Contents
- 1. Why AI recommends Vercel — 3 reasons
- 2. Six questions to decide
- 3. How far the free Hobby plan goes — rough numbers
- 4. How Pro is billed — what $20 covers and the overage rates
- 5. Five typical ways the bill grows
- 6. Setting a cap with Spend Management
- 7. Four alternatives when it doesn't fit
- Summary
- FAQ
Ask Claude Code or ChatGPT "where should I deploy this web app?" and you'll get an almost reflexive answer: "Push it to Vercel." For experienced developers that's a natural suggestion. For beginners it raises questions: "What is Vercel? It says free — really? How big could the bill get?"
The short answer: for a personal, non-commercial site built with Next.js, starting on Vercel's free Hobby plan is fine. Three conditions decide the rest. ① Hobby is for non-commercial personal use only (running ads, selling something, or building the site as paid work all count as commercial). ② What happens when you exceed the allowance is opposite on Hobby and Pro (Hobby stops; Pro keeps running and moves to usage-based billing). ③ What drives cost is "how long your functions wait" and "how many images and requests you serve". This article works out how many page views and AI calls the Hobby allowance covers, how Pro is billed and at what rates, where bills grow, and how to set a cap — all checked against Vercel's official documentation as of November 12, 2026.
This article's job is deciding whether to use Vercel when an AI recommends it, what it will cost, and where to stop it. What a PaaS is and how the providers compare lives in What is PaaS (Vercel etc.)?, whether to use Next.js at all lives in When AI tells you to use Next.js, and the actual deployment steps live in Deploying to Vercel from Claude Code and Cursor.
Even when AI recommends it, check the plan terms first
— Can you use it for free, what happens past the limit, and where to stop it
Sources: Vercel official documentation (Hobby, Pro, Spend Management). As of November 12, 2026.
1. Why AI recommends Vercel — 3 reasons
Vercel is a PaaS that builds your project and gives it a URL every time you push to GitHub. It bundles static file delivery (CDN), Vercel Functions for server-side code, domains and SSL, and a preview URL per branch. How it differs from shared hosting, a VPS or the cloud is covered in What is PaaS (Vercel etc.)?. There are three structural reasons AI suggests it by reflex.
Why AI reaches for Vercel by reflex
Sources: Next.js "Deploying", React "Creating a React App" (states that Vercel maintains Next.js)
In other words, the AI is picking the path where beginners stumble least. That's reasonable, but it doesn't check whether it fits your use (commercial or not, long-running work, budget). The rest of this article shows where that gap appears, using the numbers from the source documents.
2. Six questions to decide
Don't decide "because the AI said Vercel." Answer these six in order. Q1 is the only hard gate: if the answer is yes, Hobby is not an option. For the rest, if one applies, check the numbers in sections 3 onward before deciding.
Six questions for choosing Vercel
Asking for donations alone is not commercial.
No (a static site, say) → other providers do the same job.
Sources: Fair Use Guidelines, Functions Limits, Flat Rate CDN, Pro (as of November 12, 2026)
The Q1 line is wider than most people expect. Vercel's Fair Use Guidelines define commercial use as a deployment "used for the purpose of financial gain of anyone involved in any part of the production," including a paid employee or consultant writing the code. The listed examples are taking payments from visitors, advertising a product or service, being paid to create, update or host the site, a site whose primary purpose is affiliate linking, and ads such as Google AdSense. A blog that earns a little ad income on the side is commercial under this definition. If you're unsure, the guidelines point you to Vercel Support.
3. How far the free Hobby plan goes — rough numbers
Hobby limits are written in units like "100 GB" and "360 GB-hrs," which are hard to map onto your own site. The table below converts the official allowances into page views and call counts. The second-to-last column depends on the assumptions listed under the table; real numbers depend on how your site is built.
