Prepared for Mistral AI. As the data licensing portfolio scales beyond the AFP grounding agreement, every new data access agreement triggers the same ASC 350-30 question. The Art. 53(1)(d) summary you published already references those agreements. One register feeds both.
Under ASC 350-30, the classification of a licensed data asset turns on one question: does the license grant a finite term or a perpetual right? Term-based is expensed over the term. Perpetual capitalizes as a separate intangible. The ratio across the portfolio determines how much of the data spend is a COGS line versus a balance-sheet asset.
| License Type | Accounting Treatment | P&L / Balance Sheet |
|---|---|---|
| Term-based12-mo, 3-yr, renewable | Expensed over the license term | COGS |
| PerpetualNo expiry, or renewal at nominal cost | Capitalized as identifiable intangible asset | Balance Sheet |
| Acquired via M&AIP assumed with the company | Capitalized at fair value as of acquisition date | Balance Sheet |
| Unlicensed / disputedNo executed agreement | No asset recognized. Legal liability if applicable. | Off-BS Risk |
The "alternative future use" test (ASC 350-30-25): if purchased data can be used beyond the immediate training run (e.g., fine-tuning, RAG, product delivery), it qualifies for capitalization even if the license is term-based. In practice, most AI training data has alternative future use — which means the term/perpetual distinction is the only variable that determines the classification.
A perpetual license may still be finite-lived if technological obsolescence or regulatory change limits its useful life. The useful life determination must be reassessed each reporting period.
| Source Category | Typical Structure | ASC 350-30 | Art. 53(1)(d) |
|---|---|---|---|
| Grounding / RAG feedsAFP, news agencies, APIs | Multi-year term, for retrieval not training | Expensed over term. No capitalization. | Disclose as "licensed data source" with category, size, date. |
| Training data (licensed)Rights holders, publishers, universities | Term-based or perpetual, negotiated per source | Term → expense. Perpetual → capitalize. | Disclose as "data access agreements" with category and size. |
| Web-crawled / scrapedCommon Crawl, public web | No license (or TDM opt-out under Art. 4(3) DSM) | No asset. Compliance cost only. | Disclose crawler identifiers, purposes, content types, top 10% domains. |
| Enterprise co-creationASML, Stellantis, industry partners | Co-development agreements. Data stays with the enterprise. Access, not ownership. | No asset on Mistral's books. Service revenue, not data asset. | Disclose as "partnership data" with category and scope. |
| Synthetic / internally generatedMistral model outputs, augmentation | Internally generated | No capitalization (ASC 350-30-25-3). | Disclose as "synthetic" category. |
| Acquired via M&AEmmi AI (May 2026) | Perpetual IP, assumed with the company | Capitalized at fair value. Indefinite-lived if no foreseeable expiry. | Disclose as "acquired data" with category and size. |
| Data Source | What It Is | License Type | Status | Open Question |
|---|---|---|---|---|
| AFPJan 2025 | Real-time feed + archive back to 1983 → Le Chat grounding. For grounding, not training. | Multi-year term. Grounding-only. | Expensed | None. Classification is clear. |
| Training data access agreementsPer Art. 53 summary, July 2025 | "Data access agreements with rights holders or their representatives for access to non-publicly available datasets." | Undisclosed. Term or perpetual? | Unresolved | What are the terms of each agreement? What is the term/perpetual ratio? |
| Common Crawl / webOngoing | Publicly accessible text, images, code, metadata for model training. | No license. TDM under Art. 4(3) DSM. | No Asset | None. Compliance cost only. |
| Enterprise co-creationASML (Sept 2025), Stellantis (Oct 2025) | Proprietary manufacturing data for post-training / fine-tuning. Data stays with the enterprise. | Co-development. Access, not ownership. | No Asset | None. Service revenue, not data asset. |
| Emmi AIMay 19, 2026 | Industrial engineering simulation models. | Perpetual IP (assumed via M&A). | Capitalized | What was the allocated fair value in the PPA? |
The data licensing function is in build mode. As the training-data portfolio scales beyond the AFP grounding agreement — publishers, universities, research institutions — every new data access agreement triggers the same classification. The term/perpetual ratio across the growing portfolio will determine the balance-sheet position. Building the register before the agreements are signed is cheaper than retroactively classifying them after.
The Art. 53(1)(d) summary published for Mistral Large 3 (July 2025) already references "data access agreements with rights holders or their representatives for access to non-publicly available datasets." The obligation is met. The register behind that summary is the same register the ASC 350-30 classification requires.
The key insight: every data source you document for the Art. 53 template (category, license type, size, acquisition date) is the same data you need to determine whether it's a COGS line or a balance-sheet asset. Build the register once. Feed both obligations. Keep it current as the portfolio scales.
The AI Office's enforcement powers (including fines) began August 2, 2026 — already in effect. Models placed on the EU market before August 2, 2025 have until August 2, 2027. The Art. 53 summary must be updated for each new model release. As the data licensing portfolio grows, the register behind the summary must grow with it.