ANTHROPIC FILES THE PAPERS — $1 TRILLION AI LAB TELLS THE SEC IT IS READY FOR MAIN STREET
Less than a week ago, Anthropic convinced the private market it was worth $965 billion. That apparently was not enough. On Monday the company behind Claude quietly submitted a confidential draft registration statement to the Securities and Exchange Commission, meaning they are officially going public. Not someday. Now. Well, soon enough that they had to file the paperwork, which in SEC terms is about as official as it gets before the actual listing.
Let me put this in perspective. These guys burned through billions building what they called an AI safety lab, spent years telling everyone they were the responsible adults in the room, too principled to rush, too careful to race. And then in the span of roughly four weeks they closed a $65 billion funding round, got a papal audience at the Vatican, watched Andrej Karpathy defect from OpenAI to join them, released Claude Opus 4.8, and now filed for an IPO at a valuation closing in on one trillion dollars. For a group that built its reputation on caution, they are moving at an absolutely unhinged pace.
The IPO season we are walking into right now is genuinely something. Anthropic is filing. OpenAI is expected to file. SpaceX is out here targeting a $2 trillion valuation, which is to say Elon’s rocket company wants to be worth more than every car company on earth combined. Three of the most consequential private companies of the last decade are all lining up to go public at the same time, in the same market, and they are all selling different flavors of the same thesis, which is: artificial intelligence is going to change everything, and you should give us money before it does. Whether they are right depends entirely on how you define everything.
What makes Anthropic’s filing particularly interesting is the timing. They just closed the biggest private AI funding round in history a few days ago. They did not need to go public for cash. They filed because there are strategic and competitive reasons to become a public company when your main rival is OpenAI and your main rival is also planning to file. Two near-trillion-dollar AI labs racing each other to Wall Street while simultaneously racing each other for talent, compute, and enterprise contracts is either going to be the most exciting thing to happen to public markets in a decade, or the most expensive lesson in collective optimism since 2000. Possibly both at once.
The S-1, when it drops publicly, will actually be worth reading. We will learn how much Claude earns in real dollars, who holds the voting power, what the genuine risks are as the company sees them, and where the money is actually going. Right now the pitch is essentially “Claude is very popular and the AI boom is real.” The SEC disclosure will force something more specific. And that is always when things get interesting, because the difference between “we project explosive growth” and the numbers underneath that projection is exactly the thing the market will spend the first six months of trading trying to understand.
The paperwork is in. The curtain is going up. And if you thought the last five years of AI coverage were intense, wait until the two largest AI labs are both public companies competing for the same investors while also competing with each other for everything else. It is going to be a ride.
Source: TechCrunch
GOOGLE CALLS WARREN BUFFETT AND RAISES $80 BILLION — SUNDAR SAYS EVEN THAT MIGHT NOT BE ENOUGH
Alphabet, the company that owns Google, announced this week that it is raising $80 billion through a stock sale because apparently $180 to $190 billion in capital expenditure this year alone was not enough runway to build all the AI servers it needs. The round includes a $10 billion private placement directly with Berkshire Hathaway, which is Warren Buffett’s firm, meaning the most famous old-school value investor alive just wrote a ten-figure check to fund GPU clusters in 2026.
Let that sit for a second. Warren Buffett. The man who said for decades he only invests in things he understands, who famously watched the dot-com bubble inflate and pop from the cheap seats, who avoided technology stocks the way most of us avoid eye contact with strangers on the subway. That guy just handed Google $10 billion specifically to build AI infrastructure. When the world’s most celebrated cautious investor decides it is time to get in, you are either looking at the confirmation that this is real or the clearest possible top signal. Maybe both.
The structure of the raise is worth understanding. Thirty billion dollars in underwritten equity offerings, including preferred stock. Forty billion through an at-the-market program starting in Q3, meaning they will sell shares gradually into the open market over time rather than dumping it all at once. And the $10 billion Berkshire placement at a fixed price below where the stock was trading, because Buffett negotiated a discount, which is very on-brand for a man who has made his entire fortune by buying good things at slightly below fair value.
Sundar Pichai told the Google I/O audience last month that the company expects to spend between $180 and $190 billion on capital expenditure this year alone. That number deserves to be said out loud slowly. One hundred and eighty billion dollars. In a single year. On infrastructure. That is more than the annual GDP of roughly two-thirds of the world’s countries. That is also more than Google’s entire revenue was in 2019. They are spending in capital expenditure what used to be their total business, just to keep up with AI demand.
The reason they need an additional $80 billion on top of internal cash flows is simple: demand is outrunning supply. Google said explicitly that customer demand for its AI infrastructure is currently exceeding available capacity, which sounds like a complaint but is actually the best problem a company can have. The problem is that “building faster” costs money they have to raise from someone. That someone is now, partly, Warren Buffett.
The real question is not whether Google can build the servers. It can. The question is whether the return materializes before the market loses patience with a company spending a hundred and ninety billion dollars a year on what might still be the early innings of a long investment cycle. You do not raise $80 billion without expecting a massive return on the other side. The bet is that AI demand keeps growing. So far, everyone who made that bet has looked smart. The day that changes is going to be a very interesting day to be an infrastructure investor.
