Summary
- Vercel prices a function on two metered quantities and one counter: Active CPU, Provisioned Memory and invocations. The CPU meter pauses while code waits on external services; the memory meter is billed for the lifetime of the instance.
- The vendor's worked example re-derives exactly in both directions from the printed São Paulo rates of $0.221 per CPU-hour and $0.0183 per GB-hour, so a reader holding the same page can rebuild that line to the precision printed.
- The interval in which memory accrues and CPU does not is printed as a number and named nowhere: it is the six seconds a reader subtracts between the example's 4 seconds of active CPU and its 10-second instance lifetime.
The example is small enough to check on the back of an envelope, and that is where the audit should start. A function with 4 GB of memory, 4 seconds of active CPU and a 10-second instance lifetime, running in São Paulo: the Fluid compute pricing page prints $0.0002456 for the CPU, $0.0002033 for the memory, and $0.0004489 per invocation.
Rebuild it. $0.221 per CPU-hour is the printed gru1 rate, four seconds is 4 ÷ 3,600 of an hour, and the product is $0.00024556. Memory is 4 GB held for 10 seconds, so 4 × 10 ÷ 3,600 = 0.0111 gigabyte-hours at $0.0183, which is $0.00020333. The sum is $0.00044889. Those are my divisions on the vendor's printed rates, run on the vendor's printed inputs.
Run it backwards, which is the harder test of a worked example. Divide the printed CPU line by $0.221 and multiply by 3,600: 4.0007 seconds. Divide the printed memory line by $0.0183 and multiply by 3,600: 39.99 gigabyte-seconds, which at the example's 4 GB is 9.998 seconds. Both arrive back at the example's own inputs, give or take the rounding already inside the printed figures. The line is auditable from the page that carries it, and that is rarer in this genre than it ought to be.
What the page does not print is the interval that makes the arithmetic interesting.
Two definitions that do not share a stop condition
The CPU meter is defined by exclusion. On the pricing page: This is the CPU time your code actively consumes in milliseconds, followed by You are only billed during actual code execution and not during I/O operations (database queries, like AI model calls, etc.) and the bullet Pauses billing when your code is waiting for external services. The compute explainer says the same thing in one line: Active CPU: The CPU time your code actively consumes. Billing pauses while your code waits on external services.
The memory meter is defined by allocation and by lifetime. The same explainer: Provisioned Memory: The memory allocated to your function instances, billed for the lifetime of each instance. The memory and CPU configuration page states it differently: Provisioned Memory: Memory allocated to the function instance while requests are in progress. The Active CPU troubleshooting guide states it a third way, in a table row: Provisioned memory | Memory allocated while requests are in flight.
Three pages, three perimeters, one quantity. Lifetime of the instance, while requests are in progress, while requests are in flight. Those describe the same meter and not the same moment at which it stops, and none of the eight documents sets out which phrasing governs.
The difference is not pedantry, because the execution model puts several requests inside one instance. The Fluid compute page states it plainly: Fluid compute allows multiple invocations to share a single function instance. An instance can therefore outlive the requests it served, and requests can overlap inside one lifetime. Which perimeter the memory meter actually closes on is exactly the question the three phrasings leave open.
The stop rule, in four sentences on two documents
No reviewed page prints the full stop rule. It assembles, like this. The CPU meter pauses, per the pricing page: Pauses billing when your code is waiting for external services. The memory meter runs for the instance's life, per the explainer: billed for the lifetime of each instance. Both meters end with the instance, per the pricing page: After all requests complete the instance is paused and no CPU or memory charges apply until the next invocation. And nothing is charged in between, per the explainer: You never pay for idle CPU, and you pay nothing between requests.
Four sentences, two documents, one rule. A reader who lands on the pricing page alone gets the pause and the instance stop; a reader who lands on the explainer alone gets the pause and the between-requests sentence. Neither document prints the rule whole, and the quantity the rule creates appears on neither as a line item.
The six seconds
The example's own figures expose the interval without naming it. Ten seconds of instance life, four seconds of active CPU. What remains is six seconds in which memory accrues and CPU does not. The subtraction is mine; the page prints 4 and 10 and leaves the difference to whoever wants it. Those six seconds are 60 percent of the example's wall-clock life, billed to the memory meter alone. No page among the eight puts a name, a rate or a rule on that interval, and no page prices it separately, because it is not a separate charge: it is the part of the memory line that has no CPU line beside it.
