Anthropic Readies a Fall IPO at a Reported $965 Billion Valuation
On September 12, 2026, Yahoo Finance reported that Anthropic is preparing to go public this fall. BlockMedia carried the story the next day with the figures that matter: roughly $130 billion raised from about 300 institu
On September 12, 2026, Yahoo Finance reported that Anthropic is preparing to go public this fall. BlockMedia carried the story the next day with the figures that matter: roughly $130 billion raised from about 300 institutional investors, and a valuation of $965 billion set in a May 2026 round. Bankers are reportedly floating $2 trillion as a listing number. Terms and timing can still change. What cannot change is the arithmetic already locked into the cap table, and what it says about where capital goes when money is cheap and the exit is a public market.
The Cap Table Behind the Number
Anthropic was founded in 2021. Its valuation history is a short, steep line. The company was marked around $61.5 billion in March 2025, then roughly $183 billion in the September 2025 round. The May 2026 mark of $965 billion is a fivefold move in eight months. If the IPO clears at $2 trillion, early money from 2021 and 2022 is looking at returns that do not fit on a normal venture chart.
Amazon is the largest strategic holder. Its disclosed commitments reached $8 billion by late 2024, structured as convertible notes alongside a commitment to use AWS Trainium silicon. Alphabet has put in several billion more across multiple tranches, starting in 2023 and adding roughly $1 billion in early 2025, while supplying TPU capacity. Neither company took a controlling stake. Both bought something more useful than control: a guaranteed customer for their own compute, booked as revenue on the way back in.
That circularity is the part worth staring at. A cloud provider invests $8 billion in a model developer, the model developer spends the proceeds on that cloud provider's chips, and the cloud provider books the spend as revenue growth that supports its own multiple. The cash makes a round trip. The revenue is real in an accounting sense and hollow in an economic sense, because it was funded by the recipient's own balance sheet. Nvidia, Microsoft, and OpenAI have run versions of the same loop. Anthropic's IPO will be the first clean test of what an outside buyer will pay for the output of that loop.
The 300 institutions in the source figure are the other half of the story. That list reportedly spans technology firms, asset managers, hedge funds, sovereign wealth funds, and Silicon Valley venture capital. Note what is missing from that description: households. The entire $965 billion of value was created in private markets, accessible to accredited and institutional capital only. Retail gets its turn at the $2 trillion mark, if that is where the book clears. This is not a conspiracy. It is the predictable result of a securities regime that has spent thirty years making private capital easy and public listing expensive.
One caveat on the headline number. It is not clear from the reporting whether the $130 billion is cumulative capital raised since 2021 or a single pre-IPO round. Anthropic's publicly disclosed rounds through 2025 total well under that figure. If $130 billion is cumulative and includes committed but undrawn compute credits, the number means one thing. If it is fresh 2026 money, it means something much larger, and it would rank among the biggest private capital raises in history. The prospectus will settle it.
Where $130 Billion Actually Goes
Frontier model training is not a software business in the old sense. It is a capital-intensive industrial business wearing software margins as a costume. The money goes to GPUs, custom accelerators, data center shells, transformers, substation upgrades, and multi-year power purchase agreements. Gross margins on inference are respectable. The capex line underneath them depreciates on a three to five year schedule, and the useful life of a training cluster is shortening as each hardware generation arrives.
This is where the bull case and the bear case actually separate, and it is not about whether the models work. Both sides agree the models work. The question is the denominator.
The bull case: enterprise adoption is compounding, coding and agentic workloads have found genuine product-market fit, and per-token costs have fallen by orders of magnitude since 2023 while usage has risen faster than prices have fallen. A company growing annualized revenue at triple digit rates with a defensible technical position and two hyperscalers as anchor customers can grow into almost any multiple. At $2 trillion, the buyer is underwriting a decade of that.
The bear case: depreciation. If a $965 billion valuation rests on $130 billion of capital that converts into hardware with a four year life, the company must generate enough free cash flow to refresh the entire fleet before the multiple can mean anything. Model capability is also converging. The gap between the best available model and the third best has narrowed at every generation. Commoditization at the model layer would push margins toward the cost of compute, which is to say toward the margins of the cloud providers who already own the compute and also own the stock.
Both cases are defensible. Neither is knowable from the outside before the S-1 lands. What is knowable is that the market will price it in an environment where the risk-free rate has been the policy variable, not a market outcome, for most of the period during which this valuation was built.
The Monetary Plumbing Underneath the Valuation
Here is the Austrian reading, and it is not a claim that Anthropic is a bad company.
Capital does not appear from nowhere. In a sound money system, investment is funded by deferred consumption. Someone saves, the savings are lent, and the interest rate is the price that clears the market between savers and entrepreneurs. That rate carries information: it tells builders how long a production process the economy can actually support.
