$1.5 billion.

On July 20, a U.S. federal court in San Francisco formally approved the copyright settlement between Anthropic and a group of authors — the largest class-action copyright settlement in U.S. history. Half a million copyrighted works, roughly $3,000 per work, totaling $1.5 billion (~10.2 billion yuan). The judge signing off changed from Alsup to Martinez-Olguin, but the core logic remains: AI training itself is fair use, but pirating the data you train on is illegal.

One sentence sums it up: Bought books are fine to train on. Stolen ones aren't.

This seemingly simple conclusion could reshape the entire AI industry's data supply chain. Over the past few years, all LLM companies have been racing through the same gray area — scraping the entire web, downloading pirated e-books, hoovering up social media data, with no one knowing where the legal red line was. Now the first red line has been drawn.

The Core of the Ruling: Fair Use, but Illegal Acquisition

The most fascinating part of this case is that both sides won — sort of.

In his preliminary ruling last year, Judge Alsup sided with Anthropic on the most critical question: he ruled that training AI models on copyrighted text counts as fair use. This is the answer the AI industry wanted most: training itself isn't infringement. If training were infringement, the foundation of every major LLM company would collapse — OpenAI, Google, Meta, Anthropic would all get sued into bankruptcy.

But Anthropic celebrated too early. Judge Alsup also found that how Anthropic obtained those books was problematic — its training corpus came from two sources: books it legally purchased and scanned, and books it downloaded from pirate sites like Library Genesis and Pirate Library Mirror. The second channel is illegal on its own terms.

Here's the analogy: if you borrow a book from the library, read it, and remember what you learned, that's not stealing — that's learning, fair use. But if you steal a book to read it, the act of stealing is itself a crime, regardless of what you do with the knowledge afterward.

Judge Alsup was going to let a jury decide the "pirated acquisition" question, and Anthropic rushed to settle — because no one knows how much a jury might award in damages. $1.5 billion is a staggering sum, but it beats a jury slapping you with $50 billion or $100 billion out of anger.

So the real meaning of this ruling is: the legality of AI training is confirmed, but the compliance cost of data acquisition just went up. From now on, LLM companies can't just scrape books from pirate sites willy-nilly. They either pay for licenses or use only legally available data sources.

Is $1.5 Billion Actually Expensive?

$1.5 billion sounds like an astronomical number. But do the math, and it might be "cheaper" than you think.

500,000 works, $3,000 per work. For a bestseller, $3,000 might not even cover a single year's royalties. But for the vast majority of ordinary writers, a book might never earn $3,000 in its entire lifetime. Many niche authors, academics, and poets — whose books might have sold only a few hundred copies total — are suddenly getting $3,000 in compensation. It's like money falling from the sky.

And for Anthropic? What's $1.5 billion? Its most recent funding round valued the company at over $60 billion; $1.5 billion is about 2.5% of that. Put another way: Anthropic just signed a $19 billion, 20-year compute deal with TeraWulf — the $1.5 billion settlement is less than one-tenth of a single compute contract.

More importantly, this settlement didn't set a precedent. Because it's a settlement, not an appeals court ruling, it's not binding on other cases. Google, Meta, and OpenAI are still fighting their own copyright battles, and every judge can reach their own conclusions. Just last week, major publishers including Hachette, Cengage, and Elsevier, along with author Scott Turow, filed a class-action lawsuit against Google, alleging the company used their copyrighted works to train Gemini.

So is $1.5 billion expensive or cheap? It depends on your perspective. From the writers' angle, $3,000 per book is certainly better than getting nothing, but it's probably a drop in the bucket compared to the value AI companies extract from their content. From Anthropic's angle, $1.5 billion buys a vague confirmation that training is legal and avoids a potentially astronomical jury award — not a bad deal at all.

An Unexpected Win for Chinese Model Companies

After this ruling, the ones secretly smiling might not be Anthropic — they might be China's LLM companies.

Why? Because Chinese LLM companies have faced stricter data compliance requirements from day one. China's regulatory environment means domestic LLM companies have always been relatively "disciplined" about training data acquisition — you can't just scrape everything, you can't use whatever you want, and copyright issues have been a sword of Damocles from the start. The result? Chinese LLM companies might actually have lower data compliance costs than their American counterparts — because they followed the rules from the beginning and never accumulated massive legal risk.

