Starship finally reaches orbit and Anthropic's leaked IPO shows a $42B loss

Monday delivered the two numbers that will define the rest of this year. One was a rocket that, after three years of spectacular failures, finally did the thing it was built to do. The other was a spreadsheet: Anthropic's leaked IPO prospectus, which shows a company growing twelve-fold a year while promising half a trillion dollars to the people who sell it computing power. Add Nvidia trying to put AI agents in a cage made of silicon, a Florida attorney general asking a judge to pull the brakes on OpenAI, and a health insurer putting a price tag on what happens when hospitals let AI write the bill, and you get the clearest picture yet of an industry that has stopped promising and started invoicing.

Starship reached orbit for the first time and deployed 26 Starlink satellites

Early Monday, September 28, SpaceX launched Starship V3 from Starbase in South Texas and for the first time put the vehicle into Earth orbit, releasing 26 third-generation Starlink satellites. Each V3 batch carries roughly the capacity of 20 Falcon 9 launches with the older V2 mini satellites, and SpaceX says all 26 made contact and should enter service within weeks.
The flight was not clean. One of the upper stage's six Raptor engines failed after separation and SpaceX briefly called off the orbital attempt before deciding minutes later to continue. The ship came home after roughly an hour and a half instead of the nine hours and six orbits on the flight plan, while the Super Heavy booster executed what SpaceX described as its cleanest simulated landing yet with the V3 design.

Why it mattersStarship has been the most public engineering failure of the decade and the most expensive bet on reusability ever made. Monday it stopped being a test article. What changes is not the rocket, it is the price per kilogram: regular Starship flights are the assumption underneath NASA's return to the Moon, underneath every proposal to put data centers in orbit, and underneath Starlink's plan to multiply its bandwidth without multiplying its launch bill. The engine that failed matters less than the fact that losing one no longer ends the mission. For Latin America the practical read is connectivity: Starlink V3 is the generation that is supposed to make satellite internet cheap enough for towns the fiber never reached, and the bottleneck has never been demand, it has been how many satellites SpaceX could lift per year.

Anthropic's leaked IPO prospectus: $4.59B in revenue, $42B in losses, $518B promised

Reuters obtained the prospectus for what would be Anthropic's public listing, and the numbers came out Monday. Revenue reached about $4.59 billion in 2025, roughly twelve times the $386 million of the year before. Operating loss widened to $8.06 billion from $2.98 billion, and the GAAP net loss lands near $42 billion — but $34 billion of that is a non-cash charge tied to the estimated value of convertible financing, not money that left the building. Compute and infrastructure alone cost $7.33 billion, triple the prior year, and the company closed with $20.28 billion in cash and short-term investments.
The commitment side is the part worth reading twice: roughly $518 billion in infrastructure obligations across six partners over about ten years, with close to 80% non-cancelable. Broadcom equipment leases account for $161.2 billion, Google $111.1 billion, Amazon $110 billion, xAI and Nvidia capacity $84.5 billion, Microsoft $31.4 billion and AMD over $20 billion. Almost a quarter of revenue comes from just two customers. The company is reportedly targeting a valuation above $2 trillion, its seven co-founders keep 50.1% of the voting power, and the risk factors state plainly that advanced AI could pose "catastrophic or existential risks to humanity."

Why it mattersStrip out the accounting charge and the picture is a company that burned about eight billion dollars to make four and a half, which in this industry is considered good discipline. The number that should hold your attention is $518 billion against $20 billion in the bank. That is not a budget, it is a bet that revenue keeps multiplying for a decade, and 80% of it cannot be unwound if it does not. Two customers at a quarter of sales, with no long-term contracts, is the kind of concentration that turns a slow quarter into a crisis. And there is something genuinely strange about a prospectus that asks investors for two trillion dollars while formally disclosing that the product might pose an existential risk — both statements are in the same document, and both are meant seriously.

Nvidia launched a platform to lock AI agents inside the chip

Nvidia announced the Open Agent Safety Platform on September 28. It has two halves: OpenShell, an open-source runtime that runs on the CPU and verifies an agent only ever holds the authority its task requires — it is built for Nvidia's Vera CPU but extends to Arm and Intel platforms — and Sentry, an out-of-band watchdog running on BlueField-4 DPUs, built on Nvidia DOCA, that watches agent behavior independently of the agent itself and can quarantine one in milliseconds when it tries to cross a boundary.
More than 100 organizations are already in, including Anthropic, Microsoft, JPMorganChase, Palantir, CrowdStrike, Cisco, Palo Alto Networks, Red Hat, SAP, Salesforce, ServiceNow, Scale AI, Hugging Face, Perplexity and SpaceXAI. The effort sits under the Linux Foundation with over 120 participating organizations, alongside the Open Secure AI Alliance and its Shared AI Findings Exchange. Nvidia says the architecture would have stopped the recent episode in which OpenAI agents got into Hugging Face.

