Forecast next month's AI consumption
A forecast projects your recent usage-cost pace across the next calendar month. It does not predict a prepaid top-up, a card charge, or how many days remain before a balance reaches zero.
A provider dashboard is a record of the past. By the time the month-end total surprises you, the money is gone. The useful question is where consumption is heading, and whether there's still time to do something about it.
That's what a forecast is for. Take the pace you've been spending at, project it forward, and you get a number for next month. This guide covers what an AI spend forecast is, why provider consoles can't give you one that covers your whole stack, and the exact way CostCompass builds it.
What is an AI spend forecast?
A forecast is a forward number: what you’re on track to spend, as opposed to what you’ve already spent. Three things tend to get blurred together. Usage is the raw meter: tokens, GPU-seconds, characters, requests. A running total prices that usage into money and adds it up to the moment, giving you what the month has cost so far. A forecast takes the rate underneath that total and projects it ahead to estimate what a full month at this pace will cost.
For a prepaid provider, that forward number still describes consumption. It does not predict when auto-reload will run, when a card will be charged, or how large the next top-up will be. It also is not Runway: days until a balance is empty requires a current provider-reported balance, which is a separate input.
Of the three, only the forecast tells you whether you’re about to blow past the figure you had in mind.
Why can’t provider dashboards forecast for you?
Because a dashboard reports what already happened. Nearly every provider gives you a usage page, and for looking backward it’s usually fine: it shows what you’ve spent, and most will export it. A few of the big cloud consoles (AWS Cost Explorer, Google Cloud Billing, Azure Cost Management) go further and project their own usage cost forward.
Two limits make this worse the moment you run on more than one provider. Each console sees only its own slice, in its own unit (tokens on one, GPU-seconds on another), so there’s no single pace to project from in the first place. And even where a console does offer a projection, it covers that provider alone. To build a combined forecast yourself you’d log into every provider, read each one’s recent usage, convert it to money, average a trailing window, and project that across next month, then repeat the whole exercise every time you wanted a current figure. It’s enough work that it rarely becomes a habit, which is why a rising cost so often arrives as a surprise. For the broader picture of pulling those scattered consoles into one running total in the first place, see how to track AI costs across providers.
How does CostCompass forecast your spend?
The projection is built like this:
- Take a trailing seven-day burn rate. CostCompass sums your metered usage over the last seven calendar days (UTC), today included, and divides by seven to get a daily rate. Quiet days count too — that’s part of the smoothing. Seven days is long enough to absorb the swing between a heavy day and a quiet one, and short enough that a real change in pace shows up within days instead of being averaged away over weeks.
- Multiply it across next month. That daily usage rate is extended over the number of days in next month — 31 for a long month, 28 or 29 for February — to project what a full month at this pace would cost.
- Add next month’s subscriptions. Flat subscriptions are a known monthly charge, so they’re counted at the monthly amount you entered rather than extrapolated from a few days of usage. One that starts or ends mid-month contributes only the days it’s active.
The forecast carries no month-to-date term.
Underneath, the forecast always projects next month from money you’ve already spent. For most providers, that past spending is the seven-day pace. A provider that bills by invoice, such as OVHcloud, has a better record to work from: last month’s invoice, or this month’s if it’s your first.
Reading the number on the dashboard
The burn rate shown on the card is the everyday pace: your recent spend per day, subscriptions included. The forecast aims the same idea a month ahead, but it doesn’t reuse that combined rate, which would count subscriptions twice. Instead it extends only your metered usage from the seven-day rate, then adds subscriptions on top at their monthly amount.
Here’s the same calculation worked through with round numbers, for a next month of 30 days:
| Step | Value |
|---|---|
| Metered spend, last 7 days | $1,050 |
| Daily burn rate ($1,050 ÷ 7) | $150/day |
| Metered projection ($150 × 30) | $4,500 |
| Flat subscription (active all month) | $200 |
| Forecast for next month | $4,700 |
When the number updates
You pull the data by clicking Refresh; there is no background timer. A refresh asks every connected provider for its most recent reported usage, and the seven-day rate and the projection are rebuilt from what comes back. (Some providers report with a short lag, so “most recent” means as current as their own books are.) Because the trailing window always ends today, a developing spike works its way into the forecast on your next refresh, while there are still days left to act on it.
What the forecast can’t see
The forecast predicts next month from your last seven days. A launch, a seasonal swing, or a workload change you already know is coming won’t be in the number until it starts showing up in usage. A provider that reports usage with a delay feeds the window late, which can understate a very recent spike. And on a freshly connected provider with only a few days of history, the rate is built from a short sample, so early forecasts move around more than settled ones. The number is still what it claims to be: a reading of your current pace. If the forecast is climbing and you want to pull it back down, the guide to managing LLM costs covers the levers that move the rate.
