← the reads

the ai brief · oct 4, 2026 · mmxxvi

Anthropic's Bet on Inevitable Scarcity

The week's defining signal was financial: Anthropic's IPO prospectus revealed a $518 billion, decade-long compute commitment — about 80 percent of it payable whether the capacity is used or not — against 2025 revenue of roughly $4.6 billion and a net loss near $42 billion. For CTOs, that is the clearest picture yet of the take-or-pay economics underpinning the frontier, and it landed in the same week that Washington, rival labs, and even OpenAI's own staff turned the AI-safety argument into open political combat.

written

The clearest signal of the week came from a prospectus. Anthropic's long-awaited IPO filing, first reported by Reuters, put hard numbers on what chasing the frontier actually costs: revenue grew roughly twelvefold to nearly $4.6 billion in 2025, yet the company has committed about $518 billion to computing power over the next decade — and roughly 80 percent of that must be paid whether or not Anthropic uses the capacity. The same filing showed a 2025 net loss of about $42 billion (some $34 billion of it an accounting charge) and points toward a listing that could value the company above $2 trillion. For any CTO weighing vendor dependence, the subtext is blunt: the economics of frontier AI now rest on take-or-pay compute contracts at a scale few balance sheets have ever absorbed. Daniela Amodei, Anthropic's president, holds steady on the index as the operational face of that bet.

The politics arrived in the same news cycle. Senator Bernie Moreno of Ohio sent Dario Amodei a letter on Monday accusing Anthropic of hypocrisy — tying ordinary investors' money to an IPO while its chief executive warns the technology could drive humanity toward extinction. Moreno urged him against an "alarmist approach" and "panic-inducing commentary," writing that he was "concerned that you are failing to meet the moment." It is the latest instance of Trump allies targeting Amodei, and it followed a week in which six tech companies signed a voluntary White House accord on superintelligence and Sam Altman argued the world must accept the "bad things" done with AI while adding that "catastrophic risks" should not be accepted.

Fei-Fei Li holds the top spot, steady, on the strength of the week's biggest deal. AMD agreed on September 28 to buy her startup World Labs for about $8.2 billion in an all-stock transaction expected to close by year-end; Li will become AMD's executive vice president and chief scientist, reporting to chief executive Lisa Su. The purchase folds World Labs' "world model" research into AMD's silicon roadmap as it tries to close the gap with Nvidia.

The sharpest riser is Greg Brockman, and for uncomfortable reasons. According to a Semafor scoop, OpenAI's president pulled back a pledged second $25 million gift to Leading the Future, a pro-AI super PAC — he and his wife, Anna, had given $25 million last year — after internal backlash. In a Slack message, Brockman said he felt "terrible" about the group's conduct; employees had flagged that his money would help fund an account attacking Altman. It is a rare case of staff forcing a founder's political hand.

Among those cooling, Arthur Mensch of Mistral took direct aim at the safety camp, arguing that "the debate that we've seen in the U.S. has been a cover for the negligence of some of our competitors" — pointed language as rogue-agent incidents mount, and a sign Mistral has no plans to slow down. Liang Wenfeng also slipped despite good news: DeepSeek's annualized revenue run rate reached $1 billion, more than double its level of a few months ago, as it finalizes a roughly $7.5 billion round and eyes a Shanghai listing, with Liang telling investors that more than 70 percent of its compute still goes to training. Yang Zhilin fell too, buoyed only by softer coverage of his TIME100 Next listing.

Two risers point the other way. Andrej Karpathy drew wide attention for a practical post urging users to make models explain things in ASD-STE100, the controlled English of aircraft maintenance manuals, then escalate to diagrams, HTML pages, and explainer videos — a reminder that output legibility is becoming its own engineering problem. And Nick Bostrom resurfaced with a fresh warning on a New York Times podcast, more than a decade after his doom predictions first shaped the field.

