Anthropic hits $65B in revenue, Amazon destroys rare books for AI training, and a powerful Chinese model drops

Something quietly unsettling is running through today's news - and it's not one story, it's a pattern. AI is eating into things we assumed were safe: rare books, courtroom evidence, policy reports. Meanwhile the money flowing into this space just keeps climbing to numbers that didn't seem real two years ago.

Let's get into it.

Z.ai's new open-weight model making headlines today is built on technology originally developed by which Chinese AI company?

The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇

Anthropic's annualized revenue has hit $65 billion - and the jaw-dropping part isn't the number itself, it's the pace. The company added $18 billion in annualized revenue in just two months. To put that in context: that's faster growth than almost any software company in history at this scale.

Claude has become the go-to model for enterprise coding, research, and document analysis workflows. The revenue trajectory suggests businesses aren't just experimenting anymore - they're building Anthropic's API into production pipelines at serious scale.

If you thought OpenAI had a lock on enterprise AI revenue, this number says the race is far more competitive than the headlines suggest - and Anthropic is closing fast.

Read the full story →

Here's an uncomfortable irony: Amazon - the company that built its empire selling books - is now reportedly destroying rare texts to digitize them for AI training data. Rare books are considered especially valuable for training language models because they contain knowledge that never made it online, giving models access to ideas and vocabulary that web-scraped training sets simply can't provide.

The destruction process involves cutting the spines of physical books so they can be fed through high-speed scanners. Once digitized, the original is gone. For rare or one-of-a-kind texts, that means a permanent loss of an irreplaceable artifact in exchange for training tokens.

Libraries and archivists have spent decades preserving these books precisely because they can't be recreated. The argument from Amazon's side is presumably that digital preservation beats physical decay - but critics point out that "preservation" and "destruction for commercial AI training" are very different things.

This story sits at the intersection of AI copyright, cultural preservation, and corporate incentives - and it's the kind of quiet, unglamorous decision that rarely makes headlines until the books are already gone.

Read the full story →

Z.ai has released a new open-weight model that cybersecurity experts have been tracking with serious concern. The model was both anticipated and warned against - researchers flagged it as capable enough to help security professionals harden systems, but equally capable of landing in the hands of malicious actors looking to find vulnerabilities.

Because it's open-weight, there's no API gate or usage policy that can restrict access once it's out. Anyone can download, fine-tune, and deploy it - including for offensive cybersecurity operations. That's the core tension with powerful open-weight releases: the same properties that make them valuable for researchers make them difficult to contain.

This isn't a hypothetical anymore. As we've covered in our AI safety coverage, the debate over open versus closed model releases has been simmering for two years. Z.ai's release is the kind of real-world test case that moves that debate from theory to practice. Wired notes that experts had been both waiting for and warning about exactly this model.

If your organization handles sensitive systems, this is a good week to revisit your threat model - because the capabilities bar for what a motivated attacker can access just moved.

Read the full story →

Australia's teen social media ban - one of the most aggressive AI regulation and platform regulation moves in the world right now - may have a credibility problem. A Senate hearing this week heard that a report used to justify the ban's underlying technology contains links to academic articles that simply don't exist. The Guardian's analysis found the citations appear to be ChatGPT hallucinations - fabricated references that look real but lead nowhere.

The authors of the report have conceded that ChatGPT was used in its preparation, but deny that the fake references were AI-generated. That distinction may not matter much politically - the damage to the report's credibility is already done, and opponents of the ban now have a very concrete line of attack.

Using AI to help draft policy documents about AI regulation, and ending up with hallucinated citations as a result, is the kind of story that will be cited in AI ethics courses for years.

Read the full story →

Groq - the company that built its reputation on blazing-fast AI hardware with its LPU chips - just raised $350 million at a $3.5 billion valuation. The twist: Groq is pivoting away from being a chip company and toward being a "neocloud" provider. That means expanding its data center footprint powered by - somewhat ironically - Nvidia GPUs.

The speed advantages Groq built on its custom silicon are real, but competing in the chip design business against Nvidia and custom silicon from Google and Amazon is brutally hard. Pivoting to a cloud provider that focuses on inference speed - regardless of the underlying hardware - is a more defensible business position.

If you're building applications where inference latency is your bottleneck, Groq's neocloud play is worth watching - especially now it has $350M to expand capacity. If you want to understand how token costs and speed trade-offs work for your own projects, our Token Calculator can help you model the numbers.

Read the full story →

A UK Sainsbury's store has paused its Facewatch facial recognition system after Matt Arnold, 46, was wrongly identified as a shoplifter and ejected from the store. "I was embarrassed, mortified even, and felt quite humiliated and powerless," Arnold said. The supermarket chain attributed the incident to "human error" rather than a failure of the Facewatch technology itself - a distinction that may not hold up to scrutiny.

False positive rates in face recognition systems are well-documented, and they tend to be higher for certain demographic groups. The Sainsbury's case is a useful reminder that "AI-assisted" doesn't mean "accurate," especially in high-stakes real-world environments where a wrong call has immediate consequences for a real person.

As AI-powered surveillance scales into everyday retail spaces, the question of who bears the consequences of a false positive - and whether "human error" is an acceptable explanation - deserves a much clearer answer from the companies deploying these systems.

Read the full story →

Speaking of AI tools actually worth using - if you need to spin up a website fast, 60sec.site lets you build and launch an AI-generated website in under a minute. No code, no fuss. Worth bookmarking for your next project.

The answer is B) Zhipu AI (Z.ai itself). Z.ai is the company formerly known as Zhipu AI, the Beijing-based research lab behind the ChatGLM series of models. Their latest open-weight release is the model that experts have been monitoring for its dual-use cybersecurity implications - powerful enough to be genuinely useful, and open enough to be genuinely worrying.

Amazon destroying rare books to train AI models - does that feel like a fair trade-off to you, or does it cross a line? Hit reply and tell me what you think. I read every response and genuinely curious where readers land on this one.

That's all for today. See you tomorrow with more from the fast-moving world of AI - and if you want to browse everything we've covered, the full Daily Inference archive is always there.