THE AI ARBITRAGE

A Federal Judge Just Priced AI's Copyright Risk at $1.5 Billion

Anthropic's landmark settlement is final now. What it actually resolves, and what it very much doesn't, matters more to every AI-using business than the headline number does.

On Monday, July 20, U.S. District Judge Araceli Martinez-Olguin granted final approval of Anthropic's $1.5 billion settlement with a class of authors and publishers, closing out what's believed to be the largest copyright settlement in U.S. legal history. The deal pays roughly $3,000 per affected work, TechCrunch reports the class covers an estimated 500,000 titles, while Dataconomy puts the approved scope at more than 480,000 works, and it requires Anthropic to destroy the pirated copies it used to build Claude's training library.

The case traces back to a 2024 lawsuit accusing Anthropic of building its training library two ways: books it purchased and scanned, which the original presiding judge, William Alsup, ruled counted as fair use for AI training, and roughly 7 million books downloaded from pirate sites like Library Genesis, which he ruled was straightforward infringement no matter what happened afterward. Facing a jury trial on damages for the piracy claim alone, with exposure that could have run into the billions, Anthropic settled in 2025. Alsup has since retired; Judge Martinez-Olguin inherited the case and signed off on the final terms this week.

That distinction is the one to sit with. Alsup's fair-use ruling was persuasive, not binding, and it applies to one judge's read of one company's facts. Similar suits are still working through the courts against Google, Meta, Midjourney, and OpenAI, including a fresh one filed against Google last week by a group including Hachette, Cengage, Elsevier, and author Scott Turow, per TechCrunch. Every one of those cases could land differently, on fair use, on damages, or both.

The read for operators: downstream liability doesn't stop at the model builder. A business that uses an AI tool trained on infringing data can carry its own exposure, separate from whatever the vendor eventually pays. Before this issue fades from the news cycle, it's worth asking every AI vendor a plain question: where did your training data come from, and does your contract indemnify us if that answer turns out to be wrong.

Compute Scarcity Is Becoming the Real AI Tax

Two stories from the same week show what happens when demand outruns GPU supply. Even the best-funded labs pay up or lock out users. That's a preview of what AI vendor risk looks like for everyone downstream.

Meta is in talks to lease Anthropic up to $10 billion in AI computing capacity over two years, paid in monthly installments with an early-exit option for either side, according to a New York Times report cited by multiple outlets. Anthropic's web traffic share has reportedly nearly doubled since March, and the company is already paying an estimated $45 billion over three years under a separate compute deal signed earlier this year with SpaceX. Even a company approaching a $965 billion valuation can't simply buy its way out of a GPU shortage. It has to negotiate capacity contracts the same way an SMB negotiates warehouse space or freight capacity: ahead of need, and at a price set by whoever has the scarce resource.

The same dynamic hit a smaller player days earlier. Moonshot AI's Kimi K3, a 2.8 trillion-parameter open-weight model released July 16, drew so much demand that the company paused new consumer sign-ups on July 19, just three days after launch. "To protect the experience of existing subscribers, we're temporarily pausing new subscriptions and prioritizing compute for current members," Moonshot said. Existing subscribers were unaffected, but the model positioned as the value alternative to Fable 5 and GPT-5.6 Sol is now capacity-constrained too, and full open-weight release, which could ease that pressure by letting others self-host, is still targeted for July 27.

The read for operators: this is exactly the kind of single-vendor exposure our token and model routing cost governance work is built to catch before it becomes an outage or a surprise invoice, not after.

Brussels Tells Google: Open Up Android's AI Doors

The EU's latest Digital Markets Act order doesn't just target Google's dominance. It's designed to seed a market of AI assistant alternatives on the world's biggest mobile platform.

The European Commission issued two binding decisions on July 16 under the Digital Markets Act. The first requires Google to let rival AI assistants use 11 Android features, including voice activation, booking actions like taxis, and generating reply suggestions in chat apps, changes that must ship with Android 18 by August 1, 2027. The second requires Google to share anonymized search query, click, ranking, and view data with eligible competing search engines and AI chatbots starting January 2027, under a multi-layered anonymization framework. The decisions carry no fine today, but noncompliance risks penalties of up to 10% of Google's global annual turnover. Google's Android president, Sameer Samat, publicly pushed back, calling the Commission "on the wrong track" and raising privacy and security concerns.

The read for operators: nothing here changes a U.S. SMB's AI stack this quarter, the deadlines run through mid-2027. But it's an early signal that the Android AI assistant market won't stay a one-vendor default forever. Worth a note on the calendar, not a reaction today.