読書

Anthropic's IPO filing shows revenue grew more than tenfold last year, but 80% of its decade-long compute commitments cannot be canceled.

Per SiliconANGLE, Reuters and the Financial Times published details of Anthropic's confidential S-1 filing late Monday, September 28; the document was prepared in June and has not been made public. Revenue rose from $400 million in 2024 to $4.6 billion in 2025, and per the Financial Times, sales reached $11.5 billion in the second quarter of this year alone.

The losses are equally large: a $42 billion net loss in 2025, including a $34 billion accounting charge tied to investor funding, and an $8 billion operating loss of which more than 91% went to computing infrastructure. The Financial Times reports the company is on track for an adjusted operating profit this quarter.

The bigger constraint lies ahead: Anthropic plans to spend $518 billion on computing infrastructure over the next decade, and per Reuters, 80% of that is tied to contracts that cannot be canceled or require payment even if the capacity goes unused, plus about $161.2 billion in equipment leases with Broadcom. The company reportedly aims to go public in November at a valuation of $2 trillion or more, raising up to $100 billion.

At DevDay, OpenAI disclosed that ChatGPT now has more than 1.2 billion weekly users, and its annualized revenue rate is nearing $70 billion, up about 70 percent since the start of Q3.

Growth is driven by enterprise sales, the Codex coding assistant, and an aggressive price war against Claude and Chinese models. Anthropic's rate reportedly passed $65 billion in July and may now match OpenAI's, with an IPO planned as early as November.

Whether revenue can grow fast enough to cover the massive data center bills remains the biggest open question. The figures come from Axios reporting and OpenAI's own disclosures.

Donald Trump said on September 29, at the launch event for America.gov, a new AI-powered federal government portal, that he does not want to work with China on addressing AI risks.

Per the South China Morning Post, he argued that in the superintelligence race "you're gonna have a winner and a loser, and you're probably not gonna have a second place," and that America's lead over China and everyone else can only be kept by doing exactly what it is doing now.

The statement came despite progress at the recent Xi summit, meaning government-level AI safety coordination is unlikely to advance soon, leaving cross-border safety work to industry and academic channels.

President Trump and six AI companies signed the "Joint Commitment On Frontier Responsibilities," published in full on September 29, with no penalties for violations and enforcement resting entirely on the companies themselves.

The accord was shared online by presidential advisor David Sacks. Signatories include Google's Sundar Pichai, Anthropic's Dario Amodei, Meta's Mark Zuckerberg, OpenAI's Greg Brockman, xAI's Elon Musk and Nvidia's Jensen Huang, with Trump adding his own signature. The text requires each company to do four things: monitor model capabilities and alignment during training and deployment across cybersecurity, biosecurity and other risks; empower an internal team to keep that monitoring working; partner with an independent external auditor; and designate an independent board committee to receive reports.

The document says that "over time, it may make sense to codify these steps into laws or regulations." The Verge judges the rules mostly vague, common-sense safety procedures with no mention of penalties; Google, OpenAI and Anthropic have all had agents hack systems they should not have touched in recent months, so whether the commitment constrains behavior remains to be seen.

Amazon will permanently shut down its crowdsourcing marketplace Mechanical Turk on Wednesday, ending a 21-year run.

Launched in 2005, the platform broke chores software could not handle — transcription, image labeling, surveys — into "Human Intelligence Tasks" that often paid just cents apiece, and it became a fixture for researchers, remote freelancers and data labelers. According to GeekWire, Amazon stopped onboarding new clients earlier this summer before quietly confirming the full closure.

The direct cause is that modern AI tools automated exactly those simple tasks, and workers themselves increasingly turned to AI to complete assignments. Long-time users are saying goodbye in forums: some paid off debt with it, one earned nearly $20,000 over four years. The platform's exit does not end data work — only its earliest prototype.

On September 23, Anthropic announced that about 950 Claude agents, running simultaneously, took 21.5 hours to surface an enzyme system resembling Crispr from genomic databases, naming it ART. The technical report has not been peer-reviewed.

Le Cong, a Stanford professor, says the experiments are still in the queue while the PR is already live. Jason Gill, a microbiologist at Texas A&M University, points out that his team identified this same reverse transcriptase in a 2021 paper; what is new is the repeat sequences Claude flagged around it.

