Summary
- Snowflake's contractual Service Consumption Table sets four billing minimums across its compute classes, of one minute, five minutes, sixty minutes and sixty seconds, and the class that pays sixty minutes is defined in the same sentence as a type of the class that pays one.
- The credit, the warehouse-size ladder, the on-demand credit rate and the storage rate are all printed in documents that carry a date; no Snowflake document can print the number of chargeable starts, which is the customer's query cadence read through a suspend parameter the reference says is not intended for precise control.
The unit as printed
Snowflake defines its own meter in a sentence a reader can hold onto: "A Snowflake credit is a unit of measure, and it is consumed only when a customer is using resources, such as when a virtual warehouse is running, the cloud services layer is performing work, or serverless features are used." That is on the compute-cost page (Understanding compute cost).
The contract uses a different noun for the same thing, and says so: references to "Snowflake Credits" or "Credits" in order forms, agreements and documentation are to be interpreted as Platform Credits (Snowflake Service Consumption Table).
The metering rule follows immediately on the same documentation page: credits "are billed per-second, with a 60-second (1-minute) minimum", and "Each time a warehouse is started or resumed, the warehouse is billed for 1 minute's worth of usage based on the hourly rate in the Snowflake Service Consumption Table" (Understanding compute cost). Read the grammatical subject of that second sentence. It is a start. Not a second of runtime, not a query, not a session. The per-second clock is what happens after the start has already bought its first minute.
The floor as four numbers
Here is the contractual sentence, in full, because its shape is the finding:
Unless otherwise noted, when a Virtual Warehouse or a Postgres Compute Node is started or resumed, a minimum of one minute's worth of Platform Credits will be consumed; when an Interactive Warehouse (a type of Virtual Warehouse) is started or resumed, a minimum of sixty (60) minutes' worth of Platform Credits will be consumed; when a Compute Node (except for Postgres Compute) is started or resumed, a minimum of five minutes' worth of Platform Credits will be consumed.
A fourth minimum sits one line later: for Snowflake Openflow Compute, Platform Credits are charged on a per-second basis with a 60-second minimum. Thereafter, the table continues, Virtual Warehouses and Compute Nodes "are charged on a per-second basis, rounded up to the nearest whole second". A suspended warehouse uses no Platform Credits at all (Snowflake Service Consumption Table).
Two things about that sentence deserve to be read slowly. The first is its opening: "Unless otherwise noted" announces that the rule to follow is not universal, and the sentence then supplies its own exceptions. The second is the parenthetical. The class paying sixty minutes is defined, in the same breath, as "a type of Virtual Warehouse", which is the class paying one.
The documentation describes the same instrument in the singular. "The minimum billing charge for provisioning compute resources is 1 minute (i.e. 60 seconds)," it says, and then works the arithmetic of interruption: a warehouse that runs for 30 to 60 seconds is billed for 60 seconds, and a warehouse that runs 61 seconds, stops, restarts and runs less than a minute is billed for 121 seconds, written out as 60 + 1 + 60 (Warehouse considerations). There is no benefit to stopping before the first minute is over, the same page says, because those credits are already billed.
My divisions, not Snowflake's characterisation: sixty minutes divided by one minute is 60, and five minutes divided by one minute is 5. The sixtyfold figure is a ratio between two printed durations and nothing else. The documents I read print no hourly credit rate for Interactive Warehouses, so I attach no dollar figure to that class and stop the comparison at the ratio.
Which document governs if a reader finds a discrepancy? Snowflake's own calculator page answers that in one sentence: "The Consumption Table reflects the most accurate pricing should any inconsistencies arise" (Snowflake Pricing Calculator).
The multiplier and the two ladders
The multiplier is printed, twice, and the two printings do not say the same thing. Table 1(a) of the consumption table gives standard Virtual Warehouse Platform Credits per hour as 1, 2, 4, 8, 16, 32, 64, 128, 256 and 512 from X-Small to 6X-Large. Table 1(b) gives Gen2 rates: on AWS and GCP, 1.35, 2.7, 5.4, 10.8, 21.6, 43.2, 86.4 and 172.8; on Azure, 1.25, 2.5, 5, 10, 20, 40, 80 and 160 (Snowflake Service Consumption Table). An X-Small Gen2 warehouse therefore bills 1.25 credits an hour in one cloud and 1.35 in another. The size name is identical; the multiplier is not.
The documentation's overview table prints the first-generation ladder twice, in credits per hour and in credits per second: 1 and 0.0003, 2 and 0.0006, up to 512 and 0.1422 (Overview of warehouses). Multiplied by 3,600 seconds, 0.0003 reconstructs to 1.08 credits an hour against a printed 1, and 0.1422 reconstructs to 511.92 against a printed 512. Every cell in that column behaves the same way, which is what a four-decimal rendering of a repeating per-second rate does. It is a display convention, and I would not read a second meter into it.