| What is counted | Hobby monthly allowance | Rough conversion | When exceeded |
|---|---|---|---|
| Transfer (Fast Data Transfer) | 100 GB | About 50,000 page views at 2 MB per view | Most features stay off until 30 days have passed |
| CDN requests | 1 million | About 33,000 page views if each page loads 30 files (HTML, JS, images) | Same as above |
| Function invocations | 1 million | 1 million page views for a page that calls one API | Same as above |
| Active CPU | 4 hours | 144,000 calls at 0.1 s of computation each | Same as above |
| Provisioned Memory | 360 GB-hrs | About 32,000 calls for a 2 GB function that waits 20 s for an AI response | Same as above |
| Image Transformations | 5,000 | 100 source images in 5 widths and 2 formats is 1,000 | New images fail with a 402 error and the alt text is shown (already transformed images keep working) |
Assumptions: 2 MB and 30 files per page view is typical of an image-heavy blog; 2 GB is Vercel's default function memory; a 20-second wait is roughly one call to a text-generating AI API. The Provisioned Memory figure uses the most conservative counting, where each call occupies its own instance; when one instance handles concurrent requests, usage goes down. Image transformations are counted again each time a cached image expires and has to be regenerated. Sources: Vercel Hobby Plan, Functions Limits (2 GB default memory), Image Optimization Limits and Pricing.
What this tells you
For an ordinary blog, the CDN request count tends to run out before transfer does. For an app that calls AI, Provisioned Memory runs out first, not CPU. While the function waits for the AI response, CPU billing pauses but memory billing continues, so the allowance shrinks even though the computation is light. If you publish an AI chat on Hobby, watch memory usage on the Usage page for at least the first few days.
4. How Pro is billed — what $20 covers and the overage rates
A Pro monthly bill is "$20 platform fee + extra seats + add-ons + on-demand usage beyond the credit." The $20 fee includes one seat that can deploy and $20 of usage credit. Unused credit expires at the end of the month and does not carry over. You get an automatic notification at 75% of the credit; once it's used up, your team switches to on-demand usage and you start receiving daily and weekly usage summary emails (Vercel Pro Plan).
CDN usage (requests and transfer) is paid through fixed tiers called Flat Rate CDN. Pro includes the lowest tier (1 million requests and 1 TB a month) at no extra cost. If your total for a billing cycle exceeds the tier, you move to a higher tier at the start of the next cycle. A one-day spike doesn't trigger an upgrade.
| Flat Rate CDN tier | CDN requests per month | Transfer per month |
|---|---|---|
| Included with Pro | 1 million | 1 TB |
| $20/month | 10 million | 50 TB |
| $100/month | 50 million | 50 TB |
| $300/month | 150 million | 50 TB |
Source: Flat Rate CDN. Above the top tier you move to on-demand pricing. With Flat Rate CDN turned off, CDN usage is also billed on demand: in Washington, D.C. (iad1), $0.15/GB of transfer and $2.00 per million requests. Bulk file distribution, media delivery that makes up most of the bandwidth, and using the CDN as file storage are out of scope for Flat Rate CDN.
Everything other than the CDN is billed on demand once the credit is used. These are the rates for the items that most often show up on a beginner's bill.
| Item | What is counted | On-demand rate (iad1) |
|---|---|---|
| Active CPU | Time your code is actually computing | $0.128/hour |
| Provisioned Memory | Memory allocated to the function × time the instance is alive (waiting included) | $0.0106/GB-hr |
| Function invocations | Requests to your functions (success or failure) | $0.60 per million |
| Fast Origin Transfer | Transfer between the CDN and your functions or origin | $0.06/GB |
| Image transformations | Transforming images not yet in cache | $0.05 per 1,000 |
| Blob storage | Monthly average size of files stored in Vercel Blob | $0.023/GB |
Sources: iad1 pricing, Fluid compute pricing, Fair Use Guidelines. Rates vary by region. Functions run in a single region, iad1, by default; Active CPU is $0.128/hour there, versus $0.202 in Tokyo (hnd1) and $0.221 in São Paulo (gru1).
Function billing (Fluid compute) boils down to "CPU only while computing, memory for as long as requests are in flight." While the function waits for a database or an AI API, CPU billing pauses, but memory billing continues until the last in-flight request finishes. Once all requests complete, the instance pauses and nothing is billed until the next request. One instance can handle several requests at once.
Worked example: an AI chat API on Pro (iad1), 100,000 calls a month
Assumptions: 2 GB memory (the default), a 20-second wait per response, 0.1 s of computation. Conservative counting where each call occupies its own instance.
