Tech Subtext - September 22, 2026

Trump wants an AI Force. A lawsuit says the labs colluded to slow down. The safety gospel just hit reality.

THE SLOWDOWN BACKFIRE

Trump wants an "AI Force." Nobody knows what it does.

I woke up to a president promising a new branch of government for robots.

On Saturday, Trump said he'll name an AI czar and stand up an "AI Force." He said he would appoint an artificial intelligence czar and create an "AI Force" to help monitor the technology, though he gave almost no details about either plan. He compared it to Space Force.

Here's the wild part. He also called fears of superintelligence a hoax. So the new force monitors a threat he says isn't real.

The timing is not an accident. Trump's plan surfaced just as pressure to rein in AI has been building from multiple directions at once, including inside his own party. The whole industry spent the last two weeks screaming about danger. Now Washington is scrambling to look like it's doing something.

Where Today Points

Two weeks ago the AI bosses preached the brakes. Now the brakes are getting tested by lawyers, regulators, and their own balance sheets.

The story today is the gap between what these companies say and what they do. Slow down, but ship. Warn about extinction, but chase a two trillion dollar IPO. Watch the contradiction, not the press releases.

Across the Industry

⚖️ Lawsuit says AI giants colluded to slow down

A group of plaintiffs just sued four of the biggest AI names. The claim is strange and clever. The lawsuit, filed Friday in the U.S. District Court for the Northern District of California, argues that the leading AI companies violated antitrust laws when they agreed to coordinate slowdown efforts, and that doing so would reduce the value consumers get for paid AI subscriptions.

The hook is a public one. The coordination largely took place on Sept. 12, when Anthropic CEO Dario Amodei published an essay urging industrywide cooperation on decelerating advancements, and that same day Sam Altman, Elon Musk and Demis Hassabis each publicly responded in agreement.

Why it matters: antitrust law, what stops rivals from secretly teaming up to fix a market, normally punishes companies for racing too little on price. Here the claim is they teamed up to race too little on capability. Slowing down together can look a lot like colluding together. And a senator already rejected giving the labs a legal pass to talk.

Source: ABC News

🇨🇳 China's CXMT cracks advanced memory without ASML

A Chinese chipmaker just did a thing US export rules were built to stop. CXMT announced mass production of its fifth-generation DRAM platform at a convention in Hefei, built at 11.95nm via quadruple patterning as a workaround for denied EUV access.

DRAM is the short-term working memory your phone and PC use to run apps. The banned machine here is EUV, the ultra-precise light tool that etches the finest chip circuits. CXMT didn't have one. It didn't need the banned machine. It needed four passes instead of one, and a government willing to fund the difference.

Why it matters: Washington bet that cutting China off from EUV would freeze its chipmakers at a ceiling. For the first time in more than a decade, Samsung, SK Hynix, and Micron together account for less than 90% of the global DRAM market. A fourth real supplier changes the long-run math on memory prices, which are already spiking.

Source: Eastern Herald

🕶️ Meta bets its whole hand on smart glasses

Meta Connect opens tomorrow, and for once the headset is not the star. Published September 21, this event is expected to be all about smart glasses. Zuckerberg is betting the future runs on your face, not a VR visor.

The rumored lineup: new Ray-Ban models, camera-free glasses aimed at privacy worries, and a pricey pair codenamed Phoenix. Don't expect a new Quest. Meta Quest 4 has been pushed to 2027, positioned as a gaming-focused flagship.

Why it matters: glasses are the form factor Meta wants for AI agents that stay with you all day. But there's a catch. Thanks to the ongoing RAM crisis, prices for new models of Oakley Meta and Ray-Ban Meta glasses could be higher than the current generation. The chip shortage from the last story lands right on the gadget in this one.

Source: Gizmodo

🧠 Alibaba ships a model that watches video and acts

Alibaba's Qwen team quietly shipped something that points at where AI is going. Qwen released Qwen3.8-Omni-Flash on September 18, redirecting engineering focus toward agentic task planning and delivery rather than static image description. It reads text, images, audio and video in one giant window.

The pitch isn't the smarts. It's the bill. QwenCloud lists $0.15 per 1M input tokens, and audio input costs over 98% less per hour than the prior model.