Right as this ruling dropped, another story is worth noting: The Verge published a long piece headlined "China delivers a one-two punch to America's AI dominance." The article argues that Moonshot's Kimi K3 and Alibaba's Qwen are pincering American AI companies from two directions — one hitting the high end with a closed-source model approaching GPT-5 and Claude performance, the other hitting the ecosystem with a fully open-source model. Together, they're rapidly narrowing America's AI lead.

On prediction market Polymarket, the probability of Anthropic reaching a $1.5 trillion valuation by year's end has dropped sharply, with traders directly linking the shift to Kimi K3's release. Meanwhile, Hong Kong-listed Zhipu AI surged over 30% in afternoon trading — the company launched a 1GW domestic compute center and completed its acquisition of Zhongke Jiahe.

The deeper impact of the copyright ruling is this: if American AI companies have to pay higher compliance costs for training data while Chinese companies are already lean on data compliance, the cost gap between them narrows further. Everyone used to assume American AI companies were "faster," but the people running in front are also carrying more legal landmines. As those mines start going off, the people behind might just catch up.

And let's not forget: China has the world's largest Chinese-language corpus and a unique regulatory framework. China's generative AI management measures were introduced early, with relatively clear boundaries for data compliance. From this perspective, markets that regulate first might benefit first — like how places that build roads first get traffic first.

Endgame: The Era of Data Costs Has Arrived

Put the $1.5 billion copyright settlement, the rise of Chinese models, and the evolution of the MCP protocol together, and a clear inflection point emerges: AI competition is shifting from "who has the bigger model" to "who has cleaner data, more controllable costs, and a better ecosystem."

2023 was about parameter count. 2024 was about performance. 2025 was about price. What's 2026 about? Compliance, data quality, infrastructure.

This ruling will have several profound effects:

First, the price of high-quality copyrighted data will skyrocket. Previously, publishers' content was taken by AI companies for free. Now, with the $1.5 billion precedent, publishers will grow a spine — and licensing fees will definitely go up. The content controlled by the Big Four publishers (Penguin Random House, HarperCollins, Simon & Schuster, Hachette) will become the "oil drilling rights" of the AI era. Building a high-quality model will first require negotiating licenses with these publishers — which circles back to the old rule: he who has the money calls the shots.

Second, the relative advantage of open-source models will grow. Closed-source model companies need to pay massive costs for training data, but the open-source community naturally has more legal data sources — code voluntarily contributed by developers, open-access academic papers, Creative Commons-licensed content, and more. If closed-source costs keep rising due to copyright issues while open-source performance keeps approaching, what happens to the market landscape? The answer speaks for itself.

Third, synthetic data will become a serious field. If real-world data has copyright problems, what about training AI with AI-generated data? That's the synthetic data approach. There's already plenty of research exploring using LLMs to generate training data for next-generation models. If this path works, copyright won't be the bottleneck anymore — AI will grow its own food. Of course, synthetic data has its own problems, like model collapse and lack of diversity, but compared to billions in copyright fees, these technical problems are cheap.

Fourth, global AI regulation enters the "detail-sorting" phase. South Korea's AI Framework Act takes effect today (July 21). The EU AI Act went fully into effect on July 1. China's generative AI management measures have been in place for a while. The US is still figuring out boundaries through lawsuits. Different regions have different regulatory frameworks and different compliance costs — which itself becomes a form of competitiveness. Wherever regulation is clear, predictable, and reasonably priced, that's where the AI industry will cluster.

The $1.5 billion settlement isn't the end — it's just the beginning. It's the first landmark event marking AI industry's shift from "wild growth" to "compliant development." There will be more copyright lawsuits, more regulatory policies, more compliance costs. AI companies can't just stare at model performance and funding valuations anymore. The legitimacy of the data supply chain is becoming the next core competitive advantage.

After all, in an era where data is oil, you first need to make sure the oil you're using — isn't stolen.

See you tomorrow.