The throughline is that money and politics have become the same conversation. In a single week the field's biggest acquisition, its largest pending IPO, and its loudest safety fight all turned on the same question — who pays, who profits, and who gets blamed — and the people moving on the index are the ones answering it out loud.

read-aloud

Start with a number, because this week a number said everything. Anthropic (an-THROP-ik) finally filed to go public, and its prospectus — first reported by Reuters — spelled out what chasing the frontier really costs. Revenue grew about twelvefold last year to nearly four-point-six billion dollars. But the company has committed roughly five hundred and eighteen billion dollars to computing power over the next decade, and here's the part that should make any technology chief sit up: about 80 percent of that has to be paid whether or not Anthropic actually uses the capacity. The same filing shows a 2025 net loss of around 42 billion dollars, and points to a stock listing that could value the company above two trillion. The operational face of that bet is Anthropic's president, Daniela Amodei (dan-ee-EL-uh uh-MOH-day), who holds steady near the top of the index. The lesson for a CTO is blunt — the economics of frontier AI now run on take-or-pay compute deals at a scale almost no balance sheet has ever carried.

And the politics landed in the very same news cycle. Senator Bernie Moreno (moh-REH-noh) of Ohio wrote to Anthropic's chief executive, Dario Amodei (DAR-ee-oh uh-MOH-day), on Monday, accusing the company of hypocrisy — tying ordinary investors' money to an IPO while its own boss warns the technology could push humanity toward extinction. Moreno told him to drop the "alarmist approach" and the "panic-inducing commentary," and said he was, quote, "concerned that you are failing to meet the moment." That came right after six tech companies signed a voluntary White House accord on superintelligence, and after Sam Altman said the world has to accept the "bad things" being done with AI — though he added that "catastrophic risks" should not be accepted.

Now, the biggest single deal of the week belongs to the person at number one, steady, Fei-Fei Li (fay-fay lee). AMD (A-M-D) agreed on September 28th to buy her startup, World Labs, for about eight-point-two billion dollars, all in stock. Li will join AMD as executive vice president and chief scientist, reporting to CEO Lisa Su (LEE-suh soo). The point of the deal is World Labs' so-called world-model research, which AMD wants to fold into its chips as it tries to catch Nvidia (en-VID-ee-uh).

The sharpest climber this week is Greg Brockman (BROK-mun), OpenAI's president — and it's for an awkward reason. According to a scoop from Semafor (SEM-uh-for), Brockman pulled back a pledged second twenty-five-million-dollar gift to a pro-AI super PAC called Leading the Future. He and his wife, Anna, had already given twenty-five million last year. This time, after an internal revolt, he backed off — telling staff in a Slack message that he felt "terrible" about the group's conduct, after employees flagged that his money would help fund an account attacking Altman. Staff forcing a founder's hand like that is genuinely rare.

On the way down: Arthur Mensch (AR-thur MENSH), chief executive of the French lab Mistral (mee-STRAHL), went after the safety crowd directly. He said, quote, "the debate that we've seen in the U.S. has been a cover for the negligence of some of our competitors" — and he made clear Mistral has no plans to slow down. Also cooling is Liang Wenfeng (lee-AHNG wun-FUNG) of DeepSeek (DEEP-seek), even with good news to report: the company's annualized revenue run rate just hit one billion dollars, more than double a few months ago, as it closes a roughly seven-and-a-half-billion-dollar round and lines up a Shanghai listing. Liang told investors more than 70 percent of his computing still goes to training new models. And Yang Zhilin (yahng jih-LIN) slipped too, carried only by gentler coverage of his spot on the 2026 TIME100 Next list.

Two risers lean the other way. Andrej Karpathy (AHN-dray kar-PAH-thee) got a lot of attention for a very practical tip — ask your model to explain things in a controlled English built for aircraft maintenance manuals, then move up to diagrams, web pages, and explainer videos. It's a reminder that just reading AI output is becoming its own engineering challenge. And Nick Bostrom (BOS-trum), the philosopher whose doom predictions shaped this whole field a decade ago, is back with a new warning on a New York Times podcast.

Here's the throughline. This week, money and politics became the same conversation. The biggest acquisition, the largest pending IPO, and the loudest safety fight all turned on one question — who pays, who profits, and who gets blamed — and the people moving on the index are the ones saying it out loud.

sources (91)