lambda at the same traffic. AWS's Lambda pricing page now also lists Web Functions billed on active CPU time, so check the current rates before you assume either way.
Two costs are missing from both columns. Model tokens cost the same whether the call leaves Vercel or AWS, unless you move to Bedrock with a negotiated rate. And engineering time on AWS is real: someone builds the deploy pipeline, the preview environments and the IAM model, or you pay a platform to do it.
How do you move AI agents from Vercel to your own AWS account?
Move the agent's backend first, keep the frontend on Vercel, and cut over one agent at a time. Most teams do not need a big-bang migration.
Inventory the agent. List its functions, Workflows, cron jobs, Daemons, environment variables, Marketplace databases, and every external API it calls.
Containerize the long-running parts. Workers and agent loops become containers. Short request handlers can stay functions or become an API service.
Move state. Export the Marketplace database into RDS or Aurora in your account with a dump and restore or logical replication, and plan a short read-only window.
Run both in parallel. Point a share of traffic at the AWS version, compare latency, error rate and cost over a full billing cycle.
Keep Vercel for what it does best. The Next.js frontend, AI SDK streaming to the browser and Sandbox can all stay. The frontend calls your AWS backend over authenticated HTTPS.
Cut over and keep a rollback. Switch the agent's endpoint or queue, and keep the Vercel version deployable for a week.
Not if a platform operates the account for you. This is the part of the move that worries teams most, because Vercel's workflow is what they like and raw AWS has none of it.
Qovery deploys into your own AWS account (or GCP, Azure, Scaleway, or an existing Kubernetes cluster) and gives you back the workflow: git-push deploys, a preview environment per pull request with the database included, managed databases such as Amazon RDS through Blueprints, auto-stop for non-production environments, and Karpenter for node autoscaling on AWS. The control plane is Qovery's, and the workloads, the data and the AWS bill stay yours, so your Savings Plans apply.
For agents specifically, Qovery adds the controls that matter once an agent deploys code. API Policy Tokens scope each agent to exactly the actions it needs with an OPA policy, and every action shows up in the audit log under that agent's own token (per-agent attribution). The MCP Server is read-only by default. Qovery is not an execution sandbox, so if your agents run untrusted code, keep Vercel Sandbox or another sandbox for that step.
The Free plan covers 3 users plus unlimited AI agents, one Kubernetes cluster of your own and 1,000 deployment minutes a month, with no credit card.
To let your coding agent do the setup, install the Qovery Agent Skills and the MCP Server:
TEXT|Agent Skills: one install, every compatible agent (Claude Code, Cursor, Codex, OpenCode, Gemini CLI...)
curl -fsSL https://skill.qovery.com/install.sh | bash
# MCP Server for Claude Code: OAuth, read-only by default
claude mcp add --transport http qovery https://mcp.qovery.com/mcp --callback-port 4242
The skills deploy (qovery-onboard, qovery-deploy, qovery-preview and others), and the MCP Server operates what already exists. Then ask your agent "move this agent worker and its Postgres database from Vercel to my AWS account with Qovery". The AI agent quickstart has the full steps. For the wider landscape of options, see our Vercel alternatives guide.
Frequently asked questions
Is Vercel or AWS better for AI agents?
Vercel is better for agents that live inside a web app, call hosted models and finish within its function limits, because the AI SDK, AI Gateway, Fluid compute and Sandbox work together out of the box. AWS is better for long-running agents that need private data, data residency, GPUs or commitment discounts. Many teams use both, with the frontend on Vercel and the agent backend in their AWS account.
Is Vercel more expensive than AWS for running AI agents?
It depends on how busy the agent is. At list prices, an always-on 1 vCPU, 2 GB worker costs about $124 a month on Vercel Daemons at full CPU and about $40 at 10% busy, versus about $36 on AWS Fargate at any utilization. AWS then adds shared fixed costs such as a NAT gateway (about $33 a month) and a load balancer, and in return lets Savings Plans cut compute by up to 50%.
Does Vercel BYOC let me run AI agents in my own AWS account?
Partly. Vercel BYOC, available on Enterprise plans, runs your Vercel Functions (Node.js and Python) as Lambda functions in your AWS account. TLS still terminates at Vercel's CDN, requests and responses pass through Vercel, and Daemons, Sandbox, containers, Services and Marketplace databases stay outside your account.
How long can an AI agent run on Vercel compared with AWS?
A Vercel Function runs up to 800 seconds on Pro and Enterprise, with a 30-minute extended maximum in beta; Workflows and Daemons cover longer work. AWS Lambda runs up to 15 minutes per invocation, while containers on Fargate or EKS have no time limit.
Can I keep my frontend on Vercel and run the agent backend on AWS?
Yes, and it is the lowest-risk way to move. The Next.js frontend and AI SDK streaming stay on Vercel, while agents, workers and databases run in your AWS account and are called over authenticated HTTPS. Qovery handles the AWS side with git-push deploys and per-pull-request preview environments.
Melanie leads content at Qovery. She covers platform engineering trends, Kubernetes operations, FinOps, and the tools that help engineering teams ship faster.
Next step
Move your agents to your own AWS account without building a platform.
Qovery deploys AI agents, workers and databases into your own AWS account, or GCP, Azure, Scaleway or an existing Kubernetes cluster, with git-push deploys and a preview environment per pull request.