Voltage read NVIDIA's FAQ, its API Trial Terms and the forum replies. The key is free, the credits are gone, the rate limit isn't printed, and customers are off limits.
I'm Voltage, and I like free things with the fine print read out loud. The pitch for build.nvidia.com is simple: sign up, grab a key, call dozens of hosted models. Plenty of blog posts say the key comes with 1,000 credits and a 40-requests-a-minute cap. Some of that is stale.
So on 3 October 2026 I read NVIDIA's own pages: the NIM FAQ, the API quickstart, the API reference and the API Trial Terms. I also read NVIDIA staff replies on its developer forum. I haven't signed up, called the API or timed any model, and NVIDIA doesn't pay us. This is a reading of the rules. If you're deciding where a model should run, our guide to cloud versus local GPUs for AI covers the wider question.
What the free key is
The free key is API access to models NVIDIA hosts, for trying them out. NVIDIA's FAQ says Developer Program members "have free access to NIM API endpoints for prototyping". You get the key by clicking "Get API Key" on a model page and entering your email, which also joins you to the Developer Program.
NIM is NVIDIA's name for packaged model microservices. The hosted version answers at https://integrate.api.nvidia.com, and the LLM reference lists POST /v1/chat/completions there, the same shape many chat clients already speak. Our guide to running a model as an API server shows why that matters: a tool that talks to one such endpoint usually talks to another.
How many models? NVIDIA's public /v1/models list returned 80 IDs on 3 October 2026, 32 of them owned by NVIDIA, with Google, Meta, Mistral, IBM and Writer among the rest. That list includes models that aren't chat models. I can't tell you every ID on it works for free, because the pages I read don't say.
NVIDIA FAQ and public model list, read 3 October 2026. · aliteq research
Credits and rate limits: what is and isn't printed
The credits are gone and the rate limit isn't on any page I read. NVIDIA staff wrote on 10 September 2025: "We no longer use a credit-based system for build.nvidia.com. This has been replaced by rate limits for trial usage." So any post promising 1,000 credits is out of date.
The same staffer said "You can see your maximum rate limit in the top right of build.nvidia.com". She also said the trial "is not limited by a time period". Nothing in the quickstart, the FAQ or the model reference page I read states a number.
A number does appear in the forum. A reply on 8 June 2026, from a user whose role isn't shown, says "The 40 RPM limit is a global hard cap" and that NVIDIA can't raise it for free personal accounts. Another thread quotes a moderator saying the limit depends on the model, the use case and current traffic. So treat 40 requests a minute as a figure reported on a forum, not a published guarantee. Your own account page is the source that counts.
Rate limit and credits, by source
Credits
What it says
Gone; replaced by rate limits for trial usage
Where
NVIDIA staff, forum, 10 Sep 2025
Trial length
What it says
Not limited by a time period
Where
NVIDIA staff, forum, 17 Sep 2025
Your limit
What it says
Shown top right of build.nvidia.com
Where
NVIDIA staff, forum
A number
What it says
40 requests a minute, called a hard cap
Where
Forum reply, 8 Jun 2026, not on a docs page
Docs pages
What it says
No limit printed
Where
Quickstart, FAQ, model reference
What it says
Where
Credits
Gone; replaced by rate limits for trial usage
NVIDIA staff, forum, 10 Sep 2025
Trial length
Not limited by a time period
NVIDIA staff, forum, 17 Sep 2025
Your limit
Shown top right of build.nvidia.com
NVIDIA staff, forum
A number
40 requests a minute, called a hard cap
Forum reply, 8 Jun 2026, not on a docs page
Docs pages
No limit printed
Quickstart, FAQ, model reference
The no-production rule
The free key can't power something real people use. NVIDIA's Trial Terms say access is "for limited trial purposes only and without use of the API Service or Generated Content in production". They add that without a subscription you "may only use the API Service for internal testing and evaluation purposes, not in production."
The FAQ defines production as "any use of NIM for purposes other than development, testing, research or evaluation such as conducting business transactions and any non-testing activity". NVIDIA's route for that is an AI Enterprise licence, which "start[s] at $4500 per GPU per year or ~ $1 per GPU per hour in the cloud." The FAQ also lists a free 90-day AI Enterprise trial that includes a commercial-use licence.
Two more clauses matter for a side project. The terms call the service possibly "pre-release", with "reduced or different security, privacy, availability, and reliability standards", and say NVIDIA may end it "at any time without liability". Don't build something you can't rebuild.
NVIDIA NIM FAQ and API Trial Terms v. 19 Sep 2025, read 3 October 2026. · aliteq research
What happens to your prompts
Assume anything you send is seen by NVIDIA and may be used to improve its models. Section 3.3 of the terms says NVIDIA collects "User Content and Generated Content to improve NVIDIA products and services, including AI models". It also says use is logged for abuse monitoring and shared with third-party providers for that purpose.
The terms contradict themselves a little. Section 2.3 says NVIDIA won't store or use your content after a session "except" as stated or disclosed. NVIDIA staff said in June 2025 that content is used "solely to provide users with the API Service". I can't reconcile those, so I'd trust the stricter reading and send nothing you'd mind sharing.
Section 2.6 goes further: don't submit "confidential information, controlled or sensitive data", including health data, personal data and card data. For a work document, a customer list or a private repo, the free key is the wrong tool. A model on your own machine never sends a word out.
Free API, your own card or a rented GPU
Pick by who uses the result and how private it must be. A free key suits an afternoon of trying models. Your own card suits daily private work. A rented GPU suits a big model or a burst of heavy use. A paid, licensed route suits anything with customers.
Trying a model for an afternoon: use the free key with throwaway prompts and stay under your account's rate limit.
Using one daily with private data: run it on your own card. Our [cost-to-run tool](/cost-to-run) checks whether a model fits your VRAM.
Need a big model or a burst: rent a GPU by the hour. Compare live prices on our [cloud GPU page](/cloud-gpu).
Serving real customers: NVIDIA says production needs AI Enterprise; other providers sell their own paid plans.
Self-hosting NVIDIA's own containers is a middle path. The Developer Program lets members download NIM microservices and run them on up to 16 GPUs for research and development, on any infrastructure. It's the same production fence, just with your hardware doing the work.
Yes, for prototyping. NVIDIA's FAQ says Developer Program members get free access to NIM API endpoints for prototyping, and you join by getting a key. It's a trial under NVIDIA's API Trial Terms, not a free production tier.
What is the rate limit on the free NVIDIA API?
NVIDIA's docs don't print one. Staff say your maximum shows in the top right of build.nvidia.com. A forum reply from June 2026 says 40 requests per minute is a hard cap, but that is not a docs-page figure, so check your account.
Does the free NVIDIA key still come with credits?
No. An NVIDIA staff member wrote in September 2025 that the credit system was replaced by rate limits for trial usage. Posts that promise 1,000 or 5,000 credits describe the old system.
Can I use NVIDIA's free hosted models in a product?
No. The Trial Terms limit use to internal testing and evaluation, not production. NVIDIA says production use of NIM requires an NVIDIA AI Enterprise licence, listed from $4,500 per GPU per year.
Does NVIDIA keep or train on what I send?
The terms say NVIDIA collects submitted and generated content to improve its products, including AI models, and logs use for abuse monitoring. Another clause says content isn't stored after a session unless disclosed. Send nothing sensitive.
Is the NVIDIA endpoint compatible with OpenAI-style clients?
The LLM reference lists a chat completions endpoint at integrate.api.nvidia.com, path /v1/chat/completions. I didn't test it with any client, so check each tool's docs.
