GPU and Model API Costs in the Books: Cost of Revenue or R&D?

GPU compute and model API costs: cost of revenue or R&D?

Sort compute by what it does. GPU compute and model API calls that answer paying customers' requests (inference) and keep the live product running are cost of revenue. Compute spent training, testing and improving models or building software is research and development. One cloud invoice often holds both, so split it by usage.

The split decides gross margin, which investors read first, but it does not decide the tax treatment. The return asks a different question: was the spending research under section 174A, and does it meet the narrower research credit tests?

A case is not simple when:

• one cluster or API account runs both customer traffic and training, with no tags;
• free cloud credits cover part of the bill, so the books show less cost than the business will pay later;
• the company owns GPU servers, or reserves a dedicated cluster;
• compute runs in a data center outside the US;
• an audit or a priced round is coming, and past months need restating.

Gather: cloud and model API invoices by account, usage reports by project or tag, credit grant letters and balances, prepaid or committed-spend contracts, and GPU server purchase records.

Which compute costs go in cost of revenue (COGS)?

Cost of revenue holds the compute needed to deliver what customers pay for. For a product built on machine learning models, that usually means:

The SEC's income statement rule for public companies shows costs applicable to revenue, including cost of services, as a separate caption (17 CFR 210.5-03); it is a useful model for a private company's layout. Write the policy down and apply it every month. For cost of revenue in general and the board package, see monthly financials for investors and the board.

Which compute costs go in research and development?

Research and development holds compute used to create or improve the product rather than deliver it: training new models, running experiments, evaluating model versions, and the development and test environments engineers use. Model API calls that engineers make while building or testing features belong here too.

The line between the two moves over time. A model being trained is research; the same model serving customers after release is cost of revenue. Tag each workload when it starts, so the invoice can be split by usage hours or tokens instead of by estimate at month-end.

Model training costs: capitalize or expense in the books?

Model training compute is usually expensed as research and development when incurred. US accounting standards generally expense research and development costs, and capitalizing training costs as an asset fits only narrow software accounting cases. Whether a particular model qualifies is a judgment for the company and its auditor; settle the policy with the auditor before the first audit rather than after.

The book choice does not change the tax answer. Under the tax rules, whether a cost is a research expenditure depends on the nature of the activity, not on how the books record it (26 CFR 1.174-2).

How do we book prepaid compute commitments and cloud credits?

A prepayment for compute is an asset in the books and becomes expense month by month as the compute is used, on the cost of revenue or research line that matches the use. A commitment to spend a set amount over several years is not booked as expense until the compute is used; an expected shortfall the company must pay for unused commitment is an expense; agree with the auditor when to record it. Disclose large commitments to the board.

Free promotional credits cost nothing, so the books show no expense for the compute they cover, and there is no tax deduction either: a deduction requires an expense paid or incurred (26 U.S.C. 162). Credits that come with an investment or as part of a larger contract may need a value in the books; ask the auditor.

On the return, a prepayment under the cash method is generally deductible only in the year it applies to, unless it qualifies for the 12-month rule. Under the accrual method, the cost of services generally becomes deductible as the services are provided (Publication 538).

Gross margin with and without training: what will investors rebuild?

Investors rebuild gross margin from usage reports and invoices, so the reported figure should survive that check. Show it in three views:

Training compute does not belong in reported cost of revenue, but investors in model-heavy businesses may ask for the third view because training spend recurs. Showing all three, with the policy behind each line, answers the question before diligence does. Moving a cost between lines changes the margin without changing the loss, so a policy change should be disclosed to the board.

Owned GPU servers: where does the depreciation go?

Depreciation follows the server's use. A server serving customers puts its depreciation in cost of revenue; a server used for training puts it in research and development. A server used for both is split by usage hours. Book depreciation spreads the cost over the server's useful life; the tax write-off follows separate rules, covered in writing off GPU servers.

For section 174A, the purchase price of a server is not itself a research expenditure, but the depreciation on a server used in research counts as one (Rev. Proc. 2025-28, restating section 174A(d)(1)). For the research credit, depreciation on servers the company owns does not count.

How does the split tie to the section 174A deduction?