- Provisioned Memory: 2 GB × 20 s × 100,000 = about 1,111 GB-hrs → about $11.78
- Active CPU: 0.1 s × 100,000 = about 2.78 hours → about $0.36
- Invocations: 100,000 → $0.06
The total is about $12.20, which fits inside the $20 credit if other usage is small. Over 90% of the cost is memory, so if the wait doubles, the cost roughly doubles. The AI API fees themselves (what you pay OpenAI or Anthropic) are separate from Vercel.
5. Five typical ways the bill grows
When a Pro bill goes past expectations, it's usually one of the following. All of them come from Pro's nature of "keep running, don't stop."
What grows, and how to prevent it
Prevention: keep
maxDuration to what you need and lower memory to fit the job (Pro lets you set the default memory). Move work that takes more than a few minutes to Vercel Workflows, which has no duration limit.Prevention: put your own timeouts on external calls, and don't leave
maxDuration at the maximum.Prevention: limit the number of widths you serve. Images that don't need optimizing (icons, SVG) can have optimization turned off per image or per project.
Prevention: check the Usage page by project and by region, and narrow it down with WAF custom rules (up to 3 on Hobby, 40 on Pro).
Prevention: shorten the retention period for old deployments (30 days by default on Hobby and Pro) and delete Blob files you no longer use. Manage AI Gateway usage separately on the key side.
Sources: Fluid compute pricing, Configuring max duration, Image Optimization, DDoS Mitigation, Spend Management, Deployment Retention
The "about $94" in item 2 is 4 GB × 800 s = about 0.89 GB-hrs, times 10,000 calls, times $0.0106. When an AI writes your functions, it sometimes sets a generous maxDuration to be safe. Before deploying, read the maxDuration and memory values the AI set. The problem of an AI pushing every small fix and multiplying builds is covered in the pitfalls section of Deploying to Vercel from Claude Code and Cursor.
6. Setting a cap with Spend Management
Once you're on Pro, set up Spend Management first. Hobby doesn't have it, and on Pro it costs nothing extra. Only people with the Owner or Billing role on the team can configure it.
- In your team dashboard, open Settings in the sidebar and select Billing
- Turn on the Spend Management switch
- Enter an amount in US dollars under On-Demand Budget (only on-demand usage beyond the $20 credit counts toward it)
- Choose what happens when the amount is reached: notifications (web and email at 50%, 75% and 100%; SMS at 100%), a webhook, or pausing production deployments
- To actually stop, set Pause Production Deployments to Enabled, type your team name to confirm and select Continue (changes save automatically; this switch may already be on)
You can do the same from the Vercel CLI. vercel billing budgets inspect compares this cycle's on-demand spend with your amount, and vercel billing budgets set --amount 50 --pause (vercel billing budgets) sets "pause production on all projects when on-demand spend reaches $50." To see which project is using what, run vercel billing usage --since 30d --group-by project (Manage and optimize usage).
Four things to know about how it stops
- Setting an amount doesn't stop anything on its own. The docs say plainly that it does not stop usage; only turning on the pause does. The switch may already be on, so if you want only notifications and a webhook, check that it's off (with the CLI,
--no-pause). - Checks run every few minutes. Pausing can happen several minutes after you cross the amount, and usage keeps accruing meanwhile. The docs recommend setting the amount below your true maximum.
- When it stops, production for every project on the team returns a 503 (
DEPLOYMENT_PAUSED) to visitors. Previews aren't affected. - Raising the amount doesn't bring them back. You resume each project manually from the dashboard or the REST API.
New Pro teams start with a notification set at $200 per billing cycle. That only notifies; it doesn't stop anything. For a commercial site where "a bigger bill is better than going offline," you can turn the pause switch off and keep only notifications and a webhook. Decide by comparing what you lose if the site goes down with what you lose if the bill grows (source: Spend Management).
