What you buy when you buy Anthropic
Buying a stock means buying management's choices about how to run the business and what the business even is. That matters most with Anthropic. Its S1 filing should come soon, followed later by OpenAI's. Grok already showed its hand through the SpaceX IPO. All three run a "closed weight" model.
Closed weights hide how the AI reaches its answers. Open weights let you reach inside and change things. Picture asking for music: 128 beats, a female singer, drums, guitar, bass, and a synthesizer in the background. You might get back something like Aretha Franklin, Alicia Keys, or Olivia Dean. With an open-weight model you can adjust the sound directly, swap the language from English to French, or cut the drums out completely. You control the parts.
That control is why corporate America and Silicon Valley are shifting toward open weight for complex jobs. Companies want to see how the AI produced its answer. Anthropic does not offer that. Neither, for the most part, does OpenAI or the others. Open-weight Chinese models, very different from ChatGPT, Claude, and Grok, are winning real success on this exact point. Usage patterns are changing fast, and that change will hit these IPOs. These companies plan to sell what will be hundreds of billions of dollars worth of stock. They have to convince buyers that usage numbers, good so far, stay good in the future while the models stay closed.
Why the lead may hold anyway
AI moves on "what have you done for me lately." Two years ago OpenAI looked unrivaled. Now it has arguably slipped, and Anthropic has taken the lead in how people see the field, though OpenAI is still very impressive.
Anthropic's lead runs deeper than perception. I meet Anthropic customers daily, or at least most days, and have for weeks. Claude Code was a game changer. It lets people build their own software products, doing even more than Cursor.
Cursor shows how fast this can move. It went from under $500 million in yearly revenue six months ago to over $4 billion in projected annual recurring revenue now, growing as fast as any software company ever has. Elon Musk bought it for $60 billion, which may have been too much, but the business itself is fantastic. Both Cursor and Claude Code do the same thing: use AI to write code.
This reshapes how big the market can get. Total addressable market figures at IPOs have always looked silly and rarely turn out true. But AI agents could outnumber humans. More robots writing code than people writing code changes the size of Anthropic's business. Claude Code already builds agents that use Claude, creating a self-feeding loop of growth. Reports say Anthropic has at points reached profitability in ways OpenAI has not. It spends heavily on marketing and on everything else, because it is pulling in a lot of money and growing fast.
The demand case against a bubble
Supply constraints are the strongest argument against calling this a bubble. Hyperscalers report capacity selling out before it is even racked. ASML's latest numbers showed huge spending by chip makers, and ASML charges so much that its gross margins look ridiculous for a hardware maker. It can pick its price and sell its goods.
Taiwan Semiconductor is not thrilled with these high prices, per reporting from The Information, but pays them anyway, and so does Intel. If Taiwan Semi balks, Intel steps in. ASML's machines, arguably the most incredible machines in human history, sometimes cost more than $200 million each and sell like crazy. They sell because hyperscalers want the chips, which means the hyperscalers are growing.
Every point in the supply chain shows expectations for strong growth: hyperscalers, chip makers, chip equipment makers, all the way down to Anthropic. Nobody is saying demand is slowing. Everyone could be wrong, but probably not. Usage numbers keep getting bigger, forcing more spending to build capacity to meet demand.
There are worries. When OpenAI or xAI built out the Colossus data centers, Grok turned out to be poor and usage was terrible, whatever the S1 tried to say. But they rented that capacity out anyway, to Anthropic, Google, and others. Facebook has looked at building data centers and renting out the capacity it does not use. Across every part of the AI buildout, the pattern is buildout on top of buildout, with no sign of slowing.
Strike while the iron is hot
Bloomberg reported Anthropic is scheduling meetings, possibly targeting an IPO as soon as October. Should they move fast to get ahead of OpenAI? Yes, and that is worth interrupting a question over. That is how hot the iron is. They have to get out and take the deal, because the IPO window opens and closes, and outside threats can slam it shut.
The US is at war with Iran right now. That seems unrelated to AI, but it could close the AI IPO window. If the window closes, these companies lose much of their ability to raise money, which would hurt AI development broadly and US AI in particular. The cause might have nothing to do with the technology and everything to do with the wider economy. Talk of risk in private credit adds to that. Trouble in the overall market feeds straight into whether these IPOs can happen and whether people are willing to buy the new stock and place their bets beside these newish companies.
