Good morning. It's Wednesday, November 11, and Anthropic has built its business around a compelling promise: making powerful AI safer. The company warns that advanced AI could pose serious risks, while positioning its safety research as a competitive advantage. But there’s a tension at the heart of that strategy. More below.

Plus today's top stories: Jaron Lanier challenges Silicon Valley’s AI narrative, China builds data centers at a fraction of U.S. costs, and Google brings humanoid robots to Mercedes-Benz factories.

And don’t miss the new Real or AI challenge.

Anthropic has a remarkable pitch: Artificial intelligence could transform civilization, potentially threaten it, and the company best positioned to help manage those risks is…Anthropic!

This is a smart way to differentiate a business in a market where the technology is advancing faster than most users can keep up. But it’s a pretty uncomfortable message coming from one of the companies racing to build increasingly powerful AI. 

And now, Anthropic has effectively acknowledged the commercial logic behind that positioning.

In its latest IPO prospectus, filed November 10 as the company prepared to market its shares to investors, Anthropic describes safety research as central to its mission, but also as a source of customer trust, enterprise adoption and competitive advantage. The company is targeting an initial offering price of $70 to $78 per share.

That is Anthropic explaining its own business model, not an accusation from a skeptical competitor. 

Two things can be true at the same time: Anthropic's researchers genuinely believe advanced AI presents extraordinary risks including biological weapons and advanced cyberattacks. And Anthropic has built a commercial strategy in which being the company that takes those risks most seriously is a competitive advantage.

But here's the paradox: Anthropic believes AI is dangerous enough that it needs to help lead the race to develop it. Staying competitive means building increasingly powerful AI, which introduces new risks and makes its safety mission even more important.

When the solution to dangerous AI is building even more powerful AI, who decides when enough is enough?

Read the full story →

Google DeepMind says Gemini Robotics 2.5 helped Apptronik’s Apollo 2 humanoid sort metal parts with 96% success on active Mercedes-Benz production lines. The vision-language-action model turns visual information and spoken instructions into robot movements, and Google says it completed tested tasks 20% to 30% faster than its previous systems.

Google Cloud launched Gemini for healthcare, an AI platform that connects electronic health records, billing systems, clinical guidelines and government datasets. The preview is designed for HIPAA-compliant work and will be available to Gemini Enterprise Standard and Plus customers through subscriptions or usage-based pricing.

The U.S. Energy Information Administration expects electricity use to rise from 4,216 billion kilowatt-hours in 2025 to 4,306 billion in 2026 and 4,370 billion in 2027. AI data centers and wider electrification are driving the increase, while renewables are projected to supply 28% of U.S. generation by 2027.

Anthropic could begin trading as soon as next week with a potential valuation above $1 trillion, making it the first major standalone AI model developer to enter public markets. The Claude maker reported $33.2 billion in sales through the first nine months of 2026, but five customers supplied one-quarter of that revenue and the company says its current growth rate will not last.

Computer scientist and virtual-reality pioneer Jaron Lanier argues that Silicon Valley is making useful AI harder to accept by describing it as a conscious creature, a god or an extinction threat. His alternative is simple: treat AI as a powerful tool built from human knowledge, then design products that keep people in control instead of wrapping models in autonomous personalities.

Lanier’s sharper criticism is aimed at AI agents, which he says are often unnecessary and create many of the security and control problems that labs cite as evidence of existential risk. His proposal is to regulate autonomous agents directly and favor interactive systems that help people explore choices. → Read Jaron Lanier’s NYT essay. (Paywall)

SemiAnalysis toured Galaxy, an AI data-center campus in Inner Mongolia planned for 800 megawatts of computing capacity—roughly the scale of OpenAI’s core Stargate campus in Abilene, Texas. Developer Envision delivered the first 120-megawatt phase within a year at about one-fifth of U.S. construction costs, while building its own wind farms, batteries, power equipment and dedicated transmission line.

The campus also runs a live five-megawatt pilot that converts high-voltage alternating current directly into 800-volt direct current for server racks, removing several layers of power equipment. China’s lower costs and tighter vertical integration could reshape the infrastructure race. → Read the full SemiAnalysis report.

MIT and University of British Columbia researchers built WorldParticle, one AI system that simulates materials ranging from flowing water and sliding sand to folding proteins. Instead of hand-coding a different solver for each physical process, the model represents matter as particles and predicts how those particles move one step at a time.

The team says WorldParticle handled six types of physical phenomena more accurately than conventional numerical methods and modeled protein motion about 100 times faster. → Read MIT’s WorldParticle report.

Ring Launches a Security Drone and Rover: Indoor Drone Cam costs $249.99, business Rover Cam $999.99; both use AI Pilot navigation.

Claude Starts Working Before Meetings End: Claude Meetings enters beta for Team and Enterprise on Google Meet, with new Whiteboards and Sheets.

Meta Gives Muse Its Own Email Address: Paid U.S. and Canadian users can add Muse to email threads or forward messages for replies and follow-ups.

eBay Blocks Some Outside AI Shopping Agents: Meta's Muse and other third-party agents face limits on browsing listings over unauthorized scraping.

Adobe Opens Creative Cloud Tools to AI Assistants: A new MCP server lets ChatGPT, Claude and Gemini use Photoshop, Illustrator and Premiere.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team