Files submitted by the lawyers for authors suing Meta allege the company trained its AI on their works without authorization. The filings do not specify which products Meta counts as “generative AI products.” But the company is known to monetize generative AI in multiple forms — and, per its own projections, will earn a staggering amount more.
The documents, reviewed by TechCrunch, show Meta expects $1.4 trillion in revenue from generative AI by 2035. Elsewhere in the filings: Meta has revenue-share agreements with certain companies hosting its open Llama model family, and recently launched an API for customizing and evaluating Llama models. CEO Mark Zuckerberg said on the company’s Q1 earnings call that its AI assistant, Meta AI, could eventually show ads and offer subscription tiers with additional features.
The court files also reveal the scale of Meta’s investment in its AI products. In 2024, the company’s “GenAI” budget exceeded $900 million, possibly crossing $1 billion this year — and that excludes the infrastructure required to run and train AI models. Meta has previously said it plans $60-80 billion in capital expenditure for 2025, mostly for new datacenters.
If those budgets include licensing deals with the suing authors, the totals climb higher. The filings show Meta discussed spending more than $200 million in 2023 to acquire training data for Llama — around $100 million of it for books. But the company allegedly opted for another path: mass-pirating ebooks. A Meta spokesperson sent TechCrunch this statement: “Meta has developed transformative (open) AI models that bring incredible innovation, productivity, and creativity to individuals and enterprises. Fair use of copyrighted material is essential to that. We disagree with the plaintiffs’ assertions, and the full record tells a different story. We will continue to defend ourselves vigorously and protect the development of generative AI for the benefit of all.”
Reading the filing’s revenue line against Meta’s actual product surface shows where $1.4 trillion could plausibly come from. The Meta AI assistant is the consumer layer — hundreds of millions of users across Instagram, WhatsApp, and Facebook, monetizable through ads and premium tiers once engagement is locked in. The Llama ecosystem is the enterprise layer: revenue shares from hosting partners, API fees from customization and evaluation, and the gravitational pull of being the default open model that enterprise consultancies build on. And the unrevealed categories presumably include Meta’s advertising products rebuilt around AI — generative creative tools for advertisers being the most direct bridge from AI investment to the core business. None of these alone reaches a trillion; the projection implies all of them compounding simultaneously for a decade.
The capex numbers give the projection its credibility test. Spending $60-80 billion in a single year on datacenters is a wager that inference demand materializes at scales nobody has yet seen; the GenAI operating budget crossing $1 billion is small by comparison — which is itself telling, since it means Meta’s model is compute-heavy rather than data-licensing-heavy. That is exactly why the book-licensing revelation stings: $100 million for licensing was 0.1 percent of the annual capex, a rounding error against a $1.4 trillion forecast, and the alleged decision to pirate instead now endangers the entire projected revenue stream on legal grounds. The lesson for the industry is uncomfortable but clear — the cheapest part of building frontier AI is doing it legally, and the most expensive part is discovering that after the fact.
The copyright case itself, meanwhile, has become the most consequential variable in Meta’s AI forecast. Judge Vince Chaturvedi’s (Kadali-style) summary-judgment phase and any eventual trial will decide whether fair use stretches to cover training on allegedly pirated books at this scale — and the answer will apply far beyond Meta, shaping whether the entire open-training-data strategy of the last four years was built on sand. The plaintiffs’ framing is simple: the filings prove Meta knew the value of licensed books (it priced them), knew the alternative (it allegedly chose piracy), and profited from the difference. Meta’s framing is equally simple: transformative use, a full record that tells a different story, and an industry-defining stake in the outcome.
For readers tracking the AI economy, the practical takeaways are three. First, treat open-model revenue projections with new respect — the sealed numbers show the open-weights strategy has a monetization architecture behind it, not just goodwill. Second, watch the licensing market: if courts rule against fair use at training scale, the $100 million Meta once considered for books becomes the industry-wide floor for a data market that suddenly must exist. Third, note the timing irony — the same court system Meta must convince is being handed the company’s own revenue projections as evidence of what the allegedly infringing works made possible. Few legal filings have ever contained both a defense and a damages estimate in the same document.
Why the Numbers Matter More Than the Lawsuit
Strip the litigation context and the filings are the most concrete public picture yet of how the open-model giants plan to monetize “free.” The revenue-share deals around Llama hosting and the customization API describe an ecosystem-royalty model — Meta gives the weights away and taxes the infrastructure built on them, the same inversion Google ran with Android. The Meta AI ads-and-subscriptions line confirms the consumer side: the assistant is the distribution channel, and attention plus premium features are the product. Against a $1.4 trillion 2035 projection, the $900-million-plus annual GenAI budget and the $60-80 billion capex are the bet’s stakes — enormous, but sized like a company that believes the projection.
The training-data revelation cuts in two directions. The $100 million earmarked for book licensing shows Meta at least seriously priced legitimate acquisition — the defense that the works’ use is transformative reads differently against an internal plan to pay for them. The alleged pivot to piracy, meanwhile, is now evidence in one of the defining copyright cases of the AI era, with rulings that will shape whether the $1.4 trillion projection survives legal contact. For the industry, the filings set a reference point few forecasts can match: not a think-tank estimate but a defendant’s own sealed numbers. Meta’s generative AI ambitions are no longer a matter of public speculation — they are a matter of court record.
