Executive Summary
- AI companies look like software businesses on the surface, but their books carry usage-based revenue, heavy compute costs, prepaid credits, and harder capitalization and tax questions that a generic startup setup handles poorly.
- Revenue, compute classification, and training-data capitalization are deep enough to deserve their own treatment, so this guide summarizes those areas and links to the detailed posts, then concentrates on the parts founders most often miss.
- The high-leverage, less-covered topics are foundation-model and cloud-credit accounting, R&D tax credits under Section 174A, equity instruments like SAFEs and options, and multi-entity structure for groups whose core value is model IP.
- Compute discipline drives gross margin, and tax outcomes now depend on clean separation of domestic and foreign research, business-component records, and a chart of accounts that maps spend to economic function.
- The winning setup tracks revenue at the contract and usage level, separates inference from training and product development, documents capitalization and data-rights decisions, and keeps equity and intercompany records audit-ready month to month.
If your AI company is outgrowing a generic startup bookkeeping setup, Ridgeway Financial Services builds finance systems for usage-based revenue, compute cost discipline, capitalization policy, and R&D tax substantiation that hold up in fundraising, audit, and diligence. Explore our Accounting and Bookkeeping services.
Table of Contents
- Why bookkeeping for AI companies is different
- Revenue recognition for AI business models
- Where compute, GPU, and model costs belong
- Dataset licenses, labeling, and training data
- Foundation-model and cloud-credit accounting
- R&D tax credits and Section 174A for AI companies
- SAFEs, convertible notes, 409A, and stock compensation
- Multi-entity accounting, transfer pricing, and FX
- A sample chart of accounts for AI companies
- Controls, audit readiness, and the KPIs that matter
- The monthly close for AI companies
- How Ridgeway Financial Services helps
Why bookkeeping for AI companies is different
According to Ridgeway Financial Services, AI bookkeeping breaks on generic startup templates for three structural reasons. First, the revenue portfolio is more variable than classic subscription software, mixing API calls, token usage, seats, committed-spend contracts, enterprise licenses, implementation work, and sometimes intellectual-property licensing. Both revenue and margin depend on accurate consumption data, so the finance function needs stronger operational feeds than many software peers.
Second, cost risk concentrates in places a single hosting line never captures: GPU capacity, cloud commitments, storage, third-party model calls, evaluation and red-team labor, labeling vendors, and ongoing retraining. Third, an AI company often straddles several accounting models at once. Internal product tooling can fall under ASC 350-40, software built to be sold can fall under ASC 985-20, broad exploratory work sits in ASC 730, and acquired data rights may be evaluated under ASC 350-30 or expensed as research depending on alternative future use.
For the full industry picture, see our overview of AI and machine learning startup accounting and finance. The sections below summarize the areas with dedicated guides, then go deeper on the topics those guides do not cover.
Revenue recognition for AI business models
ASC 606 and IFRS 15 both require the same five steps: identify the contract, identify distinct performance obligations, set the transaction price, allocate it, and recognize revenue as obligations are satisfied. For AI companies the practical question is which promise the customer is paying for: access, usage, outputs, implementation, customization, a license, or a combination.
The high-frequency setups are usage-based APIs (often recognized as usage occurs and becomes billable), prepaid credits (a contract liability first, not day-one revenue, with breakage analyzed separately), enterprise minimums plus overages, distinct versus bundled licenses, and onboarding or fine-tuning services that need standalone-selling-price allocation. We walk through each model, the breakage rules, and the journal entries in our full guide to revenue recognition for AI products.
Where compute, GPU, and model costs belong
The central bookkeeping question is not only what you spent, but what economic function the spend served. A workable baseline policy: if compute directly fulfills customer requests in the current period, default toward cost of revenue; if it is primarily aimed at improving future models or products, default toward R&D unless it clearly qualifies for software capitalization; if it is internal tooling, assess ASC 350-40.