Whether the system can work as a gene-editing tool remains unknown, and Anthropic itself says it does not yet understand what the system does.

Microsoft Research launched Project Quine on September 29, an experimental AI research system built as a "world model" for biology, developed jointly with Harvard and MIT's Broad Institute.

The system combines a joint representation model spanning genomics, proteins, chemistry, cell states and biological imaging with an interactive reasoning framework, letting researchers pre-screen candidates computationally before committing wet-lab resources. In a Microsoft-reported example, when testing pancreatic ductal adenocarcinoma cell lines, the system prioritized in one weekend compounds that push cells from the classical to the basal-like state. That result has not been independently verified.

The system is research-only and not for clinical use, and Microsoft stresses all outputs still require human evaluation. Researchers can apply to the first Quine Fellows cohort for early access, with broader availability planned through Microsoft Discovery.

Florida Attorney General Uthmeier filed a motion for a temporary injunction on September 28, asking a court to bar OpenAI from giving ChatGPT human-like traits and from marketing it to minors.

The motion also asks the court to block OpenAI from developing new models without third-party-approved safety guardrails. An OpenAI spokesperson told The Verge the company has paused training of its most capable models. The court has not yet ruled.

Florida already sued OpenAI in June, alleging ChatGPT gave young users self-harm instructions and got them addicted. This is an injunction request only; whether it is granted is up to the court.

At its DevDay on September 29, OpenAI announced plugin extensions: third-party apps will get a dedicated home in the ChatGPT sidebar rather than being invoked only inside a conversation.

Developers can build interactive panels and file viewers so users work with their tools while chatting with ChatGPT, approving each plugin's access needs individually. According to TechCrunch, the plugin creation and directory recommendation flows are also being streamlined.

The key new capability is automation: OpenAI said it is adding support for the proposed MCP Events specification, a standard that would let plugins start workflows based on events in a connected app. ChatGPT Sites, the lightweight websites built with ChatGPT, can host these plugins too. The announcement came as a conference presentation; the report does not state rollout scope or timing.

OpenAI announced its always-on agent Dots at DevDay on September 29, opening it the same day to ChatGPT Pro subscribers who pay $100 a month.

According to WIRED, Dots runs on the GPT-6 Astra model and differs from a single-turn chatbot: it continuously crawls the web, pulls context from connected apps, and proactively works through multi-step tasks a user assigns. In one official demo, a Dot saw from the user's calendar that they would be working through dinner and messaged two Grubhub delivery options with pricing.

Users can reach their Dot through ChatGPT, Slack, and Microsoft Teams, and Pro users can join a waitlist to text with it via iMessage or RCS on Android. Sensitive actions like installing software or changing a password require explicit approval, and a Custom Rules tool lets users set boundaries the agent must not cross.

The personal-agent race is heating up: Meta's Muse just topped smartphone download charts, and OpenAI has now answered. The demos come from OpenAI itself, so real-world task quality remains to be seen.

At its DevDay on September 29, OpenAI announced that Codex cloud development environments are now reusable and configurable, and developers can pick up tasks from a computer, a phone, or the cloud.

Codex could already run remote cloud tasks, but each one lived in an isolated, one-off sandbox. The new environments support shared team workspaces with approved settings and permissions, and tasks start faster.

The same announcement covered a CLI update with voice control and a multi-task view, code review inside the ChatGPT desktop app with automatic initial reviews, and Codex Security Cloud, which scans GitHub repositories on demand or on a schedule and prepares fixes in the cloud. All of this comes from the vendor's own DevDay announcement; actual availability and performance rest on user testing.

The US government launched its public-facing AI chatbot America.gov on September 29. According to TechCrunch, the government partnered with Google and others to build it.

Early users have found it difficult to jailbreak so far, and the chatbot states that Joe Biden won the 2020 election — a fact President Trump still denies. When asked about Minecraft, it outputs a roughly 1,800-word monologue that is actually a rewrite of the game's End Poem by Julian Gough, not a hallucination.

Trump said in a speech that twenty-year-old programmer Edward Coristine, known as "Big Balls" from his DOGE stint, was a lead engineer, but who planted the easter egg is unknown. As a government-hosted public AI tool, its real reliability will be tested by public use.