NVIDIA NIM Pricing in 2026: Free Tier and License Cost
Nov 7, 202614 min read
You pay nothing to try NVIDIA NIM and a real budget line to run it, whether you wire it into a workflow tool like Activepieces or call it directly.
The hosted API is a prototyping allowance, and in production you pay either an AI Enterprise licence per GPU or a partner's per-token rate. Which one is cheaper depends on how many tokens you push through each month, not on the sticker price.
NVIDIA NIM pricing today, plan by plan
What NVIDIA NIM is and what you pay for
NVIDIA NIM is a set of prebuilt inference containers. Each holds a model, an optimized runtime and a standard API, packaged so you can run it on NVIDIA GPUs without assembling the serving stack yourself. You can reach it three ways.
NVIDIA hosts a free API for experimenting. You can license the containers under NVIDIA AI Enterprise and run them on your own or rented GPUs.
Or you can skip the containers and buy the same models by the token from a partner that runs them for you.
You carry the GPU bill and the on-call rota, and the licence buys only a supported, patched runtime.
Hand it to a per-token partner and both move off your plate, but you pay for that in a rate that scales with every request.
The plan table, checked 7 November 2026
The table below sets each access route against its price, inclusions and usage limit as of 7 November 2026.
| Route | Price | Inclusions | Usage limit | Checked |
|---|---|---|---|---|
| Hosted free API | Free | Hosted endpoints for trying models | Prototyping allowance, not for production | 7 Nov 2026 |
| AI Enterprise, 1-year | $4,500 per GPU | Licence per GPU, NIM containers, enterprise support | Bound to licensed GPUs | 7 Nov 2026 |
| AI Enterprise, 3-year | Contact sales | Same as 1-year, longer commitment | Bound to licensed GPUs | 7 Nov 2026 |
| AI Enterprise, 5-year | $18,000 per GPU | Same as 1-year, longest commitment | Bound to licensed GPUs | 7 Nov 2026 |
| AI Enterprise, perpetual | Contact sales | Perpetual licence per GPU; support terms vary | Bound to licensed GPUs | 7 Nov 2026 |
| Cloud marketplace pay-as-you-go | $1 per hour per GPU + CSP instance costs | Licence billed through the cloud provider | Metered by usage | 7 Nov 2026 |
| Partner per-token | Set by each partner | Hosted inference, no GPUs to manage | Per-partner rate limits | 7 Nov 2026 |
What NVIDIA Doesn't Publish
In practice you can't compare a 1-, 3- or 5-year commitment against a perpetual licence until you have a quote.
Each partner publishes its own rates, not NVIDIA, so compare those on the partner's own pricing page.
If you are running this arithmetic for your own team, see what the same workload costs on Activepieces.
Free NIM access and model calls
Before you compare any partner's rate, test the model for free: sign in at build.nvidia.com, join the NVIDIA Developer Program, generate an nvapi- key, and make one curl call.
- Go to build.nvidia.com and sign in, creating an NVIDIA account if you don't have one.
- Join the NVIDIA Developer Program when prompted. The hosted API requires it.
- Pick a model from the catalog. Choose a small one first.
- Open the model's page and click Get API Key.
- Copy the key, which begins with
nvapi-, into an environment variable or a secrets manager. Don't paste it into code. - Run one curl request against the OpenAI-compatible endpoint, shown next.
Making your first NVIDIA NIM API request
Because the endpoint speaks the OpenAI chat-completions format, any client that lets you change the base URL can point at NVIDIA later without a rewrite.
curl https://integrate.api.nvidia.com/v1/chat/completions \
-H "Authorization: Bearer $NVIDIA_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "meta/llama-3.1-8b-instruct",
"messages": [{"role": "user", "content": "Say hello in five words."}],
"max_tokens": 32
}'If a JSON reply with a choices array comes back, the key, the account and the model all work.
If you get an authorization error, regenerate the key and check that the variable is actually set in your shell.
Checking your NVIDIA NIM rate limits
NVIDIA meters hosted access against your account and rate-limits it.
The exact limits change, so read the current figures in NVIDIA's documentation and on your account page at build.nvidia.com rather than trusting a number in an article.
Monthly NIM billing at real usage
Price the same 20,000 tickets three ways. The hosted free tier costs nothing but is capped, self-hosting has a fixed monthly floor, and a partner's per-token rate scales with every ticket.
