ANTHROPIC PREPS THE BIGGEST TECH IPO IN HISTORY AS BANKERS FLOOD INVESTOR INBOXES AHEAD OF OCTOBER DEBUT
Here is a number for you: nine hundred and sixty-five billion dollars. That is what Anthropic was valued at when it closed a sixty-five billion dollar Series H round back in May, and now the company’s bankers are on the phone with every deep-pocketed fund manager on the planet, scheduling meetings ahead of what could be the biggest technology IPO in stock market history.
If everything goes according to plan, Anthropic hits the Nasdaq in October. Goldman Sachs, JPMorgan, and Morgan Stanley are leading the deal. The offering is expected to raise more than sixty billion dollars. Let that sink in for a second. Sixty billion dollars. In a single stock offering.
To understand how we got here, you have to go back a few years to when Anthropic was the scrappy underdog trying to make safety-focused AI while OpenAI hoovered up all the press. That reputation worked out pretty well financially. The company’s annualized revenue crossed thirty billion dollars earlier this year and has been growing roughly tenfold annually for three consecutive years. For comparison, it took Amazon eight years to hit thirty billion in revenue from scratch.
What is actually interesting about Anthropic going public is not just the size of it. It is what it says about the AI moment we are living in. When a company that has repeatedly told the world that its technology might be one of the most dangerous things humans have ever built then turns around and asks the public to buy shares in it, something philosophically weird is happening. The company filed a confidential S-1 with the SEC back in June, the kind of move you make when you are serious about this.
And Anthropic is serious. Very serious. The company signed a twenty billion dollar data center lease in Kentucky a few weeks ago, a twenty-year commitment to a physical facility that tells you everything about how long they expect to be in this business. They are also in early talks with Samsung to build custom AI chips, following the same playbook that Google and Apple used when they decided that buying chips from Nvidia was getting too expensive.
The question nobody is asking loudly enough yet is whether a near-trillion-dollar valuation actually makes sense for a company that is still years away from profitability. Gross margins are around forty percent right now, and the company thinks it can push those to seventy-seven percent by 2028. Maybe. But building frontier AI models is extraordinarily expensive, the competitive pressure from OpenAI, Google, Meta and a dozen Chinese labs is relentless, and the regulatory environment is getting thornier by the month.
None of that will stop the IPO from being a spectacle. When October rolls around, it will be one of the most watched stock debuts in years. And somewhere in San Francisco, Sam Altman is taking notes.
GOOGLE’S AI SEARCH TOLD KIDS EATING DISORDERS WERE NORMAL AND HANDED OUT DEEPFAKE INSTRUCTIONS. REPORT CALLS IT AN “UNACCEPTABLE RISK.”
The Youth AI Safety Institute at Common Sense Media just published a report about what happens when you give a child a Google account and let them loose on the company’s AI-powered search features. The headline finding is that the AI poses an “unacceptable risk” to children. That is not a phrase researchers throw around lightly.
Here is what they actually found. On test accounts set up for minors, Google’s AI search features failed to flag suicide risks in queries that would make any school counselor reach for the phone immediately. When test accounts asked about eating disorders, the AI responded in ways that normalized the symptoms rather than directing users to help. And in what might be the single most alarming finding in the report, the system provided actual instructions for creating deepfakes when queried by accounts registered as belonging to teenagers.
Let that last one sit for a moment. A teenager goes onto Google Search, which is the search engine that the vast majority of children in the English-speaking world use for homework, and the AI feature tells them how to make deepfakes. Not what a deepfake is. Not why they are harmful. How to make one.
Google has been pushing AI features into Search aggressively over the last couple of years because the company is terrified of losing its core business to AI-native search competitors. The pressure is real. But somewhere in the urgency to ship AI features and defend market share, apparently nobody ran a test asking “what does this tell a fifteen-year-old who says they have not eaten in three days?”
The thing about this report that makes it different from the usual tech-is-bad-for-kids discourse is that it is specific, methodical, and reproducible. They used actual test accounts for minors, ran actual queries, and documented the responses. This is not a vague moral panic. This is “we did the test and here is what happened.”
Google will almost certainly respond by saying its AI features are continuously being improved and that protecting younger users is a top priority. That is the statement they give. The statement they give every single time something like this comes out. And then a few months later another report comes out.
What makes this genuinely hard for Google is that the same AI system that tells a kid how to make deepfakes is also the system that finds them homework answers, explains historical events, and has become basically indistinguishable from their search experience. You cannot wall off “the dangerous parts” easily because the same model does all of it.
