The accounting questions an AI business model raises
Key takeaways
What it is. The accounting questions AI companies hit that SaaS playbooks do not answer: token and usage revenue under ASC 606, where training and inference compute land on the income statement, whether GPU deals belong on the balance sheet, and what goes into gross margin when compute is the largest cost.
Where it breaks. Prepaid credits booked as revenue on sale, training costs mixed into cost of revenue, a reserved-capacity contract that turned out to be an embedded lease no one screened for, and a gross margin that falls apart when diligence rebuilds it with inference fully loaded.
How we help. We document positions on compute costs, revenue recognition, and margin presentation that hold up in diligence and audit, grounded in the standards rather than in what a template assumed.
AI companies run into accounting questions that traditional SaaS playbooks do not answer. Where do model training costs sit: R&D expense, capitalized software, or cost of revenue? How is usage-based and token-based pricing recognized under ASC 606? What belongs in gross margin when compute is your largest cost?
A quick vocabulary note, because the accounting depends on it. Training is the compute spent building a model. Inference is the compute spent running it for customers: every API call and every generated response consumes inference compute. Training generally runs through R&D expense; inference belongs in cost of revenue. Companies that blur the two report a gross margin no diligence team will accept.
These are judgment calls, and investors and auditors are looking at them closely. Corviniti documents positions on compute costs, revenue recognition, and margin presentation that hold up in diligence and audit, grounded in the standards rather than in what a template assumed.
This work is led directly by a practitioner with deep, hands-on experience across the AI sector, so the positions we document reflect how these businesses actually operate rather than a generalist template applied after the fact.
AI revenue
AI revenue recognition for tokens, credits, and usage pricing
AI pricing depends on three patterns, each with its own recognition. Prepaid credits and token packs are a contract liability until consumed, with unused credits falling to breakage. Pay-as-you-go metered usage is variable consideration, and where the bill tracks the value delivered the right-to-invoice expedient recognizes what you invoice each month. Enterprise commitments layer minimums, overage, and bundled obligations that need allocation. Before any of that, one split: hosted inference is a service, while model weights delivered to customer infrastructure are a license of IP, and the two recognize differently. The underlying software revenue mechanics live on our software revenue recognition page.
The three AI pricing patterns and their recognition. Illustrative and not exhaustive.
Cost of compute
AI compute cost accounting: training versus inference
Two terms drive most of AI cost accounting. Training is the compute spent building and improving a model before anyone pays to use it. Inference is the compute the model consumes every time a customer actually uses it: each API call, each generated answer, each image. The distinction matters because the two go to different places on the income statement. Training compute is generally R&D expense as incurred, and whether any of it can be capitalized under ASC 350-40 is a live, evolving question that needs a documented position. Inference compute is cost of revenue, and it is the single largest driver of AI gross margin, which is why diligence teams rebuild margin with inference fully loaded.
The income statement split that defines AI gross margin. Illustrative and not exhaustive.
Capacity and spend
Accounting for GPU and cloud compute spend
The full walk in one flowchart. First the arrangement: hardware bought outright capitalizes as property and equipment; a contract that hands you identified machines you control can be an embedded lease under ASC 842, putting a right-of-use asset and liability on the books; everything else is a service contract expensed as capacity is used, with prepayments sitting as assets. Then, however the compute was acquired, the income statement follows what it does: training to R&D expense, inference to cost of revenue, and customer-specific fine-tuning that builds a reusable resource potentially to an ASC 340-40 asset.
Classify the arrangement first, then follow the use. Illustrative and not exhaustive.
This is for you if
You sell prepaid credits or token packs and revenue is being recognized when the cash lands.
An auditor or investor has asked where training compute sits and why.
You are about to sign a large reserved-GPU or dedicated-cluster contract.
Diligence rebuilt your gross margin with inference fully loaded and got a different answer.
What you get
The revenue map Every SKU, credits, metered usage, enterprise commitments, mapped to an ASC 606 recognition pattern.
The compute geography Training versus inference defined, written down, and reconciled to how the clusters are actually used.
The capacity screens Every GPU and colo deal screened for embedded leases under ASC 842, before signature.
