CuriosityStream Q2 2026 press release discloses $14.1M licensing revenue driven by AI training partnerships
CuriosityStream disclosed $14.1M in Q2 2026 licensing revenue on August 12, up 48% year-on-year at 72.8% gross margin, selling rights-cleared video, audio and an 880B-token private code corpus to AI labs. Full-year guidance was raised to $77-82M. The same factual-media archive that earned inside an SVOD subscription is now priced directly against a foundation-model buyer.
Where the rake sits
CuriosityStream keeps the licensing fee at roughly 73% gross margin and holds the AI-lab buyer relationship directly because it does the rights-clearing, curation and dataset construction itself rather than routing through a broker. There is no disclosed downstream claim on the models the corpus helps train, so whatever the weights permanently absorb accrues to the labs; whether the flat licence under-prices the open-ended training right is not determinable from the release.
What happened
- 2026-08-12: CuriosityStream (Nasdaq: CURI) disclosed Q2 2026 licensing revenue of $14.1M, up 48% year-on-year, described in the SEC-filed release as a record driven by content and data-corpus licensing to AI partners.
- 2026-08-12: Gross margin expanded to 72.8% in Q2 2026 versus 53.4% in Q2 2025.
- 2026-08-12: Record Q2 net income of $8.9M (up 1,033%) and Adjusted EBITDA of $11.4M reported; full-year revenue guidance raised to $77-82M.
- 2026-08-12: Licensing sits on a corpus of 3M+ hours of rights-cleared video and audio plus an 880 billion token private code corpus; the library is professionally produced factual video with narration and metadata rather than subscriber data.
- 2026-08-12: Management framed the licensing opportunity as three pillars: video licensing to traditional media; audio and video licensing for AI training; and private code licensing for AI training.
- 2026-08-12: Specific Q2 deals included licensing millions of tokens of code for AI training, RL and evaluation, a 40,000-segment HDR dataset, and synchronized multi-camera action sequences to a video research lab.
- 2026-08-12: AI-partner counterparties and per-model or per-inference terms were not disclosed; the reported figure is a licensing fee with no disclosed downstream participation.
- For historical context: CuriosityStream's FY2020 total revenue was $39.6M and it reported roughly 20M subscribers in 2021, making the Q2 2026 licensing line alone a material fraction of the historical annual base.
Who is involved
US factual-media streaming company that has repositioned as a provider of AI model training datasets leveraging a rights-cleared media corpus spanning millions of hours of premium video and audio, 850 billion tokens of production-grade code, and bespoke datasets
Nasdaq: CURI; founded 2015; headquartered Silver Spring, Maryland; Q2 FY2026 licensing revenue of $14.1M per the cited press release
The reading
The enrichment work - assembling the content library, clearing rights, and productising it into licensable datasets - sits with CuriosityStream; the Q2 call describes 17 off-the-shelf video data set products and an 880 billion token private code corpus packaged for buyers, and a specific 40,000-segment HDR data set licensed for video generation model training. The AI labs that pay then do their own training on top.
The substrate's worth to AI-lab buyers is as training material whose rights are cleared and whose provenance is known - Stinchcomb's framing is of a 'critical supplier of trusted real-world data assets' - feeding the buyer's decision of which checkpoints to train and ship. The public filings do not name individual lab counterparties, so the specific decision each buyer makes is not disclosed.
What crossed are licence rights to defined slices of the corpus (e.g. the 40,000-segment HDR set, portions of the 3M-hour library, pieces of the 880B-token code corpus) under what management describes as increasingly recurring enterprise agreements. The underlying library and the rights relationships behind it did not cross; counterparty names and per-deal terms are not disclosed.
Not determinable from the Q2 release. Management flags licensing as 'chunky' and H2 2026 adjusted EBITDA guidance of $6-10M steps down from Q2's $11.4M, which bears on timing rather than on whether individual licences were under-priced against what buyers extract.
Why it matters
CuriosityStream is a clean public example of a company whose substrate already earned inside a wrap (the SVOD subscription) and that deliberately built a sell line on top of the same asset — with rights-clearing and curation done in-house so the rake stays owned rather than shared with a data broker. The interesting move is that the same underlying asset (factual video/audio) is now priced against a foundation-model buyer's P&L rather than a consumer subscription P&L, and the code corpus was assembled specifically for that buyer — suggesting the substrate was extended, not just repackaged. Worth using to illustrate how enrichability for an AI-lab buyer (rights-cleared, domain-structured, production-grade) creates a price surface a scraped-web substitute cannot match.
The argument this deal tests: Data monetisation: three routes to value creation
Related deals
- Chegg pivots to licensing decade of proprietary STEM academic content and expert network to AI labs2026-05-13
- Google wins $10M bankruptcy auction for Spirit Airlines' enterprise data and code to train AI models2026-08-14
- Universal Music Group licenses catalogue to ElevenLabs for multi-year AI music platform collaboration2026-09-10
- Reddit's capital-light data model: $1M capex against $300M+ cash flow demonstrates non-consumption and non-rivalry2026-04-30
Sources
- nasdaq.com — primary
- sec.gov — corroborating
- en.wikipedia.org — party background
- sec.gov — party background
Announced 2026-08-12