AI inference cost: why one task gets different prices

Ask three providers to run the same AI task and you can get three different prices. Same model, same prompt, same answer. So what are you paying for?

A quoted price for AI work looks like one number. Underneath, it is a stack of decisions about power, place, timing and risk. Once you can see the stack, the differences stop looking random, and you can tell which ones matter for your work.

This piece walks through the four layers behind most quotes.

Layer 1: Power

Every AI request runs on chips, and chips run on electricity. Power is a major running cost of the machines, and its price is not the same everywhere.

Electricity costs depend on the local grid, the time of day, the contract the operator signed and how the power is generated. A site that buys power cheaply can quote lower. A site in a tight market has to pass some of that cost on.

Power also shapes what a site can do at all. A location that is waiting for a bigger grid connection cannot simply add more machines, however much demand there is. Scarcity shows up in the price.

What to ask: where does the capacity draw its power, and is the power cost fixed or does it move with the market?

Layer 2: Location

Where the work runs changes the price in a few ways.

Land, buildings, cooling and local staff cost different amounts in different places. Network links matter too. A site close to large population centres and good connectivity is valuable because it can serve requests quickly, and the operator knows it.

Then there are rules. If your data has to be processed in a named country, the set of sites you can use gets smaller. A smaller set usually means less competition on price.

None of this is good or bad on its own. A higher price for a nearby site may be the right trade if your users need fast answers or your data must stay put. The point is to know which part of the price is buying something you need.

What to ask: which location am I paying for, and do I need that location for this workload?

Layer 3: How full the machines are

This is the layer buyers see least, and it explains a lot.

An AI server costs roughly the same to own whether it is busy or idle. If a provider keeps its machines busy most of the day, the cost of each request is spread thin. If machines sit idle for long stretches, every paid request has to carry the cost of the quiet hours.

That is why the same provider may price differently by time of day, by region or by product. It is also why some providers offer lower prices for work that can wait, such as overnight batch jobs. Flexible work helps them fill the gaps.

For buyers, this has a practical consequence. If your work can tolerate a delay, you may be paying for speed you do not use. If it cannot, you are paying for the provider to keep spare capacity ready for you, and that has a cost.

What to ask: is there a different price for work that can wait, and what am I paying to have capacity ready the moment I need it?

Layer 4: How long you commit

The last layer is time.

Paying per request is flexible. You pay only for what you use, and you can walk away tomorrow. The provider takes the risk that demand drops, and prices that risk in.

Reserving capacity for a month or a year flips the risk. You commit to paying whether or not you use it, and in return the unit price is usually lower. The provider can plan its own costs around your commitment.

Neither is better in general. A team testing a new feature should rarely lock in a year. A team running a steady, predictable workload may be overpaying by staying on per-use prices.

What to ask: what would this cost on a one-month, six-month and twelve-month commitment, and what are the exit terms?

Putting the layers together

When two quotes differ, run them through the four layers:

1. Power: is one site simply buying cheaper electricity?
2. Location: is one site closer to your users or inside the country your data needs?
3. Utilisation: is one price for work that can wait, and the other for work that runs instantly?
4. Commitment: is one a per-use price and the other a reserved rate?

Often the "expensive" option is buying something specific, like a location or instant availability. Sometimes it is buying nothing you need. The layers help you tell the difference.

One more thing to check: whether the quote is dated. Prices for AI work move quickly. On Model Watch we publish list prices with the date we read them. In Model Watch #1, four paid models listed $2.00 per million input tokens, with outputs from $6 to $10, as of 22 October. A dated number is something you can compare against next month's. An undated one is a guess.

Why we care about this

Starfield is built around the idea that AI capacity should be bought the way other infrastructure is bought: knowing what you are paying for and where it runs.

We're building standardized compute close to power, connectivity and demand, connected through one programmable network for AI inference. Workloads are designed to be placed by policy, on latency, data residency, model and load, so the location and timing you choose are part of the order rather than hidden inside the price. It is designed so every workload can generate an auditable record of commitment, placement and delivery.

Related reading: Cost per million tokens: how to read an AI price list and Hidden costs of AI capacity that aren't on the quote.

If you are comparing quotes for inference and want to talk through what you are paying for, use Request capacity below.