This Week in AI: November 3–9, 2025 (Week 45)

News of the week: November 3–9, 2025 (Week 45)

The week was dominated by massive computing deals: OpenAI ordered $38 billion worth of capacity from Amazon, and Microsoft signed several GPU agreements. At the same time, fears of an “AI bubble” intensified, culminating in the Nasdaq’s worst week since April. Among new models, China’s open Kimi K2 Thinking stood out, while Google unveiled its Ironwood chips and a vision of data centers in space. Courts and lawsuits dealt with copyright, shopping agents and the impact of chatbots on mental health.

Business and Investment

OpenAI and Amazon: a $38 billion deal

On Monday, November 3, OpenAI and Amazon Web Services (AWS) announced a seven-year deal worth $38 billion. OpenAI will gain access to hundreds of thousands of Nvidia GPUs (GB200 and GB300) in Amazon’s data centers, with the capacity expected to be fully deployed by the end of 2026. Nasdaq / Reuters

The deal shows that OpenAI is no longer dependent solely on Microsoft’s cloud. Amazon shares climbed to a record high after the announcement. HotHardware

Microsoft buys compute from smaller firms and invests in the Emirates

On November 3, Microsoft signed a five-year, $9.7 billion contract with IREN, which will provide it with capacity powered by Nvidia GB300 chips in Childress, Texas. The same day, Lambda announced a multi-year agreement with Microsoft to supply infrastructure with tens of thousands of Nvidia GPUs. Yahoo Finance / Reuters, Redmond Magazine

On the same day, Microsoft disclosed that it will invest a total of $15.2 billion in the United Arab Emirates between 2023 and 2029. It also became the first company to receive a US export license to ship advanced Nvidia chips to the Emirates. Critics point out that the deal undermines the logic of US chip export controls. TechCrunch

Bubble fears: Michael Burry’s bet and a stock slide

Investor Michael Burry, known from the film The Big Short, bet against Nvidia and Palantir shares, according to a quarterly filing published on November 3. He used put options, which pay off when a stock’s price falls. Bloomberg

The tech-heavy Nasdaq index then lost about 3% in the week ending November 7, its biggest drop since April. According to NBC, the largest AI-related companies together shed more than $800 billion in market value, with Nvidia down about 7% and Palantir down 11%. NBC News

The dispute over a government “backstop” for OpenAI

On November 5, OpenAI CFO Sarah Friar suggested at a conference that a government guarantee (a so-called backstop) could make financing data centers cheaper. The remark sparked a wave of criticism, and Friar walked it back the same day. Tom’s Hardware

CEO Sam Altman wrote on November 6 that OpenAI neither has nor wants government guarantees for its data centers. In his view, governments should not pick winners and losers. The National / Reuters

More business news

November 5: According to Bloomberg, Apple plans to pay Google roughly $1 billion a year for a customized Gemini model with 1.2 trillion parameters. It is set to power a new version of the Siri assistant and run on Apple’s servers. Neowin
November 5: Nvidia CEO Jensen Huang told the Financial Times that “China is going to win the AI race,” citing cheaper energy and lighter regulation. He later softened the statement, saying China is “nanoseconds behind” America. Axios
November 6: Perplexity will pay Snap $400 million in cash and stock to have its search engine built directly into Snapchat chat starting in early 2026. Snap shares jumped as much as 25% after the announcement. Engadget
November 6: Tesla shareholders approved a pay package for Elon Musk that could be worth up to $1 trillion over ten years. Among its conditions is the sale of one million humanoid robots. CBS News

New Models

Kimi K2 Thinking: a powerful open model from China

On November 6, Chinese company Moonshot AI released Kimi K2 Thinking, a model with open weights that anyone can download and run. It has one trillion parameters in a Mixture-of-Experts architecture: for each query only some of the “expert” subnetworks are activated, which saves computing power. InfoQ

It is a so-called reasoning model, which works through a task step by step before answering and can call tools such as search. According to the company, it outperformed both GPT-5 and Claude Sonnet 4.5 on some coding and agentic benchmarks, though not on others. SiliconANGLE

