AI companies are the hardest accounting cases in ASC 606. The business models are novel, the IP licensing structures are unusual, and most auditors have never seen an Anthropic reseller agreement or an OpenAI API contract up close. This module builds a framework: five AI business model archetypes, two deep case studies (Anthropic + OpenAI), a 6-step AI contract decision tree, and guidance on the single most mis-applied rule — the usage-based royalty exception.
Traditional software revenue recognition under ASC 606 deals with relatively stable facts: software subscriptions, professional services, support. AI companies introduce a set of facts that the standard's drafters didn't anticipate:
Every AI revenue recognition question starts with: which archetype describes this contract? The five models below cover the vast majority of AI company revenue arrangements. Most real contracts are hybrids — but decompose them into components and apply the archetype analysis to each.
Customer pays once and receives a static artifact: model weights, a trained checkpoint, or a compiled AI binary that does not change after delivery. The customer runs it themselves.
Revenue recognition: Point-in-time at delivery under ASC 606's functional license guidance (ASC 606-10-55-58). The entity's IP has "significant standalone functionality" — the customer can use it independently from the entity's ongoing activities.
Key audit question: Does the entity provide updates that significantly affect the model's utility? If yes, the license may be "symbolic" (over-time). If the model is truly static at delivery, point-in-time applies.
Customer pays a recurring fee for continuous access to AI capabilities hosted by the provider. The model is updated, improved, and maintained by the provider throughout the subscription term.
Revenue recognition: Over-time via Criterion 1 (customer simultaneously receives and consumes AI capabilities). Recognize ratably over the subscription term. This is functionally identical to SaaS subscription recognition — the AI layer doesn't change the fundamental analysis.
Common example: Monthly subscriptions to AI writing assistants, AI coding tools, AI design platforms. ChatGPT Plus ($20/month) is recognized ratably over each monthly subscription period.
Customer pays based on actual consumption: per token, per API call, per image generated, per query processed. No fixed fee — the customer buys access and pays as they consume.
Revenue recognition: As usage occurs, via Criterion 1 (over-time). Subject to the usage-based royalty exception if the API grants access to licensed IP (the AI model) and usage fees are royalties on that IP. Recognize revenue as each unit of consumption occurs — no estimates, no constraint.
Key audit question: Is the predominant item a license of IP? Most AI API arrangements involve licensing the underlying model — which triggers the royalty exception and requires usage-event-by-usage-event recognition.
Customer pays for the entity to build a custom AI system: fine-tune a model on proprietary data, develop custom ML pipelines, train a specialized model for a specific industry use case.
Revenue recognition: Depends on contract structure. Over-time via Criterion 3 if (a) the deliverable has no alternative use to the entity and (b) there is an enforceable right to payment for work completed to date. If either condition fails, point-in-time at delivery/acceptance. Over-time via Criterion 2 if the customer controls the model weights during training (rare — requires customer-controlled infrastructure).
Most common outcome: Point-in-time at delivery, because most custom AI development contracts don't include right-to-payment-for-WIP clauses. Input method (cost-to-cost or labor hours) if over-time applies.
Most enterprise AI contracts include all of the above: a platform subscription fee, usage-based API charges, implementation or fine-tuning services, and ongoing support. Each component is a separate performance obligation requiring its own recognition analysis.
Revenue recognition: Decompose the bundle into individual performance obligations. Apply the appropriate archetype to each. Allocate the total transaction price using standalone selling prices. Track separately in revenue systems.
SSP challenge: Novel AI components (e.g., a custom fine-tuning service) often have no observable market SSP. Use expected cost-plus margin or adjusted market assessment — but document your method carefully.
Anthropic distributes Claude through major cloud hyperscalers (Amazon Bedrock, Google Vertex AI, Microsoft Azure) in addition to its own direct API (api.anthropic.com). Enterprise customers can access Claude through either channel. The accounting treatment differs depending on whether Anthropic is the principal or an agent in the hyperscaler distribution.
Under ASC 606-10-55-36, an entity is the principal if it controls the specified good or service before it is transferred to the customer. The 5-indicator framework:
| Indicator | Direct API Channel | Hyperscaler Channel (e.g., Bedrock) |
|---|---|---|
| Primary responsibility for fulfillment | Anthropic — provides model, handles failures | AWS — handles availability, uptime SLAs, enterprise support |
| Inventory risk before transfer | Anthropic bears model availability risk | AWS controls compute infrastructure; Anthropic provides model weights only |
| Discretion to set price to end customer | Anthropic sets API pricing | AWS sets Bedrock pricing — Anthropic has no direct pricing control |
| Credit risk from customer | Anthropic bears credit risk directly | AWS bears credit risk; pays Anthropic a royalty regardless of end-customer payment |
| Customer relationship and billing | Anthropic invoices, supports customers directly | AWS owns the customer relationship; Anthropic receives royalty from AWS |
Conclusion: In the hyperscaler channel, the hyperscaler (AWS, Google, Microsoft) is most likely the principal. The hyperscaler controls the service before transfer to enterprise customers — it sets the price, bears credit risk, owns the customer relationship, and handles fulfillment SLAs. Anthropic is the agent receiving a royalty for licensing its model.
Revenue recognition impact:
OpenAI generates revenue from two primary sources: ChatGPT Plus consumer subscriptions ($20/month) and the OpenAI API (enterprise and developer usage-based billing). These two revenue streams require different recognition treatments under ASC 606.
ChatGPT Plus is a subscription to AI capabilities (Archetype 2). Revenue recognized ratably over each monthly subscription period via Criterion 1. The customer has continuous access to GPT-4o and premium features throughout the subscription term. Recognition pattern: $20/month, straight-line.
