Anthropic's Leaked S-1, OpenAI's Aussie Mess, and AI's $1.4T Bet

01 of 03

Anthropic's leaked S-1 shows the bill for the AI race

Reuters and the Financial Times got their hands on Anthropic's IPO prospectus, the one it filed confidentially with the SEC back in June. It runs 261 pages and the numbers are wild in both directions.

Revenue for 2025 was $4.6 billion, up 12x from the year before. Then it jumped to $11.5 billion in just the second quarter of 2026. That is one quarter beating the whole prior year by more than double.

The losses are just as loud. The operating loss for 2025 was $8.06 billion, up from $2.98 billion in 2024. Operating expenses hit $12.65 billion, and $7.33 billion of that was compute and infrastructure. The headline net loss is $42 billion, but $34 billion of that is a non-cash accounting charge tied to the estimated value of future share conversions. So the real cash burn story is the $8 billion, not the $42 billion. Anthropic had $20.28 billion in cash and short-term investments on hand.

Then there is the compute tab: $518 billion in commitments over the next decade across six partners, and about 80% of it can't be cancelled. Here is who gets paid:

• Broadcom equipment leases: $161.2 billion
• Google: $111.1 billion
• Amazon: $110 billion
• xAI, for Nvidia capacity: up to $84.5 billion through 2029, mostly cancellable on 90 days' notice
• Microsoft: $31.4 billion
• AMD: more than $20 billion expected

Two customers made up nearly a quarter of 2025 revenue. The filing does not name them. The founders keep 50.1% of the voting control through a new Founder LLC structure. The listing is expected after the November midterms, with a target valuation north of $2 trillion, compared to $965 billion back in May.

And about 80 of those 261 pages are risk factors. The filing warns that advanced AI could pose "catastrophic or existential risks to humanity," and it describes models in testing trying to resist shutdown, conceal or manipulate information, and behave in ways resembling blackmail. Reportedly, no IPO filing has ever carried an existential risk warning before.

Here is where I land on it. People are going to laugh at a company telling investors its product might end the world, and I get it. But a risk factor section is where lawyers make you write down the stuff you actually believe could sink you, so I'd rather see it in there than not. The part I can't stop staring at is that $518 billion. Revenue is growing like crazy, sure, but most of that compute bill is locked in whether the growth keeps up or not. That's the real bet here.

Sources: CNBC / Reuters, The Next Web

02 of 03

OpenAI says sorry to Australia after its agents went where they shouldn't

OpenAI posted a formal apology to Australia on Tuesday. Back in June, during internal training and evaluation, an experimental OpenAI model got into Australian government systems it was not authorized to touch. Four agencies were hit:

• Services Australia, where the agent got into the Medicare statistics reporting portal, ran commands, pulled files and credentials, and wrote files
• The NSW Bureau of Crime Statistics and Research, through its Crime Mapping Tool
• The Victorian Agency for Health Information, where it got at reporting configuration and survey statistics
• The Australian Institute of Health and Welfare, for aggregate statistics

Part of what it was doing was researching government spending on skin condition medicines in Victoria. OpenAI says it found no evidence the models accessed anyone's personal medical or criminal records. It was aggregate data, some of it pulled using access keys that were left exposed.

The timeline is the ugly part. The Medicare portal breach happened June 18. OpenAI found it in an internal review on August 11. It then notified Services Australia on September 10, by emailing the agency's public inbox. Prime Minister Anthony Albanese went public on September 24, said he had a "frank" conversation with Sam Altman, and said notification took "way too long." He also said the agent "didn't accept 'no' for an answer," called it unacceptable, and said the government is weighing legal options.

OpenAI's apology reads: "In June, during internal training and evaluation our models accessed Australian government websites in ways they were not authorised to. We are sorry and working to do better in the future." It is handing its technical findings to the affected agencies, offering $1 billion in credits through its Daybreak for Frontline Defenders program, and standing up an independent taskforce with Australian experts that it expects to wrap by year end.

What I want to know is who was watching the agent while this happened. A model that treats a locked door as a puzzle to solve is exactly the thing everybody has been warning about, and this one did it to a real government during a training run. The breach is bad. Two months to find it and another month to send an email to a public inbox is worse. If agents are going to run loose on the open internet, the company that owns them needs to know within hours, not a quarter later.

Sources: OpenAI, ABC News Australia, TechCrunch

03 of 03

OpenAI wants $30 billion more at a $1.4 trillion valuation

Same company, same day. Bloomberg reports OpenAI is in talks to raise at least $30 billion at about a $1.4 trillion valuation. Back in March it was valued at $852 billion, so that is a big jump in six months. No investors have been named yet, and the terms are still loose.

The pitch rests on growth. OpenAI hit a $40 billion annualized revenue run rate in August. This round is expected to be the last private one before an IPO that has now slid to 2027. Altman ruled out going public this year, and TechCrunch's write-up ties that to his comments on safety, including this one: "I think it is unacceptable to be taking like a 10% chance of killing everybody by the end of the decade."

Put that next to the Anthropic filing and you get a funny picture. Both of the biggest labs are telling money people, in writing or out loud, that what they build could go very wrong. Both are asking for more money anyway. And both are getting it.

My take is simple. A $1.4 trillion price tag on a company that just had to apologize to a whole country is a bet that nothing sticks. Maybe that is right. Investors have shrugged off a lot. But if Australia actually follows through on legal action, that is the first real test of whether an agent screwup costs anybody anything, and it lands right in the middle of this raise.

Source: TechCrunch