AI Roundup September 2025: Nvidia Bets $100B on OpenAI, Anthropic Pays $1.5B, Claude Sonnet 4.5 Lands
TL;DR
September was the month the money got absurd and the lawyers got paid. Nvidia signed a letter of intent to put up to $100 billion into OpenAI, Anthropic agreed to hand authors $1.5 billion over pirated training books, and Anthropic capped the month by shipping Claude Sonnet 4.5, its best coding model yet. Meanwhile Alibaba pushed Qwen past a trillion parameters and Mistral pulled in fresh billions with a chip giant leading the round.
Nvidia Pledges Up to $100 Billion to OpenAI in a Circular Compute Deal
On September 22, OpenAI and Nvidia announced a letter of intent for a strategic partnership: OpenAI commits to deploying at least 10 gigawatts of Nvidia systems, and Nvidia intends to invest up to $100 billion into OpenAI, paid in progressively as each gigawatt comes online. The first gigawatt is slated for the second half of 2026 on Nvidia's upcoming Vera Rubin platform.
Read that structure again, because it is the whole story. The chip vendor is funding the customer that buys its chips. The cash Nvidia puts in flows back out as GPU orders, which lands on Nvidia's revenue line, which props up Nvidia's market cap, which funds the next tranche. It is the most explicit version yet of the circular financing that now underpins the entire AI buildout.
For builders, the signal matters more than the structure. Ten gigawatts is a genuinely enormous amount of compute, and it tells you the frontier labs are betting the next generation of models is compute-bound, not idea-bound. One caveat worth holding onto: a letter of intent is not a signed contract, and Nvidia itself later cautioned there was no assurance the deal closes on these terms. Treat the headline number as an intention, not a wire transfer.
Anthropic Agrees to Pay Authors $1.5 Billion Over Pirated Training Books
On September 5, Anthropic agreed to settle the Bartz class action for a minimum of $1.5 billion, the largest payout in the history of US copyright law. The suit alleged Anthropic trained Claude on books pulled from shadow libraries like LibGen. The deal works out to roughly $3,000 per book across an estimated 500,000 titles, and Anthropic agreed to destroy its copies of the pirated works. Judge William Alsup gave preliminary approval on September 25.
The nuance that every builder should internalize: the earlier ruling in this case found that training on legally acquired books can be fair use. What got Anthropic was the piracy, the act of downloading from illegal sources in the first place. The lesson is not that you cannot train on books. It is that how you got the data is now a billion-dollar question.
If you are building anything on scraped or borrowed corpora, this is the precedent that changes your risk calculus. Provenance is no longer a nice-to-have. Keep receipts for your data, because the courts have now put a per-item price tag on the alternative.
Claude Sonnet 4.5 Ships as Anthropic's Best Coding Model
Anthropic closed the month on September 29 with Claude Sonnet 4.5, which it positioned as state-of-the-art on coding and agentic benchmarks and capable of running long, autonomous software and business tasks. Crucially, pricing held steady at $3 per million input tokens and $15 per million output, the same as Sonnet 4. You get a meaningfully better model at the same price you were already paying.
That pricing decision is the part indie hackers should care about. Sonnet has been the default workhorse inside Claude Code precisely because it hits the speed-to-quality sweet spot, and a free capability bump with no cost increase is exactly the kind of quiet win that compounds across thousands of agent calls. If your agent stack runs on Sonnet, this was a drop-in upgrade.
Alibaba's Qwen3-Max Crosses a Trillion Parameters
On September 5, Alibaba's Qwen team released a preview of Qwen3-Max, the first model in the Qwen line to exceed a trillion parameters, trained on roughly 36 trillion tokens. The preview landed near the top of the LMArena text leaderboard, ahead of names like GPT-5-Chat, Grok 4, and DeepSeek V3.1, while sitting below the very top tier of frontier models.
The asterisk for the local-AI crowd: Qwen3-Max is proprietary, a sharp turn from the open-weight releases that made Qwen a homelab favorite. You cannot pull this one down and run it on your own rack. But it is a clear marker that the Chinese labs are now playing at genuine frontier scale, and the competitive pressure that produces tends to flow downhill into the open-weight models you actually can run.
Mistral Raises 1.7 Billion Euros With ASML Leading
Europe's open-weights champion got a war chest. On September 9, Mistral closed a 1.7 billion euro Series C at roughly a 14 billion dollar valuation, with Dutch chip-equipment giant ASML leading via a 1.3 billion euro check for about an 11 percent stake. Nvidia, a16z, DST Global, and others joined.
The interesting part is who led. ASML makes the lithography machines that make the chips, so this is a hardware supply-chain player buying directly into a European model lab. It is a deliberate bet on European AI sovereignty and a hedge against total reliance on Silicon Valley. For the open-source ecosystem, a well-funded Mistral is good news, because it remains one of the few serious labs still shipping genuinely open weights.
Key Takeaways
- Compute is the new moat, and it is being financed in circles. The Nvidia-OpenAI deal shows frontier labs and their chip supplier locking arms financially, which concentrates power and raises real questions about how durable these valuations are.
- Data provenance just got a price tag. Anthropic's $1.5B settlement makes clear that how you sourced training data is now a material legal liability, not a footnote.
- Same price, better model is the quiet trend that matters. Claude Sonnet 4.5 upgraded capability without touching the price, which is exactly the kind of efficiency gain that compounds for anyone running agents at scale.
- The frontier is global now. Alibaba crossing a trillion parameters confirms the Chinese labs are competing at the top tier, even as some of those releases close up and go proprietary.
- Europe is funding its own stack. ASML leading Mistral's round is a sovereignty play, and it keeps one of the last serious open-weights labs flush with cash.