That only works if your books mirror your cloud telemetry. Split compute into at least production inference, production training, research training, staging, internal tools, and overhead, using provider cost exports, tags, and amortized-cost views. For the forecasting model and unit economics, see GPU cost forecasting and AI unit economics. For where development compute gets capitalized or expensed, see accounting for AI development costs.
Dataset licenses, labeling, and training data
US GAAP has no standalone model for data, so entities fall back on broader intangible-asset and research guidance. Data generated or licensed solely for a particular R&D project with no alternative future use is generally expensed as incurred. Licensed data with contractual rights and alternative future use is generally capitalized under ASC 350-30 and amortized over its useful life. The same dataset contract can land in different places in different months depending on whether the related project is exploratory, capitalizable internal-use development, or post-feasibility production.
Because of that, the contract alone is never enough. Finance needs project-stage support and intended-use documentation attached to the spend. We cover the full decision tree, amortization, and audit considerations in accounting for AI training data under U.S. GAAP.
Foundation-model and cloud-credit accounting
Many AI companies are built on someone else’s hosted model, and the accounting for those fees is frequently mishandled. Most acquired foundation-model or large-language-model arrangements are service contracts rather than purchased or licensed software. When the economics are pay for access to a hosted model, upfront payments usually create a prepaid asset for future use of functionality, while ongoing usage or maintenance costs are expensed as incurred. Only when a contract conveys a separable, identifiable software or other right does the analysis move toward intangible-asset treatment, and that is not the default in current AI infrastructure arrangements.
Cloud commitments and credits need the same substance-over-form discipline. The practical treatments look like this:
| Arrangement | Likely treatment | Practical note |
|---|---|---|
| Cash paid upfront for future cloud or model access | Prepaid asset, amortized as services are received | Align amortization to the benefit period and actual usage where appropriate |
| Provider discount applied directly to usage pricing | Record lower expense as incurred | No separate prepaid asset when no standalone future right exists |
| Reservation or savings-plan purchase | Book the purchase, then allocate cost over the benefit period for internal reporting | Use amortized-cost views, not just actual cost, in product-margin reporting |
| Promotional credits | Usually reduce eligible cloud expense as applied | Do not record unvested provider marketing promises as assets until granted and controlled |
One caution worth stating plainly: treating every model-access fee as a software asset, or sweeping prepaid credits straight into expense without tracking the future right, both distort margin and the balance sheet. The arrangement type, not the vendor’s invoice format, drives the entry.
R&D tax credits and Section 174A for AI companies
Tax treatment of research costs has shifted, and the change matters most for companies that spend heavily on engineering and data. Under current IRS guidance, for tax years beginning after December 31, 2024, domestic research or experimental expenditures are generally deductible when paid or incurred under IRC Section 174A, with an elective path to capitalize and amortize domestic amounts over not less than 60 months. Software development can be included in domestic research or experimental expenditures for these purposes.
Foreign research is treated differently. It remains under IRC Section 174 and continues to be amortized over 15 years. For an AI company with offshore annotation teams, a foreign research subsidiary, or an overseas engineering center, that split creates both cash-tax consequences and deferred-tax complexity, because domestic work may be deducted immediately while foreign work is still capitalized and amortized.
The federal research credit under IRC Section 41 is separate from deduction timing. A regular credit and an alternative simplified credit are both available, and qualified small businesses may elect to apply part of the credit against payroll taxes, subject to the current statutory cap. The compliance burden is rising: Form 6765 now phases in detailed business-component information, so project-level records, wage mapping, contractor support, and data-cost substantiation are becoming far more important.
The bookkeeping consequence is direct. Book R&D, tax R&D, and capitalized software can diverge sharply, so you need a clean bridge between the GAAP ledger and the tax workpapers. If your books already separate model development, experimentation, data work, production support, and domestic versus foreign labor, your credit study is faster, stronger, and cheaper. State credits vary widely on eligibility, refundability, and definitions, so model them separately rather than assuming they follow federal. Founders weighing non-dilutive funding alongside credits may also want our guide to SBIR and STTR funding.
