Of all the questions in this series, this is the one that arrives with the most irritation attached, and with the best justification.
An organisation builds a customer-facing AI service. Somewhere in the design, someone commits to reserved capacity — a fixed monthly charge, often a substantial one, paid whether the service is busy or idle. Months later a leader looks at the usage report, sees a service handling a modest number of interactions a day, looks at the invoice, and asks the obvious question.
Why are we paying to reserve something we are barely using?
Usually there is a good answer. Sometimes there is not. Telling the difference is worth understanding, because this single decision moves more money than anything else in this series.
Two ways to pay for the same thing
Consumption AI can be bought in two shapes.
On demand. You pay for each unit of work as it happens, like a toll per trip. No commitment, no floor, no ceiling. A quiet month costs almost nothing. A busy month costs a lot.
Reserved, or provisioned, capacity. You rent a guaranteed slice of the platform's throughput for a fixed monthly fee — a lane on the road that is yours whether you drive on it or not. The lane has a size. Traffic up to that size is covered by the fee. The fee does not move when the traffic does.
The vendors give this various names — provisioned throughput, provisioned capacity, PTUs — and the details differ. The shape does not. You are converting a variable cost into a fixed one, and paying a premium for the privilege at low volume in exchange for a discount at high volume.
What the lane actually buys
Reserved capacity is not simply a bulk discount, and treating it as one leads to the wrong decision. It buys three things.
Predictable performance. On-demand capacity is shared. At busy times you queue behind everyone else on the platform, and responses slow down. A reserved lane is yours; the response time is consistent because nobody else is in it. For a service where a customer is waiting in a browser tab, that consistency has real value.
Protection from being throttled. On-demand services have rate limits, and a genuinely popular service can hit them. Reserved capacity raises that ceiling to a level you have chosen and paid for.
A predictable bill. For a service with steady, high volume, a flat monthly figure is easier to budget than a variable one, and at sufficient volume it is also simply cheaper per unit.
Every one of those is a real benefit. Every one of them is worthless if the traffic has not arrived yet.
The classic mistake
Here is what actually goes wrong, and it goes wrong in a recognisable pattern.
A new service is being designed. Somebody sensible asks what happens if it becomes very popular. Somebody else, wanting to be safe, sizes the reserved capacity for that success case. The commitment is signed before launch. And then the service launches to the volume that new services actually launch to, which is not very much, and often takes months to grow into anything.
For those months you are paying full price for a lane that is nearly empty. That money bought nothing at all — not performance you needed, not headroom anyone used, not readiness that could not have been arranged later. It bought the feeling of being prepared, which is an expensive thing to buy monthly.
Illustratively: a service might reserve capacity costing, say, the equivalent of a mid-five-figure Rand amount every month, while the actual traffic in the first quarter would have cost a small fraction of that on demand. Over four months that is a meaningful sum spent on nothing, and it is usually spent quietly, because the invoice is bundled and nobody is looking at utilisation.
The rule
Do not lease the lane before you have the trips.
Start on demand. Let the service launch, let real people find it, let the usage pattern reveal itself. Watch what it costs. Then, when the volume is genuine and sustained and you can see on a chart where it crosses over, commit to reserved capacity — and commit to the size the evidence supports, not the size the optimistic scenario suggests.
This is not caution for its own sake. Committing early is strictly worse in every dimension: it costs more, it locks in a size chosen when you knew least about the workload, and it can pin you to a particular model at a moment when the models are being replaced faster than the contracts renew.
The exceptions, honestly stated
There are legitimate reasons to reserve capacity earlier than the pure volume argument would suggest, and it would be dishonest to leave them out.
Availability of the option. In some regions — including, sometimes, in-country deployments in South Africa — the model you need is only offered on a provisioned basis. Then it is not a choice; it is the price of the constraint, and it belongs in the business case as such. The next article but one deals with this properly.
A known launch spike. If the service goes live alongside something with a hard date and a large audience — a filing deadline, a public campaign, a regulatory cutover — the traffic is not hypothetical and reserving for it is reasonable. The test is whether you can name the date and the expected volume, not whether you hope it will be busy.
A hard latency commitment. If you have committed to a response time in a service agreement, shared capacity may genuinely not be good enough from day one.
What all three have in common is that they are specific, checkable and written down. "In case it takes off" is not one of them.
What to ask before you sign
Four questions, and you do not need to be technical to ask any of them.
What does this cost on demand at our realistic volume? If nobody can answer that, the comparison has not been done and the reservation is a guess.
At what volume does the reservation become cheaper? There is a crossing point. It is a calculable number. Ask for it, and ask how far current traffic is from it.
What is the shortest commitment available? Terms vary from monthly to annual. Early on, shorter is worth a premium — you are buying the ability to be wrong cheaply.
Who reviews utilisation, and how often? A reservation with nobody watching it is a subscription nobody cancels. Monthly is right for the first year.
The point
Reserved capacity is not a trap and it is not a rip-off. At real, sustained volume it is the correct decision and it saves money.
It is simply a bet on volume, and the only way to lose the bet badly is to place it before you have any evidence about the volume. Start on demand, watch what happens, and commit when the traffic tells you to.
How CloudNala can help
We put the crossover point on a single chart before anything is signed: what on-demand costs at today's volume, what the reservation costs, and what monthly traffic has to look like before the commitment pays for itself. Then we set the review that checks it, because the expensive version of this mistake is not signing the wrong reservation — it is nobody looking at it again for a year.
Work with CloudNala
CloudNala helps organisations move from technology ambition to practical execution across cloud, AI, data, platform engineering and digital services.
Whether you are exploring AI, modernising your cloud environment, building a public-sector digital service, or turning an idea into a working MVP, we can help you shape the roadmap and deliver the next step.
Request a Cloud Review or write to us at consult@cloudnala.co.za