Accrued Expenses Journal Entry for Variable Compute Costs
How to book unbilled GPU, token, and API usage in the right period.

Variable compute costs make an accrued expenses journal entry a data problem as well as an accounting task. GPU-minutes, tokens, and API calls must be assigned to the period in which they were consumed, even when billing data remains incomplete. Reliable accruals therefore depend on three connected decisions, namely how usage is estimated, which account receives the expense, and how the estimate is cleared when the invoice arrives.
Why variable compute costs break the standard accrual entry
A payroll accrual can start with a known amount and a firm payment schedule, while the amount of a GPU-minute accrual may be unclear until the cloud provider’s billing API settles. Pay for work already done, rent for space already occupied, and interest on a fixed balance are based on amounts that are largely known. Compute consumption has no natural stopping point at month-end. The accounting cutoff divides a continuous stream of usage before the billing system has necessarily finished measuring it.
That creates a gap between metering finality and accounting close. A billing cycle may close on the 1st while usage from the last several days of the prior month remains uninvoiced. GAAP and IFRS require expenses to be recorded when incurred rather than when paid, so incomplete billing data does not remove the obligation. It changes the amount from a settled figure into an estimate.
What the matching principle actually requires when usage hasn't settled
The matching principle requires compute costs to fall in the period whose revenue they supported, regardless of when the cloud invoice arrives. Consumption establishes the obligation before the invoice documents it. Accrued expenses include costs for usage that has happened but has yet to be billed or paid. The amount takes judgment, but recognition cannot wait until that amount is certain. The later invoice supplies the basis for correcting that judgment.
For AI companies, this requires accounts that distinguish the obligation from the economic function of the spend. “Accrued cloud and model usage” records the cutoff liability for uninvoiced consumption. It is separate from “Cost of revenue — inference compute,” “Cost of revenue — third-party model usage,” and “R&D cloud training compute.” These accounts answer different questions about the same usage. The liability records what is owed, while the expense classification identifies what the compute supported.
Tax and capitalization questions sit outside that basic entry. The IRS requires businesses above an inflation-indexed gross receipts threshold to use the accrual method, but they still have to decide how to classify costs for financial reporting. Compute that directly fulfills current customer requests defaults to cost of revenue. Compute aimed primarily at improving future models or products defaults to R&D. Keeping those functions separate helps with both gross margin reporting and the separate review of how to treat the costs for tax and capitalization.
The journal entry mechanics for a variable compute accrual
Consider inference compute consumed in May but not invoiced until June. The example uses the estimated May usage cost as the amount, with no training workload included. The entry follows ordinary double-entry accounting. Both sides use the same estimate of consumption incurred but not yet invoiced, so the uncertainty affects the amount while the entry’s structure stays the same.
The May period-end entry is dated the last day of May and records the unbilled obligation. It debits Cost of Revenue, Inference Compute and credits Accrued Liabilities, Cloud and Model Usage for the best available estimate. No invoice is required to support recognition of consumption. This entry places the estimated expense in May and the unpaid obligation on the period-end balance sheet.
The June period-open entry is dated the first day of June and reverses that estimate. It debits Accrued Liabilities, Cloud and Model Usage and credits Cost of Revenue, Inference Compute for the same amount. When the invoice arrives in June, the invoice entry debits Cost of Revenue, Inference Compute and credits Accounts Payable for the actual billed amount.
Together, the June reversal and invoice leave only the estimate-to-actual difference in June expense for that May usage. An ordinary difference is a change in estimate that adjusts the current period and leaves the prior period closed. A missing reversal has a different effect. It leaves the accrual in place while the invoice records the expense again, producing double-counting.
How to estimate the accrual when the meter hasn't stopped
The appropriate estimation method depends first on what usage data is available at close. If the billing system tracks consumption per customer in real time, a customer-level rollup can aggregate in-flight usage through the close date. This is the most accurate approach described here. This approach requires the billing infrastructure to surface that consumption when finance needs to post the entry, even if billing is still underway.
If customer-level data is unavailable, provider cost data offers a direct view of incurred costs. AWS Cost Explorer, GCP Billing, and Azure Cost Management publish daily or near-daily cost data before the final invoice. Their close-date output can leave only the latest hours or days to estimate. Freshness still determines how much uncertainty remains in the entry. Provider data can capture a late-month traffic spike once that usage appears, whereas extrapolating earlier consumption would miss it.
When actual consumption is known only through day 28, or the last available meter read, a run-rate estimate works for usage that is fairly steady. The calculation divides known consumption by days elapsed and multiplies the result by total days in the period. If traffic spikes in the last three days, the earlier daily average can materially understate the period’s cost.
