Big Tuesday in AI. Anthropic finally put a number on itself, New York flipped the switch on its frontier AI law, and Google rolled out two more homegrown chips. Here is what actually happened and what I make of it.
01 of 03
Anthropic Puts a $1.4 Trillion Price Tag on Itself
Anthropic set its IPO range at $70 to $78 a share. It is offering 825 million Class A shares, which works out to somewhere between $57.75 billion and $64.35 billion raised, with an option for underwriters to buy up to 123.75 million more. At the top of the range, the company is worth $1.404 trillion on a fully diluted basis. Goldman Sachs and Morgan Stanley are among the banks running the deal, and CNBC reports it should start trading on Nasdaq under the ticker ANT around November 19.
On top of the public offering, Anthropic is selling $26 billion of non-voting Class N stock at the IPO price to seven companies: AMD, Broadcom, Google, Micron, Microsoft, Nvidia, and TSMC. Google and Nvidia are each putting in $10 billion. Up to 20% of the offering can go to retail investors.
The numbers from the filing are wild in both directions. Run-rate revenue hit $81 billion at the end of October. Q3 revenue was $17 billion, and nine-month revenue was $33.2 billion. The nine-month net loss was $104 billion. Compute and infrastructure commitments sit at $900.9 billion, and the company is holding nearly $132 billion in cash and short-term investments.
The founders are getting paid too. All seven co-founders are each selling 3,378,378 shares, roughly $250 million apiece at the $74 midpoint. That is Dario Amodei, Daniela Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and Chris Olah. They each also get a supervoting Class F share, which keeps a majority of the votes in founder hands, and the Long-Term Benefit Trust elects a majority of the board. Translation: if you buy ANT, you are along for the ride, not steering it.
Here is where I land. A $104 billion loss in nine months would sink most companies, and nobody blinks because the revenue curve is that steep. I do not think the founders cashing out a quarter billion each is a red flag. That is pretty normal for an IPO. What I want to know is how retail buyers feel about owning a company where the chip makers and cloud providers are also the investors and also the suppliers. That is a tight little circle, and it only stays comfortable as long as the growth keeps coming.
Source: TechCrunch, which links the amended S-1 on SEC EDGAR.
02 of 03
New York's RAISE Act Gets Teeth
New York opened the registration portal for its RAISE Act on Tuesday. The law covers frontier AI developers pulling in more than $500 million a year, which puts OpenAI and Anthropic squarely in the crosshairs. They have until January 1, 2027 to register.
Once they are in, covered companies have to do four things. Register with the state. Publish their safety frameworks on their websites. Put out public transparency reports on their models. And report critical safety incidents within 72 hours. Fines run up to $1 million for a first violation and $3 million for repeat ones.
The enforcer is a new state office called DIGIT, short for the Office of Digital Innovation, Governance, Integrity, and Trust. New York also launched the Center for AI Responsibility and Research at SUNY Binghamton, a $55 million outfit that will evaluate advanced AI systems and help DIGIT review those safety frameworks.
The fun comparison is California. SB 53 took effect in January and also requires published safety frameworks, plus whistleblower protections, but it gives companies 15 days to report a critical incident. New York gives them 72 hours.
My take is pretty simple. A million bucks is a rounding error for a company that just priced itself at $1.4 trillion, so the fine is not the real stick. The 72-hour clock is. That forces somebody inside these labs to decide fast and in writing whether something counts as a critical incident, and that paper trail is what regulators and lawyers will chew on later. I would rather have one federal rule than fifty state ones, but until Washington does something, states like New York are going to keep filling the gap.
Source: TechCrunch
03 of 03
Google Adds Two More TPUs to the Pile
Google announced two new chips in its eighth-generation TPU family, which first launched in April. The first is the TPU 8tx, designed with AMD and built off the TPU 8t. It is aimed at training and running models confidentially, with security built into the hardware so data and model weights stay isolated while in use. It also has a custom inference network so it can run reasoning alongside training, which Google says lets it distill and train small models on the fly. The pitch is that companies can train on their own proprietary data without shuffling it between devices.
The second is the TPU 8i Max, which packs two TPU 8i chips into one package. According to TechCrunch, that doubles memory capacity, processing power, and bandwidth, so models can hold more information while answering a query.
What I keep coming back to is the AMD part. Google designing a TPU with AMD tells you the custom silicon game is not just Google versus Nvidia anymore. It is everybody building partnerships to avoid paying the Nvidia tax. And the confidential training angle is aimed right at banks, hospitals, and anybody else whose lawyers have been saying no to cloud AI. If that hardware isolation holds up, it removes one of the biggest excuses enterprises have left.
Sources: TechCrunch and Google's announcement