The margin definition Cost of revenue defined once, with inference fully loaded, so diligence rebuilds your number and gets your number.
The instrument memos SAFEs, converts, and stock comp classified and marked at AI-round pace.
How We Help
What we deliver
Documented positions on the questions AI business models raise, grounded in the standards rather than in what a SaaS template assumed.
The revenue mapEvery SKU, credits, metered usage, enterprise commitments, mapped to an ASC 606 recognition pattern.
The compute geographyTraining versus inference defined, written down, and reconciled to how the clusters are actually used.
The capacity screensEvery GPU and colo deal screened for embedded leases under ASC 842, before signature.
The margin definitionCost of revenue defined once, with inference fully loaded, so diligence rebuilds your number and gets your number.
The instrument memosSAFEs, converts, and stock comp classified and marked at AI-round pace.
When companies bring us in
You sell prepaid credits or token packs and revenue is being recognized when the cash lands.
An auditor or investor has asked where training compute sits and why.
You are about to sign a large reserved-GPU or dedicated-cluster contract.
Diligence rebuilt your gross margin with inference fully loaded and got a different answer.
Our Experience
Where we have done this work
Engagement Notes
A model API company recognizing credits on sale
A fast-growing inference API business was recognizing prepaid credit purchases as revenue when the cash landed. We moved the balance to a contract liability, built the consumption-based recognition off the metering data the engineering team already produced, and set a supportable breakage policy for expiring credits. Reported revenue came down, but it was now defensible, which is what an investor or auditor wants to see.
Engagement Notes
A reserved-GPU contract that was a lease
An enterprise AI company signed a multi-year deal for a dedicated cluster, named machines in a named facility, booked as a simple service expense. We ran the ASC 842 screen, concluded the contract conveyed control of identified assets, and put the right-of-use asset and liability on the books with the position memo documenting the substitution analysis. The auditors tested the conclusion and moved on; the CFO now sends us capacity deals before signature.
The Detail
The gaps, and how we close each one
Issue 01
Usage, token, and credit revenue under ASC 606ASC 606
Prepaid credits get booked as revenue when sold, metered usage gets estimated when the invoice already answers it, and enterprise deals bundle capacity, support, and SLAs into one number. AI pricing pages change monthly; revenue policies usually do not keep up.
The treatment
Cash for credits and token packs is a contract liability until consumed, with unused credits recognized as breakage in proportion to usage where supportable, otherwise only when redemption becomes remote. Metered usage is variable consideration, and where billing tracks the value delivered the right-to-invoice expedient applies. Enterprise commitments get unbundled: minimums, overage, support, and dedicated capacity each carry their own pattern. And the threshold split gets documented first: hosted inference is a service, recognized as provided, while model weights delivered to the customer are a license of IP, where the usage-based royalty exception can apply. We map every SKU to a recognition pattern at launch, not at the audit.
How we handle it: We map every SKU to a recognition pattern at launch, with credits as contract liabilities and a supportable breakage policy.
Issue 02
Compute costs: training versus inferenceASC 730 / 350-40
Inference is the compute the model burns each time a customer uses it; training is the compute that built the model in the first place. Mix the two and both numbers investors care about break: burn looks better than it is, or gross margin does.
The treatment
Training compute is generally R&D expense as incurred under ASC 730, and whether any of it can be capitalized as internal-use software under ASC 350-40 is a live, evolving question in practice; we document the position rather than leave it implied. Inference compute is cost of revenue. Depreciation on owned GPUs is split the same way the machine’s time is: R&D while it trains, cost of revenue while it serves customers. Customer-specific fine-tuning that builds a reusable resource can be a capitalizable fulfillment cost under ASC 340-40. The policy is written down, applied consistently, and reconciled to how the infrastructure team actually allocates the clusters.
How we handle it: We define the training-versus-inference geography in writing and document the ASC 350-40 position.
Launching usage pricing or signing a GPU commitment? Talk to us before the contract sets the accounting for you.
GPU purchases, cloud commitments, and embedded leasesASC 842 / 360
The same GPUs can produce three different balance sheets depending on the contract: owned hardware, a take-or-pay cloud commitment, or a dedicated cluster that is legally a lease. Most teams discover the third case after signing.