Now the arithmetic that matters, and it is mine. At X-Small, one credit per hour, a one-minute minimum is 1 divided by 60, or 0.016667 credits. At the AWS US East Standard on-demand rate of $2.00 per credit printed in Table 2(a), that floor costs $0.0333. A sixty-minute minimum applied at that same hourly rate would be one whole credit, or $2.00. Twelve starts inside a first minute, each restarting the minimum as the documentation describes, bill twelve minutes for at most sixty seconds of running time: about $0.40.
The knob and the poller
The parameter that decides how many floors get bought is documented as a customer setting. AUTO_SUSPEND "specifies the number of seconds of inactivity after which a warehouse is automatically suspended"; valid values are any integer of zero or greater, or NULL; "setting 0 or NULL means the warehouse never suspends"; and "The default AUTO_SUSPEND value is 600, meaning the warehouse suspends automatically after 10 minutes of inactivity" (CREATE WAREHOUSE, ALTER WAREHOUSE).
The same reference then attaches a sentence that changes what the setting is worth: "The background process that suspends a warehouse runs approximately every 30 seconds, so AUTO_SUSPEND is not intended for precise control over suspension." AUTO_RESUME is a second customer setting, TRUE by default, and FALSE means the warehouse resumes only when explicitly resumed.
My reading of those two sentences together: a warehouse with AUTO_SUSPEND set to 60 seconds can sit idle for somewhere between 60 and 90 seconds before the suspending process reaches it. The customer prints a number the mechanism cannot honour to the second, and the quantity that comes back is a range rather than the setting. At the default 600 seconds, ten minutes of inactivity is 0.16667 credits at X-Small, or $0.333 at $2.00. If the warehouse is multi-cluster, auto-suspend occurs only when the minimum number of clusters is running and there is no activity (Overview of warehouses).
The second meter and its denominator
Cloud services are the second meter, and they carry a denominator of their own: "Usage for cloud services is charged only if the daily consumption of cloud services exceeds 10% of the daily usage of virtual warehouses" (Understanding compute cost). My arithmetic, on that sentence: one hour of a single X-Small warehouse is one credit, so that day's allowance is 0.1 credit, or $0.20 at $2.00. A day in which no warehouse credit is consumed leaves a zero allowance under the printed ratio, and no document I read addresses that case; I state it as my reading of the sentence, not as Snowflake's statement.
Storage is the third line and the simplest to rebuild. Table 3(a) prints $23.00 per TB per month on demand for AWS US East, with capacity tiers falling through $21.47, $19.94, $18.40, $16.86, $15.34 and $13.80 as the tier rises, and charges are based on the average terabytes per month of all Customer Data, using the compressed file size where data is compressed (Snowflake Service Consumption Table). The documentation's own worked example multiplies 4,464 credits by $2 for $8,928 and 65 terabytes by $23 for $1,495, totalling $10,423, and that page defers current rates to the Pricing Guide (Understanding overall cost).
One detail on that table, reported as printed: the capacity heading reads "Capacity Storage Pricing by ACV Range (TB/mo)" while the tiers beneath it are dollar ACV thresholds. I record the mismatch and leave its footing unresolved rather than supply a bridge nobody printed.
Where the dollar is, and where my arithmetic stops
The dollar is printed, and it is dated. The table is headed "Effective: November 20, 2026", its dollar values are in USD, and its changes clause reads: "This Snowflake Service Consumption Table may be updated from time to time. Changes shall be effective on the date that Snowflake announces they are effective" (Snowflake Service Consumption Table).
Where my arithmetic stops is somewhere else, and it is worth being exact about it. A reader can rebuild the billable quantity of a running warehouse from Table 1(a) or 1(b), the price of a credit from Table 2(a), and the storage charge from Table 3(a), all in one dated document. What no Snowflake document contains is the number of chargeable starts, because that number is generated by query cadence, scheduling and connection behaviour on the customer's side, filtered through a suspend parameter whose own reference says it is not for precise control.
Multiply a printed floor by an unprinted count and the product is not a printed bill; it is a customer's log, multiplied.
Three reading limits belong here rather than in a closing note. The consumption table reached me as provider-reported PDF text rather than a page-by-page binary parse, so regional rows and footnotes are less exhaustively checked than the credit ladders. Snowflake's marketing pricing page returned HTTP 403 and is therefore not cited, not characterised, and not used as evidence that anything is absent from it. And the documentation pages carry no visible date in the text I read; only the contractual table states when it takes effect.
The credit is a real unit and the floor is a real floor. Snowflake prints every price a start is multiplied by, and prices the start rather than the second. Whether your account pays one floor or forty of them in an afternoon is a fact about how your warehouse is started, which is the one input no vendor's price list can hold.
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.






