AI ROI can be calculated like any other IT project: how much working time you save, what the infrastructure costs and how quickly the deployment pays for itself. The trouble is that most calculations ignore the time spent checking AI output, or compare nothing but token prices. Below is a simple method with a worked example.
This article is for executives and managers who need to justify a budget. The example uses round, made-up numbers. They are not data from any client, only an illustration of the method: plug in your own values and see whether the result holds up.
The three costs you need to compare
Every AI deployment comes down to comparing three items:
- Labour cost: the time people spend on the task today, multiplied by the fully loaded hourly cost.
- API cost: the number of tokens multiplied by the provider’s list price. It grows linearly with usage.
- Own GPU cost: a fixed fee for the hardware (or purchase plus colocation) plus operations. It does not depend on the number of requests until you reach the machine’s capacity.
On top of that comes a one-off implementation cost: integration with your systems, data preparation, testing and user training. You pay it whether you choose an API or your own server.
How to calculate AI ROI in six steps
- Pick one process, not “AI for the whole company”. For example summarising documents, answering routine customer questions or searching internal procedures.
- Measure the baseline: how many people, how many hours a day, how many cases a month. Use your ticketing or time-tracking data rather than guesses.
- Estimate the realistic saving after subtracting the time needed to check and correct the output. AI rarely replaces the work entirely; it usually shortens it.
- Work out the volume: how many requests a day and how many tokens per request. If tokens are new to you, see tokens in large language models.
- Compare API and own-GPU costs at that volume, including operations.
- Calculate the payback period: implementation cost divided by the monthly net benefit.
Worked example: a team of 20 (illustrative numbers)
Say 20 people each spend 1 hour a day working with documents: summaries, email replies, looking things up in procedures. Every number below is an assumption for the example. Instead of plugging in a GPU price, we calculate the maximum GPU budget at which the deployment still makes sense. You can then compare it with a quote or with our ready-made GPU packages.
| Item | Assumption | Per month |
|---|---|---|
| Time spent on documents | 20 people × 1 h × 21 working days | 420 h |
| Saving after review time | 30% (assumption) | 126 h |
| Value of time saved | 126 h × €35 (assumed fully loaded hourly cost) | €4,410 |
| Operations and administration | assumption for the example, not a price list | €400 |
| GPU cost | fixed monthly fee, call it G | G |
| Net benefit | €4,410 − €400 − G | €4,010 − G |
If the one-off implementation costs, say, €8,000 (another assumption) and you want it to pay back within six months, the net benefit must be at least €8,000 / 6 ≈ €1,333 a month. So the GPU can cost up to about €4,010 − €1,333 = €2,677 a month. If the quote is below that, the business case holds.
Now the sensitivity test. If the real saving is 15% rather than 30%, the value of time saved drops to €2,205. After operations, that leaves €1,805 a month for the GPU just to break even, with nothing left to pay back the implementation. This is why step 3 matters most: measure the saving in a pilot before you sign off an annual budget. To size the hardware itself, see how much VRAM does an LLM need.
API or own GPU: where the break-even point lies
Stay with the same example. Assume each of the 20 people sends 30 requests a day, and each request is about 4,000 tokens in total (input and response). That is 2.4 million tokens a day, or roughly 50 million a month.
To find the break-even point, divide the fixed GPU cost by the volume. At 50.4 million tokens a month, every €100 of monthly GPU cost works out to about €2 per million tokens. If an API for a comparable model costs less per million tokens (blending input and output), and the data is not sensitive, the API is simply cheaper at this volume. It is worth saying that plainly.
Your own GPU wins when:
- Volume grows while the GPU cost stays flat. RAG attaches document excerpts to every request, agents make many calls per task, and batch-processing a document archive generates traffic with no human involved. Token counts can then multiply.
- One server handles several processes, for example a knowledge assistant, transcription and ticket classification. The fixed cost is shared across all of them.
- Data must not leave your infrastructure: medical records, financial data, source code. Then the API is out regardless of price, and you are comparing your own GPU only with manual work.
- You need a predictable budget. A fixed fee is easier to plan than a bill that depends on how much staff use the tool.
We compare both models in more depth in DeepSeek R1 vs ChatGPT API. Larger GPU servers are quoted individually, and renting versus buying is covered in GPU as a Service.
Common mistakes when calculating AI ROI
- Counting a 100% time saving. Someone always has to check the output, especially at the start.
- Ignoring adoption. Not everyone will use the tool from day one. Assume lower usage in the first months.
- Forgetting operations. Model updates, monitoring and backups are work that costs money, whether yours or a provider’s.
- Hardware that is too small or too big. Too small means slow responses that put users off; too big sits idle. Start with a pilot and scale based on measurements.
- Comparing token prices only. Without regulatory requirements and a one-year volume forecast, an API vs GPU comparison is incomplete.
For the full picture, from model choice to security, see our self-hosted LLM guide for business.
FAQ: frequently asked questions
How do you calculate the ROI of AI?
Calculate the monthly value of the time saved (after subtracting review time), deduct infrastructure and operations costs, then divide the one-off implementation cost by that net benefit. The result is the payback period in months.
Is it cheaper to use an API or run your own GPU server?
It depends on volume. Divide the monthly GPU cost by the number of tokens you process per month and compare it with the API price per million tokens. At low volume the API often wins; at high and growing volume your own GPU usually does.
How much does a GPU server for AI cost per month?
Our ready-made GPU packages have fixed monthly prices that depend on the card; see the packages page. Blackwell GPU servers and DGX Spark as a Service are quoted individually.
How long does it take for an AI deployment to pay for itself?
There is no single answer. The payback period depends on how much time you really save, your fixed costs and what the implementation cost. That is why it pays to start with a pilot and measure the saving on your own data.
Want to work out AI ROI for your own process? Get in touch: we will help estimate the volume, match hardware from our ready-made packages or a custom configuration, and set up a pilot.