Performance obligations: Potentially multiple — access to GPT-4o (AI model capabilities), custom GPT creation, DALL-E image generation, voice mode. If these are distinct performance obligations with different economic characteristics, they should be identified, SSPs estimated, and transaction price allocated. In practice, most bundled AI platform subscriptions treat the bundle as a single obligation given high interdependence.
The API charges per token (input and output, varying by model). Enterprise customers pre-purchase API credits and consume as they use. This is Archetype 3 — but the royalty exception application requires careful analysis.
Is the predominant item a license of IP? The API grants customers access to OpenAI's proprietary models (GPT-4, GPT-4o, o1, o3). The API call is essentially accessing licensed AI model capabilities — arguably a royalty on IP. If the IP license is the predominant item (which it appears to be — you are paying for the model's capabilities, not OpenAI's compute infrastructure), the usage-based royalty exception applies.
Result: Recognize API revenue as each token is processed. No estimates. No constraint. Track via usage metering infrastructure and recognize revenue in real-time as consumption occurs.
OpenAI's ChatGPT Team and Enterprise plans bundle API-equivalent access, advanced features, and custom GPT deployment. How do you estimate SSP for these novel bundles?
The three methods for estimating standalone selling price (adjusted market assessment, expected cost-plus margin, residual) face unique challenges when applied to AI products. Here's what auditors will push on — and how to respond.
A customer pays $3M for Anthropic to fine-tune Claude on 10 years of proprietary legal case data. There is no market comparable for "fine-tuning a foundation model on 10TB of legal case law." Adjusted market assessment fails for the custom component.
Solution: Use expected cost-plus margin for the custom development component. Document: (a) total estimated inference compute + engineering labor cost to complete fine-tuning, (b) target gross margin for services work (typically 40-60% for AI development), (c) resulting SSP = cost / (1 - target margin).
A hyperscaler bundles AI model access (Bedrock/Vertex) with cloud compute infrastructure. The combined per-token price includes both the model royalty and the inference compute cost. How do you estimate SSP for just the model license when you can't observe it separately?
Solution: Use hyperscaler's public pricing as the "all-in" market reference. Subtract the cost of equivalent compute (from comparable cloud compute pricing) to estimate the model-layer SSP. This is adjusted market assessment applied to unbundled components.
Enterprise customers often negotiate significant volume discounts on AI APIs — 40-70% off list price for committed annual spend. The observable market price (list price) is not the price actually charged, creating a gap in adjusted market assessment.
Solution: Use the range of actual transaction prices (across all enterprise deals at comparable volumes) as the basis for SSP estimation. If the range is wide, apply variable consideration constraint principles to the high end. Consider whether committed-spend arrangements contain a financing component if payment is significantly in advance of usage.
The matrix below summarizes all five AI archetypes across four key dimensions: recognition timing, measurement basis, primary audit risk, and a real-world example.
| AI Model Type | Recognition Timing | Measurement Basis | Primary Audit Risk | Example |
|---|---|---|---|---|
| Perpetual License | Point-in-time at delivery | Full contract price at delivery event | Is the model truly "functional" (static), or does ongoing update obligation make it symbolic (over-time)? | One-time license to deploy a fine-tuned model on-premise |
| AI Subscription | Over-time, ratable | Fixed fee ÷ subscription months | Bundled distinct POs (model vs. support vs. features) that should be separated and allocated | ChatGPT Plus, Claude Pro, Gemini Advanced |
| Usage-Based API | As usage occurs (royalty exception) | Per-unit (token, call, image) as consumed | Is the IP license truly the "predominant item"? Auditors test whether exception applies vs. standard output method | OpenAI API, Anthropic API, Google Vertex AI API |
| Custom Development | Point-in-time at acceptance OR over-time (Criterion 3) | Input method (cost-to-cost or labor hours) if over-time | Missing "right to payment for WIP" clause; alternative use analysis; milestone billing vs. actual progress | Bespoke fine-tuning, custom ML pipeline development |
| Hybrid Platform | Mixed: allocate by PO | SSP-allocated transaction price per PO | SSP estimation for novel components; over-bundling POs that are distinct; royalty exception on usage component | Enterprise AI platform with subscription + API + implementation |
| Topic | IFRS 15 | ASC 606 |
|---|---|---|
| Usage-based royalty exception | Applies to licenses of IP (IFRS 15.B63). Same substance as ASC 606. | Applies to licenses of IP (ASC 606-10-55-65). More detailed implementation guidance in ASC. |
| Functional vs. symbolic license | Not explicitly defined. Use judgment on whether the entity's activities "significantly affect" the IP during the license period. Principles-based. | Explicit framework: functional license = right to use IP as it exists at grant (point-in-time). Symbolic license = access to IP as it evolves (over-time). Most practical guidance for AI model licensing. |
| Principal vs. agent | Same 5-indicator framework as ASC 606 (IFRS 15.B34-B38). No material difference. | ASC 606-10-55-36 to 55-40. Identical substance to IFRS 15. Both standards require disclosing the principal vs. agent conclusion and its revenue impact. |
| Custom AI development | Criterion 3 (no alternative use + right to payment for WIP) applies identically under IFRS 15.35(c). | ASC 606-10-25-27(c). Same two-part test; identical analysis. |
| SSP estimation for novel AI | IFRS 15.79-80: Same three methods (market, cost-plus, residual). Residual permitted if highly variable or uncertain. Less SEC enforcement pressure on IFRS filers. | ASC 606-10-32-34: Residual method under IFRS 15 and ASC 606 requires same conditions. SEC staff more aggressive in challenging residual method for AI components with any observable comparables. |
Click any card to flip it. 10 key AI revenue recognition terms — final module of the course.
Three final questions on AI revenue recognition. Choose the best answer, then read the explanation.