Happy Thursday. A few things worth knowing right now. Bitcoin touched 76,000 dollars on Wednesday before sellers pushed it back under 74,000, and the options market is still leaning bearish even as the S&P 500 prints new all-time highs. Morgan Stanley's brand new Bitcoin ETF just crossed 100 million dollars in its first week, charging the lowest fee in the game at 0.14%. Tether hired KPMG for a full audit of its USDT reserves, a move that could reshape how regulators view stablecoins. And a group of authors led by Pulitzer winner John Carreyrou is suing 6 major AI companies, claiming they trained models on pirated books. We've got a lot to unpack. Let's get into it.
The legal war over what AI companies can and cannot train on is getting louder and more complicated. Let's start with the biggest new filing. A group of authors led by John Carreyrou, the journalist behind Bad Blood, has sued Anthropic, OpenAI, Google, Meta, xAI, and Perplexity. The core allegation is that all 6 companies trained large language models on pirated books pulled from shadow libraries like LibGen, Z-Library, and OceanofPDF. No licenses, no payments. The plaintiffs are citing potential damages of around 150,000 dollars per work under the Copyright Act, which across all 6 defendants could mean roughly 900,000 dollars for a single book. They're arguing for billions in total. And they're explicitly rejecting the proposed 1.5 billion dollar Anthropic settlement, which worked out to about 3,000 per title for around 500,000 books. The authors say that's a fraction of what they're owed. This case builds on a June 2025 ruling in Bartz versus Anthropic, which suggested training on legally acquired books might qualify as fair use, but drew a much sharper line against material obtained from pirate sources. That distinction is now the fulcrum of this lawsuit. Meanwhile, 3 YouTube creators just filed a class action against Apple in California, alleging Apple scraped YouTube videos to train AI models using the Panda-70M dataset. That's 70 million video-caption pairs from about 3.8 million YouTube videos. Similar suits are already pending against Meta, Nvidia, ByteDance, and Snap. And then there's the Anthropic code leak situation. Back in March, sensitive harness code connecting Claude to apps and data sources leaked online. Anthropic responded by issuing a DMCA takedown on GitHub. Critics immediately pointed out the irony: the same company that argues copyright shouldn't block AI training is now using copyright to suppress its own leaked code. It's a contradiction that sharpens the whole debate. Over in India, there's the ANI versus OpenAI case, the country's first major AI copyright litigation. The court is looking at 2 issues: whether real-time web retrieval through retrieval augmented generation infringes copyright, and whether LLMs that memorize and reproduce training data are liable. The judgment is reserved, but if the court rules that RAG is non-expressive and that memorization liability falls primarily on users who extract and publish content rather than on developers with guardrails in place, it would be a meaningful precedent. Not just for India, but globally. What's clear is that the legal framework for AI training is being built in courtrooms right now, case by case, country by country. And the outcomes will shape what AI companies can build and how much it costs them.
Bitcoin has had a solid month. It's up about 10% over the past few weeks. But the rally has stalled right around 75,000 dollars, and the divergence with traditional equities is getting hard to ignore. The S&P 500 just closed at a new all-time high of 7,022. The Nasdaq hit 24,016, its 11th consecutive up session. Tech stocks are on a tear. And yet Bitcoin pushed above 76,000 dollars briefly on Wednesday before getting knocked back below 74,000. CryptoQuant data shows why. The price is testing a level that capped the rally back in January. Large holders are positioning to sell near a key breakeven zone. Short-term traders cashed out about 63,000 BTC in profit over the past 24 hours as the price touched 76,000. That's a meaningful wave of selling pressure. The on-chain picture is mixed. Funding rates have hit their most negative level since 2023, which historically has coincided with local bottoms. So the contrarian read would be that the worst of the selling is close to done. But the options market isn't buying it. QCP Capital notes that derivatives desks still want downside protection. The put skew on BTC options has eased but remains tilted bearish. Long-end bond yields and gold aren't confirming the risk-on move either. On the institutional side, the flows are strong. Spot Bitcoin ETFs pulled in about 411 million dollars on Tuesday, the second-largest daily inflow this month. BlackRock's IBIT led with roughly 214 million. ARK 21Shares added 113 million. And Morgan Stanley's brand new MSBT fund just crossed 100 million dollars in assets after only 6 trading days, charging a rock-bottom 0.14% fee. That's already enough to overtake WisdomTree's Bitcoin fund. Goldman Sachs just filed to launch its own Bitcoin ETF using covered call strategies to generate income, targeting advisers who want yield rather than pure price exposure. If these income-oriented products attract serious institutional capital, they could actually compress Bitcoin's volatility over time by layering options strategies across the market. The leverage picture adds another wrinkle. About 6 billion dollars in leveraged positions sit in the 72,000 to 73,500 dollar zone. If prices break lower, forced liquidations could cascade. If they bounce, short squeezes could fuel a move higher. It's a coiled spring. The macro backdrop is supportive. The US-Iran ceasefire is holding. Risk appetite is up. But Bitcoin is range-bound while stocks fly, and the key question is whether sustained ETF inflows can push through the supply wall at 75,000 to 76,000 dollars.
