Today's Headlines

  • Thirty-five Sony and Warner music publishers sue Anthropic and two co-founders for copyright infringement — a complaint has been filed, and the court has yet to rule
  • Kioxia and SanDisk expect to invest about ¥5 trillion in Japan by 2032, on the premise of government support
  • The AI cloud company Lambda raises $1 billion in private short-dated debt to buy Nvidia chips it will lease to Microsoft
  • Of 120 open source projects surveyed, 37 bar AI-assisted contributions outright
  • Nvidia's advantage is read as moving from the GPU itself to the shape of the whole data centre

Today's three stories turn on a single question: who pays, and how much, for the things that hold AI up. The first answer concerns the price of the work used in training. Thirty-five music publishers filed a copyright complaint against Anthropic and two of its co-founders.

The second concerns the factories that make the memory. Kioxia and SanDisk said they expect to invest about ¥5 trillion by 2032, on the premise of government support. The third concerns the debt raised to buy chips. Lambda took on $1 billion in private short-dated debt to buy Nvidia silicon.

The recipient keeps changing: from rights owners to the state, and then to lenders. The same phrase — the cost of holding AI up — means something different depending on who is on the other side of the payment.

One note on the day itself. Sunday morning brought very few primary announcements. The official feeds of OpenAI, Google and Google DeepMind all stopped at August 28, with nothing posted on August 29. All eighteen sources responded normally, so the quiet was real rather than a gap in collection. That left room to work the first story from the complaint itself.

Today's Top Three

1. Thirty-five Sony and Warner music publishers sue Anthropic and two co-founders for copyright infringement

Twenty-four Sony Music Publishing entities and eleven Warner Chappell entities — thirty-five plaintiffs in all — filed a copyright complaint on August 28 against Anthropic and two of its co-founders in the U.S. District Court for the Northern District of California, San Jose Division.

The case is Sony Music Publishing (US) LLC v. Anthropic PBC, No. 5:26-cv-09217 (N.D. Cal.). The complaint runs to 48 pages and demands a jury trial. Everything that follows is what the plaintiffs allege in that document. The court has made no findings, and the case is at its opening stage.

Start with who the defendants are. The caption reads "ANTHROPIC PBC, DARIO AMODEI, AND BENJAMIN MANN" — the two co-founders are named individually alongside the company. The party field in the electronic docket lists only the corporate entities, so a reader working from the docket summary alone would miss this.

The alleged infringement runs along two paths. The first is acquisition. The complaint says that in June 2021 Mr Mann, at Dr Amodei's direction, used BitTorrent to download at least five million pirated books from LibGen, and that in July 2022 the defendants torrented at least two million more from PiLiMi. Because BitTorrent uploads as it downloads, the plaintiffs argue that the distribution right was infringed alongside the reproduction right.

The second path is training and output. The complaint lists scraping lyrics from the sites of authorised licensees, downloading copies from third-party datasets, destructively scanning physical books, training Claude on those copies, and generating verbatim or near-verbatim lyrics as outputs — all pleaded as direct infringement by Anthropic.

There are four counts. Direct infringement by torrenting, against all defendants. Contributory infringement by torrenting, against Dr Amodei and Mr Mann. Direct infringement, against Anthropic. And removal or alteration of copyright management information.

The relief sought is specific. Statutory damages for wilful infringement of up to $150,000 per work, up to $25,000 per violation for the removal or alteration of copyright management information, and a permanent injunction.

One further request sits alongside those. The plaintiffs ask for an accounting of the training data, the training methods and the known capabilities of Anthropic's models — an order identifying which lyrics the models were trained on and disclosing how the material was collected, copied and processed. That request goes at the contents of the system rather than at the money.

On the $1.5 billion figure from the earlier case, we followed the complaint rather than the secondary coverage. TechCrunch describes the earlier Bartz case as one in which Anthropic was ordered to pay $1.5 billion. The complaint tells it differently: after another court in the same district found that Anthropic had illegally torrented over seven million copyrighted books, Anthropic settled that class action for $1.5 billion. A settlement and an order are different things, so we used the complaint's wording.

Among the compositions the complaint names are "Ain't No Mountain High Enough," "All I Want for Christmas is You," "Eye of the Tiger," "Here Comes Santa Claus" and "Paper Rings." Exhibit A covers the compositions said to have been taken in through books; Exhibit B covers those said to have been copied in training and in outputs.

Counsel for the plaintiffs are Oppenheim + Zebrak LLP and Pryor Cashman LLP. A separate suit brought by Round Hill Music LP against Anthropic has been pending in the same district since August 17.

