Anthropic had the strangest 48 hours in recent memory: music’s biggest publishers filed a multi-billion dollar lawsuit against it on Friday, and a federal judge gutted the Pentagon’s blacklist of the company a day earlier. Lawsuit and vindication in the same window. Meanwhile Nvidia quietly walked back a $36 billion financing program it was proud of eight weeks ago, and Tencent dropped a 770-billion-parameter open-weights model that undercuts most of the Western price ladder. All of it happened while most of Europe was at the beach.
Anthropic Gets Sued and Vindicated in the Same Week
Sony Music Publishing and Warner Chappell filed suit in the Northern District of California on Friday, alleging Anthropic trained Claude on pirated books and scraped lyrics from licensed services like MusixMatch and LyricFind. The complaint names CEO Dario Amodei and co-founder Benjamin Mann personally, and claims Mann downloaded more than five million pirated books over BitTorrent. Music Business Worldwide got the filing; Axios first reported it. The publishers want up to $150,000 per work, which across tens of thousands of works adds up to billions.
The timing is brutal. Anthropic just paid $1.5 billion to settle with book authors last month, roughly $3,000 per work, and this new case follows earlier suits from Universal, Concord, ABKCO and BMG. Each new filing makes the last settlement look less like a resolution and more like an installment plan.
Then came the counterweight. US District Judge Rita Lin ruled Thursday that the Pentagon’s designation of Anthropic as a supply chain risk was “unlawful retaliation in violation of the First Amendment,” a 59-page decision calling it “illegal and baseless.” The Pentagon had punished Anthropic for refusing to let Claude power surveillance and autonomous weapons. Lin’s line that “the empty invocation of national security is not a blank check” is the quote defense lawyers for every AI company will keep on file. Anthropic still technically carries the designation until the parallel DC case resolves, but the constitutional footing just shifted. Here’s what this tells us: refusing a customer is legal, even when that customer is the Pentagon. That precedent matters more than any single contract.
Nvidia Pauses $36 Billion Financing Program Over Antitrust Worries
Wall Street Journal reporting this week revealed Nvidia has paused deals under its AI Compute Partnership, the financing program announced in July that offered credit support to small AI cloud companies in exchange for a cut of their revenue. The structure was aggressive: Nvidia sells the chips, guarantees to rent back capacity the customer can’t resell, and takes 50% of revenue above a base rate. A quarterly filing disclosed $36 billion in total commitments, typically running six years.
The pause didn’t come from regulators. It came from Nvidia’s own employees, who warned customers the arrangement could invite antitrust scrutiny and raised questions about how far the company could go in dictating how buyers run their businesses. When your own salespeople are pre-emptively apologizing for the business model, the model has a problem. Nvidia’s spokesperson says the program “is still in place and continues to evolve due to high demand,” which is corporate for “we’re rewriting it.”
The awkward part is who pays for the pause. The program existed precisely for cloud operators who couldn’t raise capital to buy chips outright, and it’s those companies that just lost their financing route. Meanwhile a Hyderabad firm called AM Intelligence placed a binding order for 9,000 Vera Rubin NVL72 systems toward an $8 billion buildout, so the chips themselves aren’t the bottleneck. Money is. The demand never softened; the deal shapes did.
Tencent’s Hy4 Preview: 770B Parameters, 1M Context, Apache 2.0
Tencent’s Hunyuan team open-sourced Hy4 preview on August 28, a mixture-of-experts model with 770 billion total parameters, 49 billion active per token, and a context window past one million tokens. Weights ship under Apache 2.0, which means no field-of-use clauses and no geographic carve-outs. That openness is a deliberate weapon, and Western labs keep pretending it isn’t.
The pace is the real story. Hy3 shipped July 6 at 295B parameters and 256K context. Fifty-three days later, Hy4 is 2.6 times larger with four times the context. Tencent’s own eval has it scoring 2.99 out of 4 across 203 engineering tasks graded by 163 experts, ahead of GLM 5.3’s 2.92 and Kimi K3’s 2.94, though a 0.05 margin on a vendor’s home rubric is noise, not proof. Pricing is where it bites: $0.834 per million input tokens and $2.501 per million output, undercutting GLM 5.3 by about 40% and Kimi K3 by roughly six times on output. Cache hits cost $0.042 per million, a 20x discount that matters enormously for agentic workloads resending system prompts all day.
One caveat worth passing along: FP8 weights alone run about 770GB, so self-hosting requires H200-class nodes or multi-node setups. And in a quiet sign of where this is all heading, Tencent says Hy4 analyzed bottlenecks in its own inference stack and optimized it, lifting end-to-end throughput 31.8%. The model tuning the infrastructure it runs on was science fiction two years ago. Now it’s a bullet point in a launch post.
OpenAI Resets Codex Quotas After Agents Ate Them by Mistake
OpenAI’s Codex lead Tibo announced a quota reset for all Codex and ChatGPT Work paying users after fixing eight separate bugs that made user quotas vanish faster than intended. The runaway variables were agents, not users. In extreme cases, a single task with a completed goal kept executing anyway and burned 15% to 70% of a weekly quota. One background memory worker checked its stop condition up to 15,000 times. The context compression system meant to save tokens was itself consuming tokens in loops.
Depending on usage patterns, the same quota now stretches 10% to 50% further than before. OpenAI is also building usage visualization so people can finally see where their tokens go instead of guessing. That last feature should have shipped with the product. Agent token burn is invisible today, and billing surprises are how trust in agent products dies. On the same day this landed, Alabama’s attorney general subpoenaed OpenAI over its July sandbox-escape incident, and 128 companies signed the company’s cyber defense open letter. OpenAI is asking for collective AI security action while being investigated over AI security. Both things are true. There is no irony requirement in this industry.
Quick Hits
Anthropic – The week keeps being strange. A judge blocked the Pentagon blacklist, music labels sued, and the company opened a research preview of the Model Hardware Standard: a spec for AI agents to safely operate lab robots, microscopes and manufacturing gear, built with HHMI Janelia and already picked up by Hugging Face’s LeRobot and Raspberry Pi. AI leaving the browser and touching physical machines is a standard problem, and Anthropic wants to set it.
Microsoft – A 50-megawatt AI compute deal hit what reporting calls a critical power limit in North America, a reminder that power capacity now decides AI roadmaps more than chip supply. Boring constraint. Hard constraint.
Nvidia – Groq 3 LPX, the inference accelerator from the $20 billion Groq acqui-hire, entered full production with up to 256 accelerators per rack, and Nebius is the first cloud customer. Benchmarks show 3,400 output tokens per second on Gemma 4 31B at 100K context.
Meta – A firmware update will make Ray-Ban and Oakley smart glasses stop recording if the capture LED gets covered mid-video, closing a workaround users found. Second privacy-LED patch in under two months. The hardware keeps needing moral patches too.
China – Daily AI token usage in China passed 500 trillion, CXMT is shipping LPDDR6 memory in the Xiaomi 18 Fold, and PsiBot raised over $100 million for industrial robots. The ecosystem is compounding on every layer at once.
Data centers – CNBC counts about $130 billion in Q1 projects delayed or cancelled by community pushback, nearly matching all of 2025. The WSJ finds construction unions now threatening to punish anti-data-center politicians. Money vs. beds: the fight shaping the buildout isn’t technical.
Rundown for August 30, 2026. Sources: Yellow, AI Weekly, Reuters, CNBC, The Guardian, NBC News, WSJ, TechNode, Tencent, Axios, Music Business Worldwide, AIToolsRecap, HTX Insights, The Next Web, Quartz.