Since 2008, that signal has been administered rather than discovered. US M2 sits around $22 trillion, roughly triple its 2008 level. Policy rates spent the better part of fifteen years near zero, then moved sharply, then moved again. Under those conditions, capital flows toward the longest-duration assets available, because their present value is the most sensitive to the discount rate. A pre-revenue frontier lab with a ten year payback is the longest-duration asset in the economy. It is exactly what an artificially suppressed rate structure funds first, and exactly what a normalized rate structure punishes first.
This is the Cantillon effect in its modern form. New money does not enter the economy evenly. It enters at specific points, through specific institutions, and whoever stands closest to the entry point gets to spend it before prices adjust. In 2026, the entry point is the balance sheet of a handful of hyperscalers and the allocation committees of 300 institutions and sovereign wealth funds. They bought at $61.5 billion and $183 billion. The public gets offered $2 trillion. Whether that turns out to be a good price or a bad one, the sequence itself is the mechanism by which monetary expansion transfers purchasing power upward. It happens whether or not anyone intends it.
Sovereign wealth funds deserve their own line. When Gulf and Asian state funds become anchor investors in American frontier AI, the capital carries a policy shadow. Those funds answer to governments. An IPO turns their private stakes into liquid positions in a company whose models will shape information access across multiple jurisdictions. Washington has spent four years restricting chip exports on national security grounds while allowing state-controlled capital into the equity of the labs those chips serve. That contradiction is unresolved, and a public listing makes it visible on a filing.
Bitcoin as the Control Group
Bitcoin had no IPO. That is not a slogan, it is the whole point, and it is the cleanest contrast available to this story.
There was no cap table, no preferred stock, no liquidation preference, no allocation to 300 institutions ahead of everyone else. Every coin that exists was either mined under published rules or bought from someone who mined it. Anyone could have participated in 2010 with a laptop. Anyone can verify the issuance schedule today with a full node and a few gigabytes of disk. About 20.1 million of the 21 million coins have been issued. The current subsidy of 3.125 BTC per block produces roughly 164,000 new coins a year, under 0.8 percent annual issuance, and that rate halves again in the spring of 2028. No committee votes on it. No allocation committee gets an early look.
Compare the two issuance regimes honestly. Anthropic's equity supply expanded from a founding cap table to 300 institutional holders through a sequence of private rounds, each one diluting the last, each one priced by negotiation among parties who all benefited from the price going up. The dollar's supply expanded by roughly $14 trillion since 2008 through open market operations decided by a committee of twelve. Bitcoin's supply expanded on a schedule published in 2009 that has never been altered, enforced by tens of thousands of nodes that reject any block violating it.
I will take the position plainly. The AI buildout is real and probably productive. The monetary conditions that funded it are not sound, and the distribution of the gains reflects proximity to the money spigot more than it reflects contribution. Bitcoin does not fix the first problem. It does fix the second, because it removes the spigot. A monetary system where no one can print the entry point is a system where returns have to be earned rather than allocated. That is worth more over thirty years than any single company's multiple.
There is a narrower Bitcoin angle too. AI training centers and Bitcoin mines compete for the same three inputs: interconnect queue position, cheap baseload power, and cooling. Several large miners have already converted capacity to AI hosting because the revenue per megawatt is higher. A $130 billion capex program accelerates that competition. Expect miner hashprice pressure to come less from difficulty and more from electricity markets where AI is the marginal bidder.
What to Watch
The S-1 disclosure of compute commitments. Anthropic must disclose multi-year purchase obligations to Amazon and Google. Watch whether those commitments exceed trailing twelve month revenue. If they do by more than 2x, the company is a leveraged bet on its own growth curve, and the risk section will say so in language the marketing does not use.
Related party revenue. The filing will break out how much revenue comes from investors. If Amazon and Alphabet account for a meaningful share of both funding and bookings, the circular financing question moves from commentary to a line item, and short sellers will build their entire thesis on it.
The gap between the $965 billion mark and the listing price. A $2 trillion debut would mean the private mark was 52 percent below the public clearing price four months later. That is either evidence of extraordinary growth or evidence that the May round was priced by parties with an interest in a low mark. Watch for a listing below $965 billion. A down-round IPO from a company this prominent would reprice every private AI mark in the market within a week.
Lockup expiry, roughly 180 days after listing. If the IPO lands in October or November 2026, the first unlock hits around April or May 2027. Three hundred institutional holders with positions bought below $200 billion will face a decision at once. That date matters more to the stock than any earnings report before it.
Bitcoin's correlation behavior during the offering. Large IPOs pull liquidity. In 2024 and 2025, Bitcoin traded as a risk asset on short horizons and as a monetary hedge on long ones. If a $2 trillion listing coincides with a sharp Bitcoin drawdown, that is a liquidity event, not a thesis break. The thesis breaks only if the issuance schedule changes, and it will not.
Interest rate sensitivity. The single largest determinant of whether this IPO is remembered as 1995 Netscape or 2000 Pets.com is the discount rate over the next thirty-six months. Neither Anthropic nor its investors control that variable. A central bank does. That is the most concise argument for owning an asset whose supply schedule no central bank controls.
Go deeper: How Bitcoin Mining Works · How to Run a Bitcoin Node
Source: BlockMedia