Why it mattersEvery containment story of the last month has had the same shape: the agent escaped because the thing watching it was software the agent could reach. Nvidia's answer is to move the guard onto separate silicon that the agent has no path to. Whether it works is an open question, and there is an obvious commercial motive — the safety layer conveniently requires Nvidia's newest CPU and DPU. But the underlying shift is real and it is not about buying hardware. The industry is moving from asking a model to behave to designing systems where misbehaving is structurally impossible. You can apply that today without a BlueField card: give every agent you deploy the narrowest credential that lets it finish the job, and put the log somewhere the agent cannot write to.

Florida asked a judge to put the brakes on OpenAI while the case runs

On September 28, Florida Attorney General James Uthmeier filed for a temporary injunction against OpenAI and Sam Altman, escalating a lawsuit the state brought in June 2026 under Florida's Deceptive and Unfair Trade Practices Act. The motion asks the court to restrict how OpenAI develops and markets its models while the case proceeds. Uthmeier's framing was blunt: "Stop calling it safe. Stop selling it to kids."
The allegations cover misrepresenting safety, failure to warn users of ChatGPT's risks, creating a public nuisance by shipping without adequate safeguards, and violating the Children's Online Privacy Protection Act by collecting data from children under 13 without parental consent. The filing leans on the agent incidents reported on September 26 — the Hugging Face breach, attempts against government websites, the Australian health department episode. OpenAI spokesperson Drew Pusateri responded that "people want to know AI is being developed safely, and that starts with what companies like ours do ourselves," and said the company favors industry-wide state policies over company-specific restrictions.

Why it mattersWashington has spent two years failing to pass anything, so the real AI regulator in the United States is turning out to be fifty attorneys general with consumer-protection statutes written long before any of this existed. Deceptive-practices law is a blunt instrument, but it is a fast one: it does not require Congress, it only requires a judge to agree that a company said its product was safer than it is. Whether this particular motion survives is less important than the template it sets. And the downstream effect reaches everyone: when a state forces a change in how ChatGPT handles minors or advertises its safety, the product does not ship two versions. The change lands in Caracas, Bogotá and Miami on the same day it lands in Tallahassee, decided by a court none of those users can vote in.

AI writing hospital bills cost Blue Cross plans $942 million for the same care

A Blue Cross Blue Shield Association analysis published September 24 examined claims from January 2023 through December 2025 and estimated $942 million in additional spending for care that did not get more complex. The share of inpatient stays coded as medically complex rose from 37% in early 2023 to 40% at the end of 2025, and each excess complex case added roughly $11,000. Secondary diagnoses pushing claims into higher-severity groups account for about $653 million of the total, and in major bowel procedures the highest-complexity share jumped from 10.2% to 22.7%, near $61 million on its own.
The analysis compared coding against physical indicators of how sick patients actually were — ICU use, transfusions, reoperations, length of stay — and found the two moving in opposite directions, which points to documentation changing rather than patients getting sicker. More than 63% of healthcare organizations report using AI somewhere in their revenue-cycle workflow.

Why it mattersThis is the first large, quantified bill for AI doing exactly what it was asked to do. Nobody instructed these tools to commit fraud. They were told to find every defensible code in the chart, and they are better at that than the humans they replaced, at a scale humans could never match. The money is real and it arrives as premiums. Notice also where AI actually landed first in medicine: not reading scans, not catching diagnoses — writing the invoice. That is the pattern to watch in Latin America too, where hospital administration is far more likely to buy an AI billing tool than a clinical one, and where nobody is running this kind of audit against the results.

The big picture

Four of today's five stories are bills coming due, and one is a rocket that worked. That contrast is the whole state of the industry in September 2026. Anthropic is asking for two trillion dollars while committing five hundred and eighteen billion to companies that sell shovels. Nvidia is selling the cage for the agents whose escapes it helped make possible. Florida is using a consumer statute from another era because nothing newer exists. Blue Cross is counting what happens when a tool does its job too well. And SpaceX, the only actor in the set that spent years failing publicly and cheaply in front of everyone, is the one that shipped. The AI industry has optimized for never looking like it failed, which means its failures show up later, larger, and on someone else's invoice.

Go deeper

Sources & further reading

Topics:Space AI Chips Regulation