The treatment
Purchased hardware capitalizes as property and equipment, and the useful-life estimate, roughly three to six years in practice with accelerator lives genuinely debated, moves margin materially. Reserved capacity is usually a service contract: expensed as used, prepayments as assets, commitments disclosed rather than booked, and US GAAP generally records no loss on committed-but-unused capacity. But a contract that hands you identified machines you control can contain an embedded lease under ASC 842, putting a right-of-use asset and liability on the books. We run the lease screen in deal review, before signature, because retrofitting ASC 842 onto a signed capacity contract is painful. The lease mechanics live on our ASC 842 lease accounting page.
How we handle it: We screen every capacity deal for embedded leases in deal review, before signature.
Issue 04
SAFEs, converts, and stock comp on an AI cap tableASC 480 / 718
AI rounds move fast and stack large SAFEs quickly, which means the standard startup classification problems arrive at larger numbers: SAFEs that are liabilities remeasured at fair value, convertible notes with embedded features, and option grants outrunning the 409A cadence.
The treatment
The classification sequence is the same as for any venture-backed company, run at AI valuations: ASC 480 first, then ASC 815-40, equity as the residual, with liability-classified instruments marked to fair value through earnings each period. Stock compensation is priced off 409A valuations refreshed often enough to keep up with how fast AI rounds reprice. The broader startup treatment lives on our tech startups page.
How we handle it: We classify the cap table instruments through the ASC 480 and 815-40 sequence and keep the marks current.
Issue 05
Gross margin investors will rebuildDiligence
Diligence teams rebuild AI gross margin with inference fully loaded: compute, hosting, third-party model fees, and support. Companies that defined cost of revenue optimistically clear that exercise slowly, or not at all.
The treatment
We define cost of revenue once, in writing: inference compute, hosting and serving infrastructure, third-party model API fees, and support, then hold the line so margin is comparable quarter over quarter. Revenue quality gets the same discipline, separating recurring usage from one-time spikes, so the unit economics investors test are the ones you already report.
How we handle it: We define cost of revenue once, inference fully loaded, and hold the line quarter over quarter.
FAQ
Frequently asked questions
What is the difference between training and inference, and why does it matter for accounting?
Training is the compute spent building and improving a model; inference is the compute spent running it each time a customer uses it. Training is generally R&D expense, while inference belongs in cost of revenue, so the split defines both your burn and your gross margin. Investors rebuild it in diligence.
Should model training costs be capitalized or expensed?
Generally expensed as R&D under ASC 730. Whether any training compute can be capitalized as internal-use software under ASC 350-40 is a live question in practice, and the answer needs a documented, consistently applied position, because it moves both margin and burn.
How should usage-based AI revenue be recognized?
Usage-based arrangements are typically recognized as usage occurs, and the right-to-invoice expedient often applies when billing tracks delivery. Prepaid credits are a contract liability until consumed, and minimum commitments, prepaid credits, and enterprise terms each complicate the pattern. We map the contract terms to ASC 606 and document the model.
Do GPU cloud commitments go on the balance sheet?
Usually not: take-or-pay capacity commitments are typically service contracts, expensed as used and disclosed as commitments. The exception is a contract that gives you identified machines you effectively control, which can be an embedded lease under ASC 842 and does go on the balance sheet. Screen before signing.
What do AI investors focus on in accounting diligence?
Gross margin definition and revenue quality. Diligence teams rebuild margin with inference compute fully loaded and test whether usage revenue is recurring. Companies with documented, consistent positions clear this quickly.
Sources & authorities
Primary sources for this page
Research costs, capitalized.IRC Section 174: capitalizing and amortizing research and experimental costs, including much software development.
Research credit.IRC Section 41: the credit for increasing research activities.
Compute and software costs. ASC 350-40 on internal-use software, which governs much of the model-development and infrastructure spend.
Revenue. ASC 606 applied to usage-based, subscription, and hybrid pricing.
This page summarizes federal tax and accounting rules for general information, and is not tax or accounting advice. Rules change; confirm the current text before you rely on it.