The workload and its token counts
Picture a ticket summariser: each of 20,000 monthly tickets goes in as a thread and comes out as a short summary. I'm not going to guess your token counts, because input length drives the whole bill.
Pull the average input and output tokens per ticket from a week of your own logs.
Cost on each route, line by line
The table shows how each route's bill is built.
| Route | Inputs | Arithmetic | Monthly cost |
|---|---|---|---|
| Hosted free tier | 20,000 tickets | No charge, but rate-limited prototyping access | $0, not a production option |
| Self-hosted AI Enterprise plus rented H100 | $4,500 per GPU per year, one GPU, your H100 hourly quote | $4,500 ÷ 12, plus hourly rate × hours the GPU runs | $375 plus GPU rental |
| Together AI at $1.04 per million tokens | 20,000 × (input + output tokens per ticket) | Total tokens ÷ 1,000,000 × $1.04 | Scales with tokens, nothing when idle |
Whether you summarise 20,000 tickets or none, the $375 licence share is owed, so it's a floor, not a usage charge.
Where the break-even sits
To find break-even, take the monthly token volume at which the per-token bill equals the self-hosted floor: (375 + your GPU rental) ÷ 1.04, in millions of tokens.
Below that volume, Together AI is cheaper; above it, each extra token on your own GPU is effectively free until the card saturates.
If you sit well under it, take the partner rate and skip the operations burden of running a GPU.
If you're near or over it, self-hosting starts to pay, but weigh the maintenance of patching, monitoring and capacity against the saving.
Activepieces can produce the measured token count that break-even needs: one step calls the NIM endpoint over HTTP and the next logs the usage from each reply.
Its pricing meters the run, not the steps inside it, at 1 credit per run, so a two-step flow and a ten-step flow cost the same.
Zapier bills per task and Make per module, and on their pricing pages every extra call, log or check step adds to the bill; here you can split the work as finely as you like without the measuring costing more than the thing measured.
Worth checking against a plan that does not meter every step: one credit covers a whole run on Activepieces.
Hidden NIM costs and production limits
None of these seven line items is in the token estimate above, and every one of them appears only after NIM leaves the free hosted API:
Hosted API licensed for development and testing only: (NVIDIA API catalogue terms; checked 2026-11-07).
A licence for self-hosting (NVIDIA AI Enterprise pricing guide; checked 2026-11-07).
GPU rental, whichever cloud you choose.
Storage and egress for model weights.
GPUs that sit idle between bursts.
Operations time to patch, scale and monitor.
Three of these come from NVIDIA's terms. The other four come from your own deployment.
Free tier is for prototyping, not production
Ship a customer feature on the hosted API and you breach the terms, because it is licensed for development and testing.
It's also rate-capped, so one busy minute can queue your users behind each other.
Ship a customer feature on the hosted API and you breach the terms, because it is licensed for development and testing.
Most teams never get past this stage. MIT's GenAI Divide study found 60% of enterprises evaluated AI tools, 20% reached a pilot, and 5% reached production.
Licence plus GPU: the two bills
Self-host and you pay NVIDIA and a GPU landlord separately. NVIDIA AI Enterprise costs $4,500 per GPU for one year, $9,000 for two, $13,500 for three and $18,000 for four.
At five years the price is still $18,000, so the fifth year's free.
A buyer who can commit for five years pays $3,600 per GPU-year against $4,500 on a one-year term, which means multi-year terms save approximately 20% per GPU-year. None of that includes the GPU itself, which you rent or buy on top.
Somewhere has to hold the model weights, and every replica pulls them. Pulling across regions or out of a cloud is billed as egress, so a multi-region rollout pays for the same download repeatedly.
At 3 a.m. a licensed GPU costs the same as at peak. If traffic is bursty, you pay for unused hours, which is where a partner's per-token rate usually wins.
NVIDIA NIM compared with the closest alternatives
Of these four, NVIDIA NIM's the only one you can run on your own GPUs, and once licensed it costs about what a mid-priced hosted API does per token.