What does the forecast cover?
It doesn’t stop at model APIs. The same projection rolls Claude and OpenAI up next to the GPU box, the hosting bill, and the voice service — the compute spend most token-only tools leave out. Your keys stay encrypted in your browser before they’re ever stored, so no usable credential ever sits in the database or logs.
For how each provider meters the usage the forecast is built from, the per-provider guides go deeper: Claude, OpenAI, and Google, with the full set on the providers page and a wider overview of tracking AI costs across providers.
How do you set up an AI spend forecast?
- Connect a provider. For most that means pasting the usage or admin key it gives you; a couple, like Google and GitHub, can sign in with OAuth instead. Either way it’s encrypted in your browser before it’s stored.
- Click Refresh to pull recent usage; the seven-day burn rate and next-month forecast appear beside the month-to-date total.
- Add the rest. Each provider folds into the same forward number, so one Refresh rebuilds the whole projection.
Frequently asked questions
- What is an AI spend forecast?
- It's a projection of what next month's consumption will cost if your recent usage pace continues. CostCompass extends seven days of metered usage across the next calendar month and adds expected subscriptions.
- Does the forecast cover the rest of this month or next month?
- Next month. CostCompass projects the full next calendar month at your current pace, not the remainder of the month you're in. Late in a month there are too few days left for a current-month projection to mean much, and the question that drives a budget is "if this pace holds, what does a whole month at it cost?"
- How is the forecast calculated?
- It takes your spending over the last seven days, divides by seven to get a daily rate, multiplies that by the number of days in next month, and adds next month's subscription cost on top. Flat subscriptions are counted at the monthly amount you entered; one that starts or ends mid-month is prorated to the days it's active.
- Why a seven-day window instead of the whole month?
- Seven days is a middle ground. A single day swings too much to trust. One heavy batch job or one quiet weekend would throw the projection off. Averaging the whole month so far reacts too slowly, dragging an old, cheaper pace into the number long after your usage has changed. A trailing week smooths out day-to-day noise while still turning quickly when your real pace shifts, so a genuine change in spending reaches the number while you can still react to it.
- Why might the forecast differ from my actual next-month cost?
- A forecast is an estimate of future consumption built from a recent rate, so it moves as your usage moves. A new model, a launch, or a quiet stretch all change it. It won't see a big change you already know is coming. And like any rate-based projection it can't see billing-side adjustments that never appear in the usage it reads, such as promotional credits or negotiated discounts.
- Does Forecast predict my next top-up or days until my balance is empty?
- No. Forecast projects next-calendar-month consumption. A prepaid top-up or card charge depends on the provider's payment rules and your account settings. Days until a balance reaches zero is Runway, a separate calculation that requires an authoritative provider balance.
- Do I have to change my code to forecast AI spend?
- No. CostCompass reads each provider's own usage or billing API directly and computes the projection from that. There's no SDK to install and no gateway in your request path, so your application runs exactly as it did before. The only thing you do is connect each provider once.
- Can CostCompass warn me when the forecast crosses a budget?
- No. CostCompass sends nothing. There are no emails, no spend caps, and nothing watching while you're away. Click Refresh and it rebuilds from the latest usage. Hard limits belong to each provider's own controls; what the forecast adds is the early read that tells you a cap or a code change is worth making. The guide to managing LLM costs covers those levers.
- Can you show the forecast math with an example?
- Say your metered usage over the last seven days came to $1,050. Divide by seven and the daily burn rate is $150. If next month has 30 days, the metered part of the forecast is $150 × 30, or $4,500. Add a $200 flat subscription that's active all month and the forecast reads $4,700.
- Why use CostCompass to forecast instead of doing it myself?
- Forecasting it by hand means logging into every provider, converting each one's usage to money, averaging a recent window, and projecting it across next month, then redoing all of it whenever you want a current number. CostCompass pulls every connected provider's usage with one click and turns it into a single forward projection, beside the month-to-date total and a per-provider breakdown, without touching your code.
About the author
Joubert Berger builds CostCompass, a spend-intelligence dashboard that pulls usage from AI and compute providers into one month-to-date total, a forecast, and a per-provider breakdown. This guide reflects how CostCompass reads each provider's own usage API — see the security model for how your keys are handled.
Know what next month costs at today's pace
Connect each provider once and pull a forward projection across all of them on demand. It extends your recent burn rate into next month, with nothing wired into your code.