SAFEs, convertible notes, 409A, and stock compensation
Equity and near-equity instruments are where AI companies most often defer work to year-end and then pay for it in audit delays. A standard legal form does not remove the accounting analysis. SAFEs should be reviewed under the liability-versus-equity and embedded-feature frameworks rather than assumed to be simple equity, because redemption economics, settlement alternatives, side letters, or nonstandard terms can change classification and measurement.
Convertible notes are simpler legally but still need a schedule that separately tracks principal, accrued interest, discounts or premiums, conversion terms, maturity, and any bifurcated features. For 409A, the regulations provide a presumption of reasonableness for several valuation methods, including an independent appraisal dated no more than 12 months before the relevant transaction, which is one reason a serious close should maintain an equity-events log.
Stock-based compensation under ASC 718 is frequently underbuilt at AI companies that lean on equity to hire senior engineers and researchers. Cap-table administration, board approvals, grant-date records, and valuation support belong inside the monthly close, not in an annual cleanup. The same records feed diligence, so keeping them current pays off twice. See our fundraising data-room checklist and our guide to managing cash runway.
Multi-entity accounting, transfer pricing, and FX
The international story is more sensitive for AI than for many startups because the group’s value often sits in intangibles. Section 482 authorizes the IRS to adjust income among commonly controlled taxpayers so that intercompany pricing for goods, services, and intangibles reflects arm’s-length outcomes. The OECD guidelines give special attention to unique and valuable intangibles and to hard-to-value intangibles, which is exactly where core model IP, training-data rights, model weights, and proprietary evaluation systems fall.
The practical task is to decide whether a non-US affiliate is a reseller, a contract research provider, a shared-services center, or an intangible owner, then align contracts and bookkeeping to that conclusion. A common trap is treating central model-development work as a low-markup service center, since the simplified approach for low-value-adding services does not apply to core R&D, including software development, when that work is part of the group’s core business.
On currency, IAS 21 frames the issue around functional currency and foreign-currency transactions. Each entity should have a documented functional-currency memo, an intercompany settlement policy, and a month-end remeasurement process for monetary balances. Without those, intercompany noise and FX volatility distort burn, margin, and subsidiary performance.
A sample chart of accounts for AI companies
A generic SaaS chart of accounts hides the things that matter most for an AI company: where compute lands, how data rights are carried, and how customer prepayments are deferred. The structure below is illustrative, designed to support revenue, compute, R&D, capitalization, and fundraising accounting in one ledger.
| Account | Why it exists |
|---|---|
| Prepaid cloud and model credits | Upfront rights to future cloud or hosted-model service |
| Capitalized internal-use software | ASC 350-40 assets once the threshold is met |
| Capitalized external-use software | ASC 985-20 production costs after technological feasibility |
| Licensed datasets and data rights | Capitalized licensed data with alternative future use |
| Accrued cloud and model usage | Cutoff for usage incurred but not yet invoiced |
| Deferred revenue, prepaid credits | Unused customer credits carried as a liability |
| Deferred revenue, minimum commits | Enterprise contract liabilities |
| Deferred revenue, services | Onboarding, fine-tuning, and support obligations |
| SAFE liability or equity-classified SAFE | Depends on instrument analysis |
| Convertible notes payable | Principal and accrued interest tracking |
| API usage revenue | Core consumption revenue |
| Breakage revenue | Separate visibility for prepaid-credit breakage |
| Cost of revenue, inference compute | Direct production compute |
| Cost of revenue, third-party model usage | Direct vendor model cost |
| R&D cloud training compute | Training and evaluation cost |
| R&D data acquisition and labeling | Dataset and annotation cost, tagged by business component |
If you want a sense of what running this kind of ledger looks like as an engagement, our bookkeeping cost estimator gives a quick read.
