Records / Negotiation levers

The labs list four inputs, and only one is bought from a holder: sell into that one

The labs' own disclosures name where training data comes from. Anthropic: "publicly available information on the Internet", "non-public data from third parties", "data provided by data-labeling services and paid contractors", opted-in users and "data we generated internally". Apple: data "licensed from publishers", public or open datasets and pages "crawled by our web-crawler, Applebot". Meta: Llama 3 on "over 15 trillion tokens of data from publicly available sources", Llama 4 on "publicly available, licensed data" and its own products. Microsoft's phi-4 "strategically incorporates synthetic data throughout the training process". Public pages they crawl themselves and generated data they make themselves; what they take from outside is non-public or licensed data and paid labour. OpenAI's own data acquisition team "manages web crawling and GPTBot services".

Dated 22 Oct 2026Source checked 22 Oct 2026Quote found on its page again 28 Oct 2026Basis: strong