Why it matters: inference cost, what it costs to actually run a model after it's trained, is the real battlefield now. A Chinese lab dropping audio-video costs by 98% while US labs preach caution is a different kind of pressure. The frontier isn't just about who's smartest. It's about who's cheapest to run.

Source: TechNode

💸 Anthropic eyes $2 trillion IPO in November

Anthropic wants to go public, and it wants to go big. It has moved its planned IPO to November as investors discuss a deal that could value the Claude developer at roughly $2 trillion and raise up to $100 billion. That would be one of the largest listings ever.

The growth is real. Anthropic raised $65 billion in May at a $965 billion post-money valuation, and its annualized revenue run rate exceeded $65 billion by July's end, rising from roughly $9 billion in late 2025.

Why it matters: a $2 trillion price tag prices in years of faster models. That sits awkwardly next to a CEO telling everyone to slow down. Public markets will force disclosure Anthropic has never had to give, on compute costs, on safety, on how fast it's really moving. The pitch and the price are pulling in opposite directions.

Source: Unusual Whales

The Numbers

~$2 trillion — Anthropic IPO valuation target

November listing could raise up to $100 billion, one of the largest ever. Source

26% — Claude's share of Anthropic's own R&D

Up from effectively zero at the start of 2026. The AI is building the AI. Source

~10% — CXMT global DRAM market share

First time in a decade the big three hold under 90% of DRAM. Source

The Main Event

The slowdown gospel is meeting the real world all at once

Ten days ago, Dario Amodei wrote an essay. Slow the frontier, he said. Bring in outside evaluators. Move with care.

The other bosses nodded fast. Within hours of Amodei publishing his blog post, Musk responded saying "Dario is right," while Altman also said he agreed. It looked like a rare moment of unity.

Then reality showed up with three bills.

Bill one: the lawyers. Plaintiffs took that unity and called it a cartel. They allege Sam Altman, Elon Musk and Demis Hassabis made an illegal business agreement among competitors by agreeing to slow the development of their AI systems, violating federal antitrust law. Amodei even saw it coming. In his essay he acknowledged potential antitrust challenges, writing the government would need to issue a narrow waiver for certain kinds of safety conversations. Congress said no. Josh Hawley asked if he'd give the most powerful companies in history an antitrust exemption so they could collude together. His answer was no.

Bill two: Washington's theater. Trump answered the panic with an "AI Force" and a czar, and no plan. He wants the optics of control without the reality of it.

Bill three: the money. This is the one that stings. Anthropic turned Amodei's essay into action, sort of. Accenture and Anthropic announced they are partnering to establish embedded evaluators to red-team models and test safeguards, with each company expecting to invest at least $1 billion over five years. Outside eyes inside the lab. Good.

But look at what else the same lab did the same week. Its first R&D Automation Index shows Claude leading a quarter of the company's AI research tasks, up from less than 1% in February, arriving five days after Amodei's call for deliberate deceleration. And it's shopping a $2 trillion IPO. And Reuters says Anthropic is weighing a new model to counter GPT-6 Astra before an anticipated IPO.

So the pace essay, the R&D-that-builds-itself report, and the record IPO all landed in the same ten days. Preach the brakes. Floor the gas. Same company.

I don't think anyone here is lying. I think the incentives are just louder than the sermons. You can believe AI is dangerous and still need to beat the other guy to the next model. Both things live in the same building.

The tell for me is simple. When a company that says "slow down" also lets its own AI run 26% of its research and chases the biggest IPO in history, the safety talk isn't the strategy. It's the packaging.

The Subtext Report

An AI Force with no mission

Trump wants an AI Force like Space Force. Cool. What does it guard against? He also says the danger is a hoax. So it's a fleet built to patrol a threat that officially doesn't exist. That's not a mission. That's a logo.

Meanwhile the only detail he offered about the czar job was a vibe. High IQ only, apparently. Résumés to the void.

The evaluator who bills you

Anthropic's watchdog is Accenture, a consulting giant that also sells to half the enterprise market. The company being watched is paying the watcher directly, because no neutral funding exists yet.

It might genuinely help. Fresh eyes beat no eyes. But calling it independent when the lab writes the check is a stretch. The safest thing Anthropic did this week was hire someone to say it's being safe.

"AI systems are becoming exponentially more powerful and have begun to automate more of the process of building themselves."

— Anthropic, in its September 17 R&D Automation Index report