Section 174A allows a deduction for domestic research or experimental expenditures paid or incurred in the year, or an election to amortize them over at least 60 months, and treats any amount paid in connection with software development as research (Rev. Proc. 2025-28). Training and development compute is therefore usually a section 174A cost, whatever line it sits on in the books.

Inference serving customers is generally not: costs of producing a product after the uncertainty about its design is resolved fall outside section 174 (26 CFR 1.174-2), and IRS interim guidance under the former section 174 treats making software available by remote access as distribution, not development (Notice 2023-63, section 5.05). It is an ordinary business expense instead.

Two traps:

• Foreign research. Section 174A covers domestic research only; costs of research conducted outside the US are spread over 15 years. IRS interim guidance looks to where the research activities are performed (Notice 2023-63, section 3.04), but does not say where training is conducted when US engineers run it on GPUs located abroad. See software development done abroad.
• Book line is not the tax line. The return's cost of goods sold, on Form 1125-A when it applies (IRS), follows tax rules, so it need not match the books' cost of revenue.

How to claim the deduction, choose amortization, or recover earlier amortized amounts is covered in deducting US software development costs.

Does rented GPU compute count toward the research credit?

Rented compute can count, under strict conditions. The research credit includes amounts paid to another person for the right to use computers in qualified research (26 U.S.C. 41(b)(2)(A)(iii)). The regulation adds that the computer must be owned and operated by someone else, located off the company's premises, and the company must not be its primary user (26 CFR 1.41-2). The amount goes on Form 6765, line 44 (instructions).

Model API fees paid per call are a further open question: they buy a service rather than the use of a specific computer, and the statute and regulation do not say whether that counts. Claiming the research credit also changes the section 174A amount unless the reduced credit is elected; a qualified small business may use the credit against payroll tax. Both are covered in the research credit against payroll tax.

What records keep the split defensible?

Records made at the time of use hold up; allocations made months later from memory often do not. A company claiming the research credit must keep records detailed enough to show each claimed cost is eligible (26 CFR 1.41-4), and investors test the margin against the same data.

Keep:

1. a written policy listing which accounts, tags and projects go to cost of revenue and which to research;
2. separate cloud accounts or tags for production, training and development;
3. monthly usage reports by tag, tied to each invoice;
4. credit grant letters, balances used each month, and the list price of the compute they covered;
5. prepaid and committed-spend contracts;
6. GPU server purchase records, where each server runs, and its usage split;
7. for the research credit, project records showing the uncertainty and experiments each training run addressed.

Example

These examples are illustrative; figures are in US dollars.

Inference vs training

A startup spends $200,000 a month on cloud compute: $120,000 serving customers and $80,000 training its next model. Revenue is $500,000 a month, with no other cost of revenue for simplicity. With all compute in cost of revenue, gross margin is 60%. With only inference in cost of revenue, it is 76%. The 76% is the reported figure; the 60% view shows the cost of keeping the model current.

Credits run out

Free credits worth $300,000 covered inference for six months, or $50,000 a month at list price. Revenue is $250,000 a month and other cost of revenue is $25,000. The books show a 90% gross margin. At full price, it is 70%. Investors rebuild the 70%, so the board package shows both, and the plan uses 70% from the month the credits end.

Same cost, two answers

The $80,000 of monthly training compute from the first example is research and development in the books. On the return it is a domestic research cost under section 174A, if the research is conducted in the US. It may also count toward the research credit, but only if the training is qualified research and the rented compute meets the provider, premises and primary-user tests.

Different for you?

• You need the board package, the close, or cost of revenue for the whole business: see monthly financials for investors and the board.
• You are claiming section 174A or recovering amounts amortized in earlier years: see deducting US software development costs and get startup tax help.
• Developers or GPUs are outside the US: see software development done abroad.
• You want the research credit against payroll tax: see the research credit against payroll tax.
• You buy GPU servers: see writing off GPU servers.
• You rent out spare GPU capacity: see renting out GPU servers, and sales tax on hosting and GPU compute for what you bill.
• An audit or priced round is coming and past months were never split: restating them is a judgment call; get startup accounting help.