Good morning. This week we’re looking at a new round of heavy investment in the AI infrastructure race, how GPT-5.1, Gemini and new “world models” are changing assistants, and why regulators suddenly care more about chips, deepfakes and child safety.
The AI infrastructure race becomes a power game
In the span of a few days, the cost of competing at the frontier jumped by another ~$100 billion, and that’s just from three announcements.
OpenAI signed a $38 billion, seven-year cloud deal with AWS, giving it immediate access to AWS data centres and “hundreds of thousands” of Nvidia chips (and diversifying from Microsoft). Sam Altman told the press OpenAI expects to spend around $1.4 trillion on AI infrastructure over the coming decade.
Anthropic announced a $50 billion investment in US data centres in Texas and New York, in partnership with Fluidstack. The sites will come online from 2026 and are described as “custom built” for Claude workloads.
Microsoft signed a $9.7 billion, five-year contract with IREN to secure Nvidia chips at a massive Texas campus.
Taken together, these deals show how quickly AI has shifted from “cheap API calls” to utility-scale infrastructure: multi-GW campuses, long-term chip contracts and power agreements that look more like national grid planning than software procurement. One analyst put it bluntly: the OpenAI–AWS deal “exposes how dependent AI is on centralized infrastructure”.
At the same time, politics is moving into the data centre. In Washington, the proposed GAIN AI Act would give US buyers first refusal on Nvidia and AMD’s most advanced chips before they can be exported, effectively turning GPUs into a strategic asset. Microsoft, Amazon and Anthropic all back the bill; Nvidia has warned it could hurt global competition.
Why it matters
The moat in AI is increasingly capital + chips, not just clever algorithms. That raises the bar for any company pretending to “build a frontier model” on the side.
For us, it means AI strategy is inseparable from vendor concentration, cloud risk and energy policy. A small number of providers will control most of the capacity — and outages or regulatory shifts suddenly matter a lot more.
For regulators, the conversation is shifting from “AI ethics” to systemic risk: power grids, export controls, and whether a handful of companies should effectively run the world’s AI infrastructure.
From chatbots to assistants and world models
While the infra race grabs the big numbers, the experience of using AI also changed meaningfully in the last two weeks — especially around assistants and new model types.
GPT-5.1 and Gemini grow up
OpenAI rolled out GPT-5.1, the new default behind ChatGPT, in two variants: Instant and Thinking. GPT-5.1 Instant is positioned as more conversational and “warmer”, with better instruction-following and new preset personalities (Professional, Friendly, Efficient, Nerdy, etc.). GPT-5.1 Thinking adapts how long it “thinks” depending on task complexity - trying to get closer to Claude’s level. A little bit more hidden: OpenAI seems to be planning a Group Chats for ChatGPT, letting teams now coordinate shared exchanges with custom prompts, potentially set for a December launch.
On the Google side, Gemini is quietly spreading across devices and modalities. We get:
Gemini for TV is now rolling out to Google TV Streamer devices;
Gemini Live, Google’s real-time voice interface with a new update: more natural prosody, adjustable speaking speed, accents, and better coaching flows for language learning and interview practice.
Google AI’s Search rolls out conversational shopping;
On the imaging side, Nano Banana - DeepMind’s top-rated image model - is now powering more of Google Photos’ editing, baking generative imaging into everyday consumer workflows.
In other words, the goal is now to move from “one chatbox on a website” toward persistent assistants embedded everywhere: phone, TV, browser, documents, photos.
Kimi K2 and Marble show where things go next
Two launches hint at the next wave beyond chatbots:
China’s Moonshot AI released Kimi K2 Thinking, an open-source “thinking model” tuned for multi-step reasoning and tool use. Benchmarks suggest it can match or beat some proprietary models on complex tasks at a fraction of the cost, sustaining 200–300 sequential tool calls for agentic workflows.
Fei-Fei Li’s startup World Labs launched Marble, a commercial world model that generates editable 3D environments from text, images or video, exportable into game engines and robotics simulators. A new competitor for Google’s Genie?
These aren’t just toys: they point toward agentic systems (models that plan and act over time) and simulation-heavy use cases (risk modelling, virtual testing, robotics) rather than isolated Q&A.
Why it matters for Zurich
The assistant space is maturing fast. GPT-5.1 and Gemini Live show what “Zurichat 2.0”-style experiences could look like: multi-modal, persistent, and much more “human” in how they speak and remember context.
Open models like Kimi K2 Thinking matter for cost, control and data residency. They make a multi-model strategy (OpenAI + Anthropic + open-source) more realistic — but also increase the governance burden.
World models and 3D environments could eventually matter for cat modelling, climate risk, mobility, and industrial insurance — wherever we rely on simulated worlds to estimate losses and test interventions.
More top news
Regulation & Policy
EU eases off the gas: The European Commission is considering a “simplification package” that would delay some AI Act obligations and grant grace periods for high-risk systems, after heavy lobbying from Big Tech and pressure from the US administration.
China’s AI security standards kick in: China’s national standards for gen-AI security, covering training-data safety and censorship of sensitive content, officially took effect on 1 November 2025. Generative-AI services must now meet detailed cybersecurity and content rules.
UK deepfake & child-safety law: The UK passed new powers to let regulators and child-protection agencies test AI models for their ability to generate child sexual abuse imagery. It also criminalises models built to generate such content.
Economy, Labour & Productivity
Fed looks at gen-AI adoption: A new blog from the St. Louis Fed finds early evidence that sectors with more reported time-savings from generative AI also show faster productivity growth relative to recent trends.
PwC’s workforce survey: PwC’s 2025 Global Workforce Hopes & Fears Survey reports that daily GenAI users are far more likely to report higher productivity (92% vs 58%), higher pay (52% vs 32%) and greater job security (58% vs 36%) than non-users - but only 14% of workers use GenAI every day.
Security, Deepfakes & Fraud
First “AI-orchestrated” cyber-espionage campaign: Anthropic disclosed that a China-linked hacking group used its Claude model not just as a co-pilot, but to execute most of a large-scale cyber-espionage campaign targeting tech firms, financial institutions and government agencies.
Tips and tricks
One of the struggle with AI adoption is finding uses cases that actually add value. For this you need to know your company and its struggle well, but also the model and its capabilities (to match both). To help with this, Anthropic released last week a “Use case library” showcasing what can Claude do. I encourage you to give it 5 minutes as you might find inspiration on how to use it for work or personal purpose.