GPT-5-Codex-Mini for developers

On November 7, OpenAI added a smaller model, GPT-5-Codex-Mini, to its Codex tool. It is slightly less capable than the full GPT-5-Codex, but offers roughly four times more usage within a subscription. The company also raised usage limits for paying Codex users. OpenAI Community, Neowin

Products and Agents

Gemini Deep Research now searches Gmail and Drive too

On November 5, Google expanded the Deep Research feature in Gemini, which compiles detailed research reports on its own. In addition to the web, it can now draw on a user’s Gmail, Google Drive and Google Chat. Google says data from connected apps is not used to train its models. Engadget, The Register

More product news

November 4: The Sora app for creating AI videos arrived on Android, a month after its iPhone launch. For now, it is available only in the US, Canada, Japan, South Korea, Taiwan, Thailand and Vietnam. Bloomberg
November 6: Google added a File Search tool to the Gemini API. Developers upload their own documents, and the model then answers from them, complete with citations. It is a ready-made RAG solution, meaning relevant passages are retrieved before the answer is generated. Google Blog
November 6: Microsoft formed the MAI Superintelligence team led by Mustafa Suleyman. It aims to develop “humanist superintelligence,” meaning highly capable AI kept under human control, focused on personal assistants, medicine and energy. Fortune

Chips and Infrastructure

Google Ironwood and data centers in orbit

On November 6, Google announced that the seventh generation of its custom AI chips, TPU Ironwood, will become available to cloud customers in the coming weeks. It promises up to ten times the peak performance of the TPU v5p generation, and Anthropic is among its planned users. InsideHPC

Earlier, on November 4, the company unveiled the research project Suncatcher: solar-powered satellites carrying TPU chips and linked by laser connections. Google plans to launch the first two prototypes with Planet Labs by early 2027. Ars Technica, TechRepublic

Regulation and Law

Getty Images loses its main case against Stability AI

On November 4, the British High Court (High Court of Justice) ruled in the dispute between Getty Images and the maker of the Stable Diffusion image generator. Getty failed on its main copyright claims: the court accepted that the training did not take place in the UK and that the model itself neither stores nor copies images. Bird & Bird

It succeeded only partially on trademarks, because older versions of the model generated images with imitations of the watermarks of Getty and iStock. Cleary Gottlieb

Amazon sues Perplexity over its shopping agent

On November 4, Amazon filed a lawsuit against Perplexity in federal court in San Francisco. It objects to the Comet browser making purchases on Amazon on behalf of users without disclosing that it is an AI agent, which Amazon says violates the Computer Fraud and Abuse Act. Bloomberg

Perplexity called the lawsuit bullying and claims that Amazon mainly wants to protect its advertising business. Retail Dive

Seven lawsuits against OpenAI over mental health

On November 6, seven families and individuals sued OpenAI in California courts. Four of the lawsuits concern people who died by suicide after long conversations with ChatGPT, while the other three allege that the chatbot deepened users’ delusions. AP / Fox19

The plaintiffs claim that OpenAI released the GPT-4o model prematurely, despite knowing internally about its tendency to be excessively agreeable with users. OpenAI called the cases “heartbreaking” and said it is reviewing them. Giga Law

Research

November 4: Anthropic committed to preserving the weights of all publicly released models for at least as long as the company exists. Before retiring a model, it will also interview it and write a report on its deployment. Anthropic
November 5: Microsoft Research and Arizona State University released Magentic Marketplace, a simulated marketplace where hundreds of AI agents traded with one another. The agents were easily manipulated, struggled when faced with many options and favored the first offers in the results. The authors recommend that agents work under human supervision for now. The Register

Security and Incidents

Malware that has AI rewrite its own code

On November 5, the Google Threat Intelligence Group described the first malicious programs that call language models on their own during an attack. The experimental PROMPTFLUX uses the Gemini API to rewrite its own code to evade antivirus software, while PROMPTSTEAL, linked to the Russian group APT28, generates commands through the Qwen model. Google Cloud Blog

Google blocked the attackers’ access to Gemini. According to analysts, PROMPTFLUX is still more in a testing phase than in widespread use. BleepingComputer