If there is too little usage data for the current period, finance can start with the prior month’s settled invoice. Finance can adjust it with a growth factor derived from product usage trends or customer count changes. This is faster, but it introduces model risk. High growth or seasonality can make the previous period a poor guide, so large corrections that keep coming back show that the method no longer fits how compute is being used.
Why cost classification — inference versus training — changes which account gets debited
The debit account depends on the economic function of the compute. Inference serving live customer requests belongs in Cost of Revenue, Inference Compute because it fulfills revenue earned in the period. Training and fine-tuning to improve what models can do in the future default to R&D, Cloud Training Compute. The liability account can remain the same in both cases. What changes is where the cost appears on the income statement.
Consider a company running a nightly fine-tuning job and a live inference API on the same cloud account. The shared provider bill does not make the workloads economically identical. The live API’s consumption supports customer service, while fine-tuning aimed at future capability supports development. The accrual thus needs to be split between the two expense accounts. For shared or ambiguous workloads, including a training run that also improves the production model currently in service, that allocation requires documentation when the workload runs so the evidence is already available at close.
Classifying training as cost of revenue understates gross margin, while classifying inference as R&D overstates it. Both errors distort the unit economics that finance teams and investors use. A settled invoice can resolve the total amount owed without resolving that classification. The allocation evidence must explain what the underlying workloads did.
What real-time metering data does for accrual reliability
Accrual reliability is bounded by the usage data available at close. A metering system that settles hours after period-end leaves less consumption to estimate than one that settles days later. This cutoff advantage leaves the accounting policy unchanged. A real-time pipeline can show consumption through the final minute of the period, while a batch system settled two days later leaves a two-day estimation gap.
Consumption pricing relies on event ingestion, aggregation, and pricing. Ingestion records the billable event when it happens, aggregation adds up events for each customer and billing period, and the pricing engine uses the rate card to price those totals. Finance’s usage estimate is only as fresh as the aggregation output. Fast capture alone does not close the gap if the totals arrive late. At AI volumes, metering needs to handle a stream of events as well as the batch calculation used for invoicing.
Idempotent ingestion prevents retries from counting the same event again, while deduplication prevents an event arriving from multiple sources from being counted twice. Events that arrive late must be put back in the billing period where they belong. Otherwise, missing events create a recurring downward bias. Timely aggregation and event handling that can be trusted mean fewer large invoice corrections and a more reliable accrued cloud and model usage balance.
How the revenue-side accrual mirrors the cost-side entry — and where ASC 606 adds complexity
Consumption can create an unbilled revenue accrual before the customer invoice is issued, just as it creates a cost obligation for the buyer. Recognition is not governed by cost accrual mechanics alone. ASC 606 adds rules on contract structure and variable consideration, so a usage total is evidence but cannot on its own settle every question of revenue recognition.
At month-end, the revenue accrual debits Accrued Revenue / Contract Asset and credits Revenue for the amount recognized. When the invoice is issued in the following period, the entry debits Accounts Receivable and credits Accrued Revenue. This reclassifies the asset while leaving the revenue recognized. That distinction separates the revenue-side sequence from the reversing expense accrual.
For pay-as-you-go contracts, the “right to invoice” practical expedient supports recognition in the period usage occurs where applicable. For prepaid credits, cash is first recorded as deferred revenue, and revenue is recognized as the credits are used. Both treatments depend on a reliable consumption cutoff. Most billing tools do not automatically create the accrued revenue entry for usage between the last billing cutoff and period-end, leaving that entry to a manual process or a system that tracks consumption in real time.
What a high-accrual-variance pattern tells finance teams to fix
Large differences between accruals and invoices that keep coming back point to a flaw in the process beyond the uncertainty of any one estimate. Diagnosis starts by checking whether the estimation method fits the usage pattern. A run-rate applied to spiky, event-driven consumption can repeatedly understate the obligation. The corrective action is to move toward a metering-data-based method. If the method already uses consumption data, the next question is whether that data arrives in time to cover the period.
Metering latency and classification drift create different symptoms. A system that settles two days after close leaves the same recurring gap, which real-time aggregation and late-arriving event handling address. Workloads moving between inference and training without updated allocations instead produce errors within expense lines. The total may look right even when gross margin is skewed. An expense line that remains too high relative to the invoice also calls for checking whether the prior accrual was reversed.
Ordinary estimation differences adjust the current period, but repeated directional errors weaken close-to-close comparisons and the reliability of cost per inference, compute as a percent of revenue, and gross margin by workload. The final question concerns the data’s path into the estimate. Separate cloud invoices, third-party model API invoices, and internal metering logs take more work to reconcile as the number of sources grows. Where fragmentation drives the discrepancy, consolidating usage events into a single metering layer addresses the cause before it reaches the journal entry.