Allocation is printed, and one sentence calls it a rate
The configuration page also states who may turn the dial. Hobby functions always execute with 2 GB and 1 vCPU and memory size cannot be configured there; Pro and Enterprise default to the same 2 GB and 1 vCPU and go up to 4 GB and 2 vCPU. Then the sentence: Larger memory settings increase the GB-hour rate for this dimension. Raising the setting does not move the unit price, which the region table sets at $0.0106 a gigabyte-hour in iad1, cle1 and pdx1 and $0.0183 in gru1. What the setting moves is the number of gigabyte-hours billed. The page called a quantity a rate.
Two printed rates, one word, a factor of ten
Here is the sentence a buyer is most likely to carry away from the vendor's own guidance. The vendor's cost-comparison guide: Vercel bills for CPU only while your code is running, pauses CPU billing during I/O waits, and charges nothing for CPU between requests. You still pay for provisioned memory while a request is in flight, at $0.0106 per GB-hour, less than 10% of the active CPU rate.
Divide the two printed unit prices: $0.0106 ÷ $0.128 = 8.28 percent. The sentence holds as a ratio of two unit prices in different dimensions, one denominated per gigabyte-hour and one per CPU-hour. In the vendor's own example the memory line is 82.8 percent of the CPU line, $0.0002033 ÷ $0.0002456. Also my division, also printed figures, also correct, and a different ratio, because its terms are bill lines rather than unit prices.
Two ratios, one word: rate. A reader who takes less than 10 percent as a statement about his own bill and a reader who takes it as a statement about the price list are both reading the sentence as written, and their figures sit a factor of ten apart. The comparison sentence also prints the $0.0106 memory figure without saying which region's active CPU rate supplies its denominator, which is a fact the sentence omits rather than gets wrong.
The free plan makes the two meters one test
Hobby includes 4 CPU-hours of Active CPU, 360 GB-hours of Provisioned Memory and the first 1,000,000 function invocations, per the Hobby plan page. Hobby functions run at a fixed 2 GB. Divide: 360 ÷ 2 = 180 instance-hours of memory. Divide that into the CPU allowance: 4 ÷ 180 = 2.22 percent. The two allowances run down together only at a workload where active CPU occupies just over one fiftieth of instance time, and no reviewed page names that shape.
Push the same allowances to the page's own ceiling and the spread widens. The limits page gives Hobby a 300-second maximum duration, default and maximum. Three hundred seconds is 1/12 of an hour, so a function that executes code for its whole life consumes 1/12 of a CPU-hour per invocation: 4 ÷ (1/12) = 48 invocations. A function that waits for its whole life consumes no CPU and 2 × 300 ÷ 3,600 = 0.1667 gigabyte-hours per invocation: 360 ÷ 0.1667 = 2,160 invocations. Forty-eight against 2,160, a factor of 45 between two shapes of the same function on the same allowance. Both multiples are mine, and both shapes are extremes no page asks anyone to build.
The third dimension is a counter rather than a clock: $0.60 per million invocations on Pro, and the pricing page says invocations count regardless of request success or failure. That is the one component of this bill a reader can predict before deploying anything.
What the eight documents do not state
None gives the memory-only interval a name. None states a minimum billing increment, a per-invocation floor or a rounding rule for either metered dimension. None reconciles the three phrasings of the provisioned-memory quantity. None names the workload shape at which the Hobby allowances deplete together. None states an applicability rule for the two rate-card regions beyond the region codes themselves. These are reading limits scoped to eight documents read on 21 November 2026, not claims about what the company publishes elsewhere.
What the audit yields
Set the two findings side by side and the instrument's shape becomes clear. The meters are published, the rates are published by region, the counter is published, and the worked example closes in both directions, so the invoice line a customer sees can be rebuilt from the same page that explains it. What remains unpublished is the quantity that comes between the clocks: the interval in which memory is the only thing accruing, which the example shows as six seconds carrying one meter and no other, and which no document names.
Measurement ownership here is ordinary and worth naming precisely. Vercel defines both meters, chooses which page prints which half of the stop rule, and supplies the rate card. The customer who wants to reconcile a past invoice is well served. The customer who wants to forecast a future one has to supply the missing quantity from his own traces: the share of instance time his code spends executing, and the concurrency his instance actually achieves. The vendor prints enough to be checked. The checking is still his work.
Vercel, Inc keeps a profile in the BTW directory: entity:vercel-inc.
Member Briefing
Read the full trend briefing
Members read the full briefing with the source notes behind every finding.
Strategic Circle
A daily briefing in your language, built from changes we have verified.
Join for USD 20/monthCancel anytime.
Members also get the full research library, a daily briefing on the networks they follow and the private forum.