Source: CNBC
JENSEN HUANG CRASHED A RIVAL’S KEYNOTE AND TURNED A $250 BILLION COMPANY INTO A $330 BILLION COMPANY IN ONE SENTENCE
On Tuesday at Computex in Taipei, Marvell Technology CEO Matt Murphy was in the middle of his keynote when Jensen Huang, the CEO of Nvidia and the most powerful person in the semiconductor industry, walked onto the stage unannounced. The audience noticed. And then Jensen said, essentially: Marvell is going to be the next trillion-dollar company.
Marvell’s stock went up 32.5% that day. It was the biggest single-day gain in the company’s history. Their market capitalization went from roughly $250 billion to over $330 billion in one trading session. Jensen Huang said a sentence, and $80 billion of market value appeared. That is not a keynote cameo. That is a superpower.
To understand why this matters beyond the theater, you need to know what Marvell actually builds. They are not a consumer brand. You have probably never thought about them in your life. What they make is the networking and connectivity chips that let GPUs communicate with each other inside massive AI data centers. As AI workloads spread across thousands of chips working simultaneously, the bottleneck is often not the raw compute power but how fast those chips can share data with each other. Marvell makes the technology that solves that exact problem. They are not the engine. They are the roads between the engines.
Jensen knows precisely what he is doing by showing up on that stage. He is not doing Marvell any favors out of goodwill. He is sending a message to the market, to hyperscalers, and to his own partners: this infrastructure layer is essential to what Nvidia is building. Nvidia’s continued dominance depends on data centers being able to scale the interconnects that link their GPUs. Marvell builds those interconnects. The endorsement is also a statement of strategic interdependency.
There is a broader story here about the AI chip ecosystem maturing. For three years, the narrative has been almost entirely about Nvidia. Jensen. The H100. The Blackwell. The GPU shortage. The waiting lists. But data centers are not just GPUs. They are cooling systems, power infrastructure, memory, storage, and the networking that ties it all together. Marvell sits in that last category, and the market has been underpricing it relative to what hyperscalers actually need to spend on it. Tuesday was the day that changed.
Whether Marvell actually reaches a trillion dollars depends on execution, competition, and a market cycle that does not turn against capital-intensive infrastructure spending. But the fact that Jensen Huang spent his Tuesday morning flying to Taipei to stand on their stage and say that publicly tells you something real about where Nvidia thinks the demand picture is going. When the person who knows the most about AI hardware spending tells you a particular company is going to be worth a trillion dollars, the smart move is to at least take notes.
The trillion-dollar chip club used to be a party of one. Jensen is apparently handing out invitations now.
Source: CNBC
MICROSOFT LAUNCHES SEVEN AI MODELS IN ONE DAY AND QUIETLY STARTS SHOWING OPENAI THE DOOR
At Microsoft Build 2026 in San Francisco on Tuesday, the company did something that looked routine but was actually a significant statement: it introduced seven new in-house AI models under the MAI brand, covering coding, reasoning, image generation, transcription, and voice. All in one announcement. All built internally. All designed to run on Microsoft’s own Azure infrastructure without paying anyone else a licensing fee.
The strategic message is not subtle. Microsoft has spent years and somewhere above $13 billion becoming OpenAI’s most important partner, investor, and infrastructure provider. OpenAI’s models power Copilot. OpenAI’s technology is woven through Microsoft’s products from Word to Azure to GitHub Copilot. And now Microsoft is shipping its own models across those exact same categories. That is not an accident. That is a plan.
The headline model is MAI-Thinking-1, Microsoft’s first reasoning model, running on 35 billion parameters and built for high efficiency at low token cost. It is aimed squarely at the reasoning segment where OpenAI’s o-series and Google’s Gemini thinking models have been competing. The coding model, MAI-Code-1-Flash, went live in Visual Studio Code on the same day it was announced, which is unusually fast for an enterprise software company. The image model, MAI-Image-2.5, debuted ranked second on the Arena AI leaderboard for image-to-image tasks. For a company that was not building frontier models six months ago, those are respectable opening numbers.
Microsoft is not ending its OpenAI partnership tomorrow. The investment is still there, the relationship is still there, Copilot still runs on GPT in many contexts. But the direction is unmistakable. You do not build seven in-house models covering the same territory as your partner unless you are building the option to reduce your dependence on them. Whether this is about leverage in contract renegotiations, genuine confidence in their internal research teams, or both is probably a question only the top floors of Redmond can answer. The answer is probably both.
There was another announcement buried in the Build keynote that deserves more attention than it got: Majorana 2, Microsoft’s next-generation quantum computing chip, achieved an average qubit lifetime of 20 seconds with instances stretching to a minute. They claim it is 1,000 times more reliable than their previous generation and that they have a path toward one million qubits on a chip that fits in your palm. Their stated target for a scalable quantum machine is 2029. Quantum computing has been “just a few years away” for about two decades, so treat that timeline with appropriate patience, but a 20-second qubit lifetime is a real engineering milestone in a field where sustained coherence has historically been measured in fractions of a second.
The overall picture from Build 2026 is a company that has decided, very clearly, that it is not going to be left behind in this technology cycle the way it was left behind in mobile. Seven models in one day, a personal work agent called Scout, and an updated quantum roadmap. Microsoft is moving. The question for OpenAI is how it responds when its biggest customer and investor starts building the same things it sells.
Source: CNBC