2. Kioxia and SanDisk to invest about ¥5 trillion in Japan by 2032 — on the premise of government support

Kioxia and SanDisk announced on August 27 that they expect to invest a total of about ¥5 trillion (over US$31 billion) in Japan by 2032, on the premise of support from the Japanese government.

The money goes into production equipment, related infrastructure and technology at the Yokkaichi and Kitakami plants. Kioxia, the flash memory business spun out of Toshiba, has run its Yokkaichi fabs as a joint venture with SanDisk for more than twenty-five years; the two companies say they have together invested about ¥9 trillion (over US$50 billion) in Japan over that period. Both the yen and the dollar figures appear in the original announcement, so no conversion of our own has been applied.

On the same day, Kioxia disclosed that site preparation has begun for a third fabrication building at the Kitakami plant in Iwate Prefecture, to be built south of the second building and targeted to start operating during fiscal 2029. Construction timing and equipment investment will be decided in light of market conditions, and that investment, too, is premised on government support.

The word "premise" recurs through the announcement. Kioxia's president, Hiroo Ota, thanked the government for its support to date and said that continued and strategic support from the Japanese government matters for the group's international competitiveness. SanDisk's chairman and chief executive, David Goeckeler, said he is confident the investment will become a symbolic success of Japan–U.S. economic cooperation.

The week gave the announcement company. MONOist reports that China's YMTC had its listing application accepted by the Shanghai Stock Exchange on August 21, with proceeds expected around ¥780 billion, and that SK hynix of South Korea broke ground on an advanced HBM packaging plant in Indiana on August 27.

Japan's growth strategy sets a target of ¥101.6 trillion in combined public and private investment in AI and semiconductors by fiscal 2040. Read together, the three items describe an industry that has moved past the point where a company's own balance sheet carries a leading-edge fab on its own.

3. Lambda raises $1 billion in private debt to buy Nvidia chips — and leases them to Microsoft

The AI cloud company Lambda raised $1 billion in private short-dated debt to buy the Nvidia AI chips it will lease to Microsoft, Bloomberg reported.

Every figure below comes from TechCrunch's account of Bloomberg's reporting. The original Bloomberg piece sits behind a paywall and has not been obtained here.

Bloomberg says JP Morgan Chase arranged the notes. TechCrunch reads the structure itself as the signal: buying chips with short-dated debt suggests Lambda expects to deploy them quickly and repay out of the revenue they start generating.

This is one of several such loans. Lambda closed a $1 billion secured credit facility in May, and this week announced the closing of a $926 million loan to fund Nvidia GB300 GPUs for a deployment it is under contract to provide to Nvidia.

Equity is moving in parallel. The company is reported to be in talks for a $3 billion pre-IPO round; its last raise was $1.5 billion in November 2025 at a $5.43 billion post-money valuation, per PitchBook.

The aggregate matters more than any single deal. On Bloomberg's own count, banks and technology companies have raised over $400 billion in AI-related debt globally so far in 2026. The expansion of AI infrastructure is running on borrowed money rather than retained earnings, and the year's running total now puts a number on it.

Other Developments

Policy and Regulation

  • A review of the AI policies of 120 open source projects finds that 37 bar AI-assisted contributions outright. The Linux kernel permits them on condition that the use of an LLM is disclosed, while GCC, QEMU, SDL, Gentoo, Zig and Ghostty refuse AI-assisted contributions, and Codeberg, Sourcehut and Flathub prohibit the use of AI. Debian has put the question, including a full ban, to a developer vote. — Why open source projects ban AI (optimizedbyotto.com)

Business

Models and Research

  • Anthropic's blog, dated August 26, describes the shape of Warp's self-improving agent. An inner base skill generates reviews, and an outer improvement skill runs on a schedule as an observer, proposing small edits to the base skill as pull requests. Because skills are plain text files, review becomes possible on the same terms as code review. — How Warp builds self-improving agents on Claude (Anthropic)
  • The inference engine vLLM released v0.28.0, with 584 commits from 270 contributors. Full optimisation for Kimi-K3 and sparse MLA support for DeepSeek V4 are the headline items. — vLLM v0.28.0 (GitHub)
  • An account of an attempt to give an LLM memory that turned, in the end, into a program analysis problem. — Giving an LLM memory turned into program analysis (pwning.systems)

Products

Source: selected by the editorial desk from the AI news inbox for August 30, 2026 — 17 items, 1 primary and 16 secondary.