Hidden costs of AI capacity that aren’t on the quote

The headline price for AI capacity is the number everyone negotiates. It is often not the number that decides what the project costs. The rest hides in the gaps around it: moving data, waiting for capacity, and paying for machines that sit idle.

None of these costs show up on the first quote. All of them show up on the invoice, or in the project plan, a few months later. Here are the ones to look for before you sign. For dated list prices to compare against, see Model Watch. Each section ends with a question you can put to any provider.

1. Moving data in and out

AI work needs data. Documents to read, images to look at, logs to summarise. That data has to travel from where it lives to where the model runs, and the answers have to travel back.

Some providers charge for data leaving their network. Others charge for storage, for logs or for keeping copies of requests. Each line looks small. Multiplied by every request, every day, it can become a real share of the bill.

There is also a time cost. If your data sits in one place and the AI runs somewhere far away, every request waits for the round trip. Teams then spend engineering time on caching and copying data around to make things feel fast.

Ask: what do you charge for data going in, coming out and being stored, and can the work run near where my data already is?

2. Waiting for capacity

The second hidden cost is time spent not running.

When demand for AI capacity spikes, the capacity you want may not be there. Requests get queued or rejected. Reserved machines have waiting lists. A launch that was planned for one month slips to the next because the hardware is not available.

The cost of waiting does not appear on any invoice. It appears as a delayed product, a pilot that runs out of budget before it shows results, or a team that keeps paying salaries while it waits.

Waiting also pushes buyers into bad deals. A team that has been queued once will often reserve far more than it needs next time, just to be safe. That leads to the third cost.

Ask: how quickly can I get the capacity I need, what happens at peak times, and is there a guarantee in writing?

3. Paying for idle machines

Reserved capacity is cheaper per unit. It is also paid for whether you use it or not.

Many AI workloads are uneven. A customer support assistant is busy in working hours and quiet at night. A document pipeline runs hard at the end of the month and idles in between. If you reserve enough capacity for the busiest hour, you pay for that peak around the clock.

Idle capacity is easy to miss because the unit price looks good. The useful number is what you pay per piece of work actually done. If machines are busy a third of the time, each request is carrying the cost of the other two thirds.

Ask: can I commit to a baseline and pay per use above it, and can unused reserved time be moved or resold?

4. Lock-in

Some costs only arrive when you try to leave.

Moving to a different provider can mean rewriting code that depends on one company's tools, re-testing models, moving stored data and paying exit fees. If the cost of leaving is high, the provider knows your next renewal is not much of a negotiation.

Lock-in is not always avoidable, and some of it is the price of convenience. It should be a choice, made with the cost in view.

Ask: what would it take to move this workload elsewhere in six months, and what do you charge if I do?

5. The cost of not knowing

The last hidden cost is uncertainty. If you cannot see where your work ran, how much of it ran and when, you cannot check the bill against the work. Finance teams then spend time reconciling invoices, compliance teams ask questions nobody can answer, and budgets get padded to cover surprises.

A clear record of what was delivered does not lower the headline price. It lowers the cost of trusting it.

Ask: what record do I get of each piece of work: where it ran, when and how much?

A short checklist

Before you sign for AI capacity, put these lines next to the headline price:

• Data transfer and storage charges for your expected volume
• Time to get capacity, and what happens at peak
• Expected utilisation of anything you reserve
• Exit terms and the work needed to move
• The record you will get of what was delivered

Add them up. A low headline price is not always a low project cost.

What we are building

Several of these costs come from distance and from capacity that is concentrated in a few places. We're building standardized compute close to power, connectivity and demand, connected through one programmable network for AI inference.

The network is designed so work can be placed by policy, on latency, data residency, model and load, which lets it run nearer to where data and users are. Every workload can generate an auditable record of commitment, placement and delivery, so the bill can be checked against the work.

Related reading: AI inference cost: why one task gets different prices and Cost per million tokens: how to read an AI price list.

If hidden costs are shaping your AI plans, talk to us: use Request capacity below.