Controls, audit readiness, and the KPIs that matter
Strong AI bookkeeping is inseparable from control design, because the key estimates depend on operational systems. Usage logs drive revenue, cloud tags drive cost allocation, and cap-table and board records drive equity accounting. When those source systems are weak, the close is weak no matter how good the accountant is.
A generic SaaS dashboard is not enough either. The metrics that actually explain an AI business include cost per inference, compute as a percent of revenue, gross margin, contribution margin by customer or workload, training spend as a percent of engineering spend, burn sensitivity to a change in compute, deferred-revenue and prepaid-credit utilization, and the capitalized-software ratio. For the control framework and the audit file, see internal controls over AI systems for financial reporting and audit documentation for AI development costs.
The monthly close for AI companies
A close that holds up for an AI company connects commercial terms, usage telemetry, and finance controls on a repeatable cadence. The recurring steps usually look like this:
- Reconcile signed contracts to billing setup and verify usage completeness for revenue cutoff.
- Roll forward deferred revenue for prepaid credits, enterprise prepayments, and services obligations.
- Import cloud usage, apply amortized commitment views, and map tags to cost centers.
- Split compute into inference, training, internal tooling, and overhead.
- Review the internal-use software threshold, project status, and data-rights treatment.
- Update notes, SAFEs, the cap table, option grants, and board approvals.
- Update tax tagging for qualified research expenses, domestic versus foreign research, and elections.
- Reconcile intercompany balances, apply transfer-pricing markups, and post FX remeasurement.
- Produce the KPI package and archive contracts, reconciliations, and memos in a month-specific file.
Companies that wait until their first audit to define these routines usually end up revisiting months of revenue and capitalization. Building the cadence early is far cheaper than rebuilding it under deadline.
How Ridgeway Financial Services helps
Most firms treat AI companies like ordinary SaaS startups, which is usually the wrong model. A specialized bookkeeping partner should do four things well. It should make revenue reliable by translating contracts into billing logic, reconciling usage data, and maintaining deferred-revenue schedules with documented ASC 606 conclusions for nonstandard deals. It should make gross margin visible by breaking cloud and model spend down by product, customer, model, and environment, with amortized commitment views where relevant.
It should make technical accounting practical by setting capitalization, data-rights, foundation-model, and equity policies early rather than at audit time. And it should make tax and diligence easier by keeping books that already separate training compute, labeling, engineering wages, contractors, and business components, with a clean domestic-versus-foreign research split. That serves AI startups, model-enabled SaaS companies, API-first AI products, and venture-backed teams preparing for fundraising, audits, or international expansion.
To talk through your setup, explore our Accounting and Bookkeeping services or our Fractional CFO services, and contact us when you are ready.
Frequently Asked Questions
Most hosted foundation-model arrangements are service contracts rather than licensed software. Upfront payments usually create a prepaid asset for future use, and ongoing usage is expensed as incurred. Intangible-asset treatment applies only when the contract conveys a separable, identifiable right.
Under current IRS guidance, for tax years beginning after December 31, 2024, domestic research or experimental expenditures are generally deductible when paid or incurred under Section 174A, with an elective path to capitalize over not less than 60 months. Foreign research remains under Section 174 and is amortized over 15 years.
Cash paid upfront for future access is a prepaid asset amortized as services are received. A discount applied directly to usage pricing is simply a lower expense as incurred. Promotional credits usually reduce eligible cloud expense as applied and should not be recorded as assets until granted and controlled.
Yes. A generic chart of accounts hides where compute lands, how data rights are carried, and how customer prepayments are deferred. An AI-specific structure separates inference from training compute, carries licensed data and capitalized software distinctly, and isolates deferred revenue and breakage so margin and capitalization stay visible.
Reviewed by Yousuf Rizvi, CPA
Principal of Ridgeway Financial Services