Comparison table on five dimensions
Three of the four routes bill by the token and only NIM bills by the GPU, as the table shows.
| Route | Free access | Billing unit | Self-hosting |
|---|---|---|---|
| NIM | Hosted API prototyping allowance | Licence per GPU (AI Enterprise) | Yes, in your own containers |
| Together AI | Not a free tier I can confirm | Per token | No |
| Fireworks AI | Not a free tier I can confirm | Per token | No |
| Amazon Bedrock | Not a free tier I can confirm | Per token | No |
For the per-token numbers, the source is a Together AI pricing comparison of a 70B model, hosted or self-run. NIM with its licence works out to $0.69 per million tokens, so that's the figure a self-hoster has to beat.
At $0.90 per million tokens, Fireworks sits about 30% above NIM, which is what you pay for not owning any hardware.
Bare vLLM, an open-source serving engine run on your own GPUs, comes in at $0.61, so the licence adds roughly $0.08 per million tokens, or 13%.
That premium buys NVIDIA's tuned, supported containers, and it's worth paying only if you would otherwise spend engineer time maintaining a serving stack.
Cheapest of all is DeepInfra, a third-party inference host, at $0.32, less than half of NIM's rate.
When NIM wins
When your volume keeps GPUs busy around the clock, or when data can't leave your network, NIM wins.
Past that point a flat per-GPU licence spreads thin, and $0.69 per million tokens holds only if the card is actually saturated.
When a per-token API wins
When traffic is spiky, small or still unproven, a per-token API wins. You pay nothing on quiet days, and there's no licence to amortise.
At low volume even Fireworks' $0.90 costs less in total than an idle licensed GPU.
What Activepieces does about this
The decision in this article rests on one number you probably don't have yet: your monthly token count. Activepieces is where we'd have you get it.
Build a flow that calls the NIM endpoint over HTTP, then add a step that writes the usage block from each reply to a sheet or database.
After a week you have measured input and output tokens per ticket, and the break-even formula above stops being a guess.
The cost of that measuring is flat. We charge 1 credit per run however many steps it has, so logging, checking and routing steps add nothing to the bill. Under per-task or per-module pricing, the same instrumentation would charge you for watching the workload.
Activepieces is also open source under the MIT licence for the community edition, so you can run it next to your own NIM containers if the data can't leave your network. That is the case where this article says self-hosting wins.
Nothing in the flow is tied to NVIDIA, so if the numbers favour a per-token partner, you change the URL and the key and keep the logging.
Frequently asked questions
Is NVIDIA NIM free for commercial use?
No: the hosted API on build.nvidia.com's a prototyping allowance, and running NIM containers in production requires an NVIDIA AI Enterprise licence per GPU. The alternative is paying a partner host a per-token rate.
A free trial that quietly becomes a product dependency is the expensive mistake here, so decide which paid route you are on before launch, not after.
Do I need a GPU to use NVIDIA NIM?
Only if you self-host. Calling the hosted API or a partner's endpoint needs nothing but an HTTP client, so your laptop is fine.
Running the container yourself needs an NVIDIA GPU that the specific model supports, and each model's documentation lists its supported hardware. Smaller models may fit on a workstation card, but check that list before you buy hardware. Don't assume a gaming card's covered.
How many free requests does build.nvidia.com allow?
I won't quote a number for the free tier's per-minute cap, because NVIDIA can change it and your account page's the authority.
What matters is the behaviour: once you pass the cap, requests are rejected or throttled rather than billed. That's fine for a demo and a poor foundation for anything with real users.
Treat the limit as a signal that you have outgrown prototyping, not as a ceiling to engineer around with retries.
Is NIM cheaper than the OpenAI API?
It depends on volume, and that's the whole decision. A per-token rate from a partner scales with usage, so it wins while your traffic is light or uneven.
The AI Enterprise licence is a fixed cost per GPU, so it only wins once you keep that card busy enough to spread the fee across many tokens.
Add the engineering time to run and patch the containers, and the break-even sits higher than a simple price comparison suggests. Below saturation, pay per token. Run the arithmetic with your own monthly token count before committing to a licence.
Can I automate NIM calls without writing code?
Yes. NIM endpoints speak an HTTP API, so any tool that can send a web request can call one.
One optional way to do it is Activepieces, an open-source workflow automation tool: you configure an HTTP step against the endpoint and add a second step that logs each call's usage.
That log gives you a running token count, which is the number you need to decide whether a licence ever pays off.
Related reading
Running the numbers
See what the same workload costs here.
Free forever plan, and every paid plan self-hosts at no extra cost.
See pricing Talk to sales