The stablecoin landscape is shifting fast, and this week brought a cluster of moves that tell a bigger story about where the market is heading. Start with Tether. The company has hired KPMG to conduct a full audit of its USDT reserves and brought in PwC to strengthen internal controls. This is not another attestation or quarterly snapshot. This is a Big Four audit. For a company that spent years fending off transparency criticisms, it's a significant step. The timing aligns with the evolving US regulatory picture. The GENIUS Act is moving through Congress, and JPMorgan analysts say the CLARITY Act is close to a final deal. Tether appears to be positioning itself for a more formalized US stablecoin regime where credible audits aren't optional, they're table stakes. Then there's the Drift Protocol situation. After an exploit earlier this month drained more than 270 million dollars from the Solana-based perpetuals exchange, Tether stepped in with partners to provide about 148 million dollars in recovery funding. The twist: Drift is relaunching as a USDT-based platform, dropping Circle's USDC. That move is directly connected to a brewing feud between Circle and its critics. Circle CEO Jeremy Allaire has been firm that USDC freezes require formal legal orders. He won't freeze funds based on informal requests during active exploits. On-chain investigator ZachXBT and others have argued this approach has allowed over 420 million dollars in stolen funds to escape since 2022. Tether, by contrast, has been more aggressive about freezing tainted funds quickly, sometimes within hours. Drift's decision to switch to USDT after the exploit is a market vote on that debate. Circle isn't sitting still, though. The company is making a major push into South Korea, where Tether currently controls over 80% of stablecoin trading. Circle CEO Allaire visited Seoul, signed agreements with major exchanges Upbit, Bithumb, and Coinone, and partnered with fintech Hecto Financial for payments infrastructure. Circle's pitch is that USDC already handles roughly 17 trillion dollars in annual on-chain volume, much of it in payments and institutional finance, compared to Tether's roughly 13 trillion. The Korea push is about real-world payment use cases, not just trading pairs. Allaire also made headlines by predicting China could launch a yuan-denominated stablecoin within 3 to 5 years, framing it as an escalation of global currency competition. Meanwhile, Tether backed a 134 million dollar funding round for the Stablecoin Development Corporation, a publicly traded firm building infrastructure to move stablecoins more efficiently across platforms and borders. Total stablecoin circulation now exceeds 300 billion dollars, with 33 trillion in transaction volume last year. That's bigger than Visa and Mastercard combined. The infrastructure race is on.
If you follow AI capabilities, you know the models keep getting better. But here's the uncomfortable question: are we actually measuring whether they're getting safer? Stanford's 2026 AI Index Report suggests the answer is no. The report finds that responsible AI benchmarks covering safety, fairness, security, and factuality are largely missing from frontier model evaluations. Most top models have empty results across these benchmarks. Only a handful, like Claude Opus 4.5 and GPT-5.2, report even a limited set. Red-teaming and alignment work happen behind closed doors, but there's almost no standardized, externally comparable disclosure. You can compare models on capability all day long, but comparing them on safety is nearly impossible. The report also closes the book on one popular assumption: the idea that the US holds a permanent lead in AI performance. It doesn't. China now trades top performance with the US, leads in publication volume, citations, and patents, while the US still produces more top-tier models. The strategic vulnerability Stanford highlights isn't about code. It's about hardware. The global AI supply chain runs through a single foundry, TSMC in Taiwan. The UK government just published an open letter to business leaders warning that AI-enabled cyber threats are accelerating faster than expected. They cite tests showing frontier models can now perform offensive cyber tasks faster and at greater scale than before, with capabilities doubling roughly every 4 months. IBM responded this week by announcing new cybersecurity measures specifically designed for what they call agentic attacks, where AI systems operate autonomously at machine speed. Their new offering includes multi-agent, vendor-agnostic security services that coordinate AI-powered defenses across an organization's stack with minimal human input. On the policy front, OpenAI and Anthropic are clashing over Illinois Senate Bill 3444, the Artificial Intelligence Safety Act. OpenAI supports it. The bill would largely shield AI labs from liability if they meet safety protocols and publish transparency reports. Anthropic opposes it as insufficiently protective. It's a microcosm of the broader tension in AI governance. Do you move fast with industry-friendly guardrails or push for stronger, more prescriptive accountability? The EU is taking the latter approach. The Cambridge Commentary on the EU General-Purpose AI Law just launched as the first scholarly deep dive into the AI Act's provisions on general-purpose AI models. It maps the interpretive landscape rather than pushing a single conclusion, which is probably the honest approach given how much ambiguity exists in the law. The bottom line: AI models are racing ahead while safety evaluation, regulation, and standardized disclosure lag behind. The gap is widening.
Here's the thing to sit with. Whether it's copyright law, AI safety benchmarks, or stablecoin audits, the common thread this week is accountability catching up to ambition. The technology moves fast, the legal and governance infrastructure moves slow, and right now we're in the messy middle where the rules are being written in real time. If you're building, investing, or just paying attention, the question isn't whether guardrails are coming. It's whether they'll arrive before the next big breakage. That's it for today. I'll talk to you tomorrow.