Google Cloud announced Ironwood, its seventh-generation TPU, for Cloud TPU customers on Nov 6, pitched at what the post calls the age of inference: 10x peak performance over TPU v5p and more than 4x per-chip performance versus TPU v6e for both training and inference, 9,216 chips in a single superpod over 9.6 Tb/s inter-chip interconnect sharing 1.77 PB of HBM. Anthropic, which said in October it plans to access up to 1 million TPUs, is the flagship customer.

The developer-facing change is the software layer: enhanced vLLM support on TPU behind a unified backend, so a serving stack can move between GPUs and TPUs with minor configuration changes, plus a GKE Inference Gateway that Google says cuts time-to-first-token by up to 96% and serving cost by up to 30% by load balancing across TPU servers. CNBC reported the same day that Google Cloud VP Amin Vahdat told an all-hands the company must double AI serving capacity every six months.

Edison Scientific, the commercial spinout of non-profit FutureHouse that took over its platform the same day, announced Kosmos on Nov 5. Kosmos runs for up to 12 hours on an open-ended objective and a dataset, executing about 200 agent rollouts that collectively write roughly 42,000 lines of analysis code and read about 1,500 papers per run. The architectural claim is that a structured world model, shared between a data-analysis agent and a literature-search agent, is what keeps a single objective coherent across tens of millions of tokens — the failure mode that capped earlier AI scientists at a handful of coherent steps.

In the technical report, independent scientists rated 79.4% of statements in Kosmos reports accurate, and every conclusion is traceable to specific lines of code or passages of literature. In a poll of seven beta users, a 20-cycle run was estimated at 6.14 months of their own research time, and estimated human-equivalent duration scaled linearly with run depth. It is available at $200 per run, with caveats the authors state plainly: Kosmos goes down rabbit holes and chases statistically significant but scientifically irrelevant findings.

Mrs Justice Joanna Smith handed down Getty Images v Stability AI on Nov 4. Getty's remaining secondary copyright infringement claim failed: the court accepted that a model can be an article under the Copyright, Designs and Patents Act, but held that Stable Diffusion is not an infringing copy because it does not store or contain the training images — once the weights stop holding a copy, they stop being one. Getty had already abandoned its primary copyright and database-right claims during trial because training happened outside the UK; it won only narrowly on trade mark, over generated images carrying its watermark.

Litigation volume continued in the same seven days: Danish collecting society Koda sued Suno in Copenhagen City Court on Nov 4, author Tasha Alexander filed a class action against Salesforce over its CodeGen and XGen models on Nov 5, and Entrepreneur Media sued Meta over Llama on Nov 6, taking US AI copyright filings past 60. The ruling follows the US arc of Judge Alsup's June 2025 decision that Anthropic's training was fair use but its retention of pirated books was not, and the resulting $1.5B settlement filed on Sept 5.

OpenAI and AWS announced a multi-year partnership on Nov 3: a $38B commitment over seven years, hundreds of thousands of Nvidia GB200 and GB300 GPUs on Amazon EC2 UltraServers, the ability to scale to tens of millions of CPUs for agentic workloads, and all target capacity deployed before the end of 2026. It is OpenAI's first contract with the largest cloud provider and the clearest break from the Microsoft-exclusive era; Amazon closed at a record, up 4%.

It is the fourth commitment in a run — roughly 10 GW with Nvidia, 6 GW with AMD, 10 GW of Broadcom custom accelerators and a reported Oracle deal — leaving OpenAI with close to $600B of cloud commitments against about $13B of expected 2025 revenue. The market started repricing that assumption two days later: Michael Burry's 13F disclosed puts on Nvidia and Palantir, and the Nasdaq fell 1.9% on Nov 6 with Palantir down more than 12% for the week.