Regulators in the EU already have frameworks that could bite here, and the US is increasingly looking at children’s online safety as a rare area of bipartisan agreement. Congress has been circling this topic for a while. Reports like this one hand them a specific, documented set of facts to point at. This will not be the last anyone hears about it.
HACKER CRACKS SUNO OPEN AND FINDS YEARS OF STOLEN AUDIO FROM YOUTUBE, DEEZER, AND GENIUS HIDDEN IN THE SOURCE CODE
Somebody hacked Suno, the AI music generation startup that has raised hundreds of millions of dollars and built a product that lets anyone create a passable pop song in about thirty seconds. And what the hacker found in the source code is exactly what a lot of people in the music industry always suspected but could never prove.
The hack itself is a supply chain attack story. Back in November, someone managed to get into an employee’s credentials through a compromised piece of software in the company’s development pipeline. Once inside, they found source code. And that source code revealed something that Suno had never publicly disclosed: the system had been scraping audio from YouTube Music, Deezer, Genius, stock music libraries, and podcast RSS feeds to train its model.
Now, to be clear, this is exactly what the music industry has been alleging in lawsuits against multiple AI companies. The record labels took Suno and Udio to court last year specifically over training data, arguing that the companies had scraped copyrighted recordings without permission or payment. Suno has maintained that its training process was lawful.
The hack gives that lawsuit a potentially significant new dimension. Source code showing the mechanism of the scraping, if it can be authenticated and introduced as evidence, is a different category of thing than inference about what a company probably did based on the quality of its outputs.
What makes this story especially awkward for Suno is the YouTube angle. YouTube is owned by Google, and Google has been in the middle of a very public debate about AI training data rights. YouTube itself has rules against scraping its content, and the platform has been increasingly vocal about protecting creator content in the AI era. Scraping YouTube Music specifically, if confirmed, would represent a fairly brazen disregard for those rules from a company that presumably needs a functional relationship with the music industry to survive.
The startup has not commented specifically on the source code findings, which is the kind of non-denial that tends to make lawyers on the other side of a lawsuit smile a little wider.
Here is the wider context worth holding onto: Suno is not unique in this situation. The entire first generation of AI companies trained on data scraped from the internet, often without clear licensing arrangements, because that was the fastest and cheapest way to build capable models. The legal reckoning for that era is happening right now, across music, images, text, and code. Suno just happens to have had its internal records accessed by someone with a grudge and good timing. The music industry’s lawyers are having a very good week.
PALANTIR’S TOP TECH OFFICER SAYS CHINA BUILT ITS AI EMPIRE BY QUIETLY STEALING FROM SILICON VALLEY AND AMERICA IS ONLY NOW WAKING UP TO IT
Palantir’s Chief Technology Officer went on record this week with something that has been whispered in national security circles for a while but rarely said out loud by someone with a C-suite title at a major American technology company. China, he said, built a significant part of its current frontier AI capability by using the work of Silicon Valley AI developers without authorization. And that now poses a real economic threat to the United States.
The statement lands with some weight coming from Palantir. The company is not a neutral party in this conversation. It does billions in government contracts and has been positioning itself as the AI infrastructure provider of choice for defense and intelligence applications. But the CTO’s comments are also consistent with what American AI researchers and government officials have been saying privately for some time.
The specific concern is about distillation. When you train a smaller model to mimic the outputs of a large frontier model, you can capture a significant amount of the bigger model’s capability at a fraction of the compute cost. If Chinese labs trained on the outputs of American models in ways that violated terms of service, they effectively got access to years of American AI research investment for the price of some API calls. That is not a theoretical attack vector. It is a practical one.
DeepSeek’s R1 model was the moment this became a mainstream conversation earlier in the year. The model was extraordinarily capable for its reported training cost, and a lot of people in the American AI industry started asking hard questions about where exactly that capability came from. The OpenRouter data we have been watching over the last few months, which shows Chinese models now accounting for more than sixty percent of tracked global token volume, suggests those capabilities have since been deployed at a scale that should be making American labs nervous.
What makes the Palantir CTO’s framing interesting is the economic lens. Most of the public conversation about Chinese AI has focused on national security, on weapons systems and surveillance infrastructure. But the economic threat is actually the more immediate one for American companies. If Chinese AI providers can offer equivalent or near-equivalent model performance at dramatically lower prices, and they currently can, the question of how they got there has real commercial implications for OpenAI, Anthropic, and Google.
The policy response so far has mostly been export controls on chips. But chips are only part of the story. If the knowledge transfer happened at the model output level, through distillation and unauthorized API access, then chip controls are closing the barn door after the horse has already learned to do calculus. Nobody in Washington has a clean answer for that one yet.