Two weeks ago the heads of the biggest AI labs agreed out loud that everybody should ease off the gas. This week four of their paying customers hauled them into federal court over it. Meanwhile Anthropic pointed a swarm of agents at a DNA database and came back with something that looks like CRISPR, and an Australian company nobody outside the data center world had heard of is trying to borrow ten billion dollars to buy Nvidia chips.
Busy week. Here is what actually happened.
01 of 04
The AI Slowdown Pact Just Became an Antitrust Case
On September 12, Anthropic CEO Dario Amodei published an essay calling on frontier labs to "pace the frontier" - coordinate on slowing capability jumps, submit to independent safety evaluations, push for international cooperation. Sam Altman backed it. Elon Musk backed it. Demis Hassabis at Google DeepMind backed it. For about six days it looked like the rare moment where the industry agreed on something.
On September 18, four people who pay for ChatGPT, Claude, Grok and Gemini filed a class action in the U.S. District Court for the Northern District of California naming Anthropic, OpenAI, SpaceXAI and Google. The case is docketed as Buist v. Anthropic. The claim is straightforward Sherman Act stuff: four competitors publicly agreed to restrain output, and the people paying monthly for that output got less product for the same money.
The complaint is careful about what it is not saying. The plaintiffs do not argue that any single lab is barred from slowing itself down for safety reasons. Their argument is that the companies took what the filing calls a shortcut, substituting collective restraint for individual accountability. Decide on your own to ship slower and that is your business. Agree with your three largest rivals to all ship slower and you have a cartel, whatever the stated motive.
The uncomfortable part for the defendants is that there is no discovery problem here. Usually an antitrust plaintiff spends two years digging for the smoking gun email. In this one the agreement was an essay with a byline and four executives publicly endorsing it.
Here is where I land on it. I think the safety argument is real and I think the lawsuit is also probably correct on the law, and those two things being true at once is exactly the mess. Antitrust does not have a "we meant well" exception. If the labs genuinely believe pacing is necessary, the honest route was always a regulator setting the pace, not four CEOs shaking hands in public and hoping nobody noticed that is literally the thing the Sherman Act was written about. Four annoyed subscribers noticed.
02 of 04
Anthropic Pointed 950 AI Agents at DNA and Found a New Enzyme System
Anthropic announced a life sciences research group and a physical lab on September 23, and led with a result rather than a roadmap. Roughly 950 Claude-powered software agents spent 21 hours combing a DNA sequence database. They surfaced more than 200,000 genes for a single enzyme type, narrowed that to 20 worth a human look, and one of those turned out to be a previously uncharacterized enzyme system with a structure resembling the DNA repeats behind CRISPR. The system appears in phages - viruses that infect bacteria.
The lab is in the Bay Area, was stood up in spring of 2026, and operates only at biosafety levels 1 and 2. No human pathogens. Every wet-lab step is run by human scientists. Anthropic was explicit that Claude is not turned loose in there.
Worth being precise about what was claimed. Anthropic is not saying Claude invented a new gene editing tool. It is saying the agents found a structure nobody had characterized, and that what it actually does is still an open question. The pitch is about the screening, not the biology: twenty candidates pulled out of two hundred thousand genes in under a day, work the company says would take an expert weeks or months.
What I want to know is whether that enzyme system does anything useful. Right now this is a very expensive, very fast literature and database search that produced an interesting lead. That is genuinely valuable - most of science is narrowing the search space - but a lead is not a discovery until somebody characterizes it. The detail I keep coming back to is the human-in-the-loop part. A frontier lab building its own biology wet lab and then volunteering that the AI does not touch the bench is a company that has read the room.
Source: TechCrunch, Al Jazeera
03 of 04
OpenAI and Anthropic Cut AI Prices on the Same Morning
September 22, Anthropic shipped Claude Opus 5.5. Input and output run $4 and $20 per million tokens, twenty percent under Opus 5, with cache reads at $0.20 per million, a sixty percent cut. Anthropic says it performs at Claude Fable 5.1 level on most work while costing about forty percent less to run, and called it the strongest-performing model it has tested.
About ninety minutes later OpenAI launched GPT-6 Sol and GPT-6 Luna. Sol is $2 per million input and $10 per million output, down from $4 and $20. Luna is $0.10 and $0.50, down from $0.20 and $1.20. That is roughly half the previous generation across the board, and per tier it undercuts Anthropic.
Ninety minutes is not a coincidence. Somebody had a launch calendar and somebody else had a finger on the publish button.
The interesting number in that pile is Anthropic's cache read price. Agentic and coding workloads are mostly cache reads, so a sixty percent cut there moves the real bill a lot more than the headline token rate does. OpenAI is competing on the sticker and Anthropic is competing on the invoice.
Honestly, this is the best week developers have had in a while and I do not think it lasts in this form. Price wars this aggressive usually end with the cheap tier getting quietly capped, rate limited, or deprecated into something less useful. Enjoy the $0.10 input tokens, but do not build your unit economics on the assumption that Luna stays that price through next year.
Source: VentureBeat, CNBC
04 of 04
A $10 Billion Loan to Buy AI Chips, Before the Company Is Even Public
Firmus Technologies, an Australian outfit backed by Nvidia, is talking to lenders about roughly $10 billion in financing to buy Nvidia chips for an Indonesian data center. The structure is about $7.5 billion in debt split between senior and mezzanine tranches, plus around $2.5 billion in equity. If it closes it is one of the largest AI infrastructure deals in Asia.
The money is pointed at a 360-megawatt Nvidia DSX facility in Batam, Indonesia, covering access to up to 170,000 Nvidia accelerators across the Grace-Blackwell, Vera-Rubin and Vera platforms during 2027 and 2028. On September 8 Firmus said OpenAI signed on as anchor customer for two planned Malaysian sites, pushing contracted capacity past 900 megawatts.
And Firmus started pitching an IPO this week. It is looking to raise up to A$7 billion, about US$5 billion, on the Australian Securities Exchange with a listing targeted for the end of October. That would be the biggest Australian float since Telstra was privatized in 1997.
So the sequence is: borrow ten billion to buy chips from your own investor, then go public in Sydney a few weeks later. Every piece of that is legal and every piece of it is how the whole buildout works right now, which is sort of the point. Nvidia is increasingly financing the purchase of Nvidia hardware, and the Firmus deal is a clean example of the loop.
The thing I would actually check before touching that IPO is the contract book. Nine hundred megawatts of contracted capacity with OpenAI anchoring it is a real number and it is the only thing standing between this and a very leveraged bet on 2028 demand. Debt does not care whether the AI trade is still hot in eighteen months. It just wants paying.
Source: Bloomberg, Startup Daily