7. Four alternatives when it doesn't fit
If the six questions or the numbers in sections 3–5 show that Vercel doesn't fit, here's where to go, grouped by the condition that didn't fit. A side-by-side comparison of each provider's free tier and pricing is in What is PaaS (Vercel etc.)?.
| Condition that didn't fit | Where to go | One thing to check first (official) | How Next.js runs |
|---|---|---|---|
| Lots of images and static files; worried about transfer or request counts | Cloudflare (Workers) | Requests to static assets are free and unlimited, and the free tier for functions is 100,000 requests a day with 10 ms of CPU per invocation (Workers Pricing) | Cloudflare's own integration (verified adapter in progress) |
| You want a static site on a free tier that stops at its limit | Netlify | New accounts since September 4, 2025 use credits: Free is 300 credits a month, 1 GB of transfer = 20, one production deploy = 15 (Credit-based plans) | Netlify's own integration (verified adapter in progress) |
| You need an always-on server or a database | Render | A free web service spins down after 15 minutes without traffic and takes about a minute to come back, and a free Postgres database expires 30 days after creation (Free instances) | Runs as a Node.js server (all features) |
| You want a flat fee and full control, not usage billing | Your own VPS + Docker | The VPS provider's fixed monthly fee, with OS updates, monitoring and backups as your job | All features with Docker |
The Next.js column follows the categories in the official Next.js "Deploying" docs: Node.js server and Docker support all features, static export supports only some, and provider integrations "vary." So dropping Vercel doesn't mean dropping Next.js. What you take on instead is either looking after a server yourself or checking the gaps in each provider's integration.
Summary
"What should I do when AI says to use Vercel?" — the answer as of November 2026: "For a non-commercial Next.js site, start on Hobby. If money is involved, use Pro or another provider. Once on Pro, set the Spend Management amount and the pause before anything else." The AI's reasons (Vercel makes Next.js, the steps are short, it's free to start) are sound, but they describe the option with the fewest steps, not an answer that has considered commercial use or a cost ceiling.
You can estimate cost from three things: the number of CDN requests, how long functions wait × memory, and the number of image transformations. As a rough guide, the Hobby allowance covers about 33,000 page views for an image-heavy blog and about 32,000 AI calls that wait 20 seconds. On Pro, the $20 credit covers even 100,000 of those AI calls a month by this calculation, but the cost grows almost in proportion to the wait. If the terms don't fit, go to Cloudflare, Netlify, Render or your own VPS. Next.js runs with all features on a Node.js server or in Docker.
Related articles: When AI tells you to use Next.js (whether to use Next.js), What is PaaS (Vercel etc.)? (the kinds of hosting and each provider's free tier), Deploying to Vercel from Claude Code and Cursor (the actual steps), Vercel AI SDK complete guide (how to build an AI chat), Can you leave infrastructure to AI?.
FAQ
Q. Can I run my personal blog on Vercel Hobby (free)?
A. Yes, as long as you're not monetizing it. Ads such as Google AdSense, selling or promoting a product or service, or accepting payments make it commercial, which requires Pro. Affiliate links count as commercial if they're "the primary purpose of the site." The Fair Use Guidelines state explicitly that asking for donations is not commercial.
Q. If I go over the free allowance, will I suddenly get a huge bill?
A. Not on Hobby. Most features stay off until 30 days have passed (Web Analytics can resume after 7 days, and for image transformations only new images fail). On Pro, you can. Once the credit is used up, billing switches to on-demand and doesn't stop unless you turn on "Pause Production Deployments" in Spend Management.
Q. How much does it cost to run an AI chat on Vercel?
A. The Vercel part is driven mainly by "time spent waiting for the response × memory." With 2 GB of memory and a 20-second wait, 100,000 calls a month on Pro (iad1) come to about $12 (the calculation in section 4). On the Hobby allowance, about 32,000 calls is the rough limit. The AI API fees are charged separately.
Q. Is it hard to move off Vercel later?
A. If you only use standard Next.js features, it's straightforward. The official Next.js docs say a Node.js server or Docker supports all features. The extra work comes from parts that depend on Vercel-specific services such as Global Config or Vercel Blob.
Q. Is it safe to let an AI agent (Claude Code, Cursor) set up Vercel?
A. Leaving the initial deploy and basic setup to it is fine. But a person should check the plan choice (commercial or not), the Spend Management amount and pause, the functions' maxDuration and memory, and any secrets you put in environment variables. AI is good at "making it run," but "where to stop it" depends only on your budget and use. How much to leave to AI is covered in Can you leave infrastructure to AI?.
