The one-screen version: Anthropic stopped being a story you read about and became one you can price. On November 10 it set an IPO range of $70–$78 a share — 825 million Class A shares, a $57.8B–$64.4B base deal, plus a separate $26B private placement of non-voting Class N stock — implying up to about $1.404 trillion fully diluted at the top. Pricing is expected November 18, trading November 19, ticker ANT. The same day, TSMC's board approved buying shares in both Anthropic and OpenAI, and Mercor raised $750M at $20B to build the environments agents train in. One theme connects them: the stack is consolidating in plain sight. Here's what each asks of you.
1. The price is the least interesting number in the filing#
Everyone will quote the valuation, so get it out of the way: the $70–$78 range works out to roughly $916 billion to $1.02 trillion on outstanding shares, and up to about $1.404 trillion fully diluted if it lists at the top — down from the $1.5–2T talk that floated earlier this fall as the sector repriced. The offering is 825 million Class A shares (801.4M new, 23.6M from existing holders), on top of a separate $26 billion private placement of non-voting Class N stock — the same class Google holds a ~$10B option on.
The number that should actually move a founder is the shape of the sale. Per CNBC's reporting, the book is already covered at this range and Anthropic is targeting a retail allocation near 20% — far above a normal IPO — into the teeth of an AI-stock selloff. That's a lab being sold hard to the public.
You don't need an opinion on the stock. You need the filing. An IPO turns your most important private vendor into one whose books you can finally read.
So treat this as diligence, not a trade. When the final S-1 lands, read three lines: gross margin (does serving frontier models actually pay?), compute obligations (how much future cash is already promised to clouds and chipmakers?), and customer concentration (how much of that $16B-and-climbing revenue leans on a handful of accounts?). Those answers tell you how much to lean on Anthropic — far more than the ticker's first-day pop will.
2. Your chipmaker just became your fellow shareholder#
The quieter item is the structurally louder one. In a November 10 Form 6-K, TSMC's board approved up to $500 million to buy Anthropic common stock at the IPO price, up to $500 million in OpenAI preferred, and up to another $500 million to stand up a vehicle for strategic semiconductor and ecosystem bets. These are approvals, not closed trades — but the direction is unmistakable.
Follow the money and it runs in a circle. Nvidia invests in the labs that buy its chips; the clouds take stakes in the labs that rent their capacity; and now the foundry that fabricates the chips is buying equity in the labs at the far end of the same supply chain. Each link makes the others look better-capitalized; none of it adds a new, independent customer. For a founder, the lesson isn't scandal — it's concentration. The entire stack you build on rests on a short list of mutually invested companies and one US–China export regime. Keep your model and inference layers portable, the way we keep arguing, so that when one node in that circle has a bad quarter, your product doesn't.
3. The agent moat moved to the data layer — and it's now a $20B line item#
The third story is the one to actually act on if you ship agents. Mercor raised $750 million at a ~$20 billion valuation, with Nvidia, Salesforce and Zoom reported in, to expand the training environments agents learn in — the simulated workflows and expert-labeled tasks that teach a model to do a job rather than describe one.
That valuation is a bet on a shift founders can feel in their own stacks: as base models converge in raw capability, the scarce and defensible input is becoming the environment and evaluation data, not the model. The team that owns a crisp eval harness and a library of realistic tasks for its domain can swap models underneath and keep winning; the team that only picked a good model has nothing a competitor can't buy next week. This is why RL environments keep getting funded like infrastructure and why, when you compare agent frameworks this fall, the interesting differentiation is moving up the stack toward how they're trained and tested.
What to actually do. Three moves, one per item. Read, don't chase: when Anthropic's final S-1 prints, underwrite it like a vendor — margin, compute, concentration — and skip the first-day theater. Stay portable: let TSMC's buy-in reset any assumption that your silicon or model supplier is a permanent fixture; keep a second path warm. Own your evals: if agents are your product, invest in the environment and evaluation data that no model swap can take from you. The stack is consolidating. The founders who come out ahead are the ones who read it clearly and stay un-captured.
