The bytecode never lies, only the intent does.
Here is a number that should not exist: $116.3 billion. That is the spread between what FTX's bankruptcy estate actually recovered from its Anthropic position and what that same position would theoretically be worth at a $1.5 trillion IPO valuation. The estate sold its entire stake in batches through 2024. The first tranche moved roughly $884 million across 24 buyers โ about $500 million to ATIC, an affiliate of Abu Dhabi's Mubadala sovereign fund; $100 million to Jane Street; $45 million to a Fidelity fund. Gross recovery landed near $1.3 billion; net realized, after the machinery of Chapter 11, closer to $800 million.
None of that gap came from a market call. It came from a legal status.
I have spent the last several years pulling apart failed protocols line by line, and the pattern is always the same: the failure is never where the marketing points. It is in the structure. FTX did not lose $116 billion by buying badly. It bought at the cheapest tier on the entire cap table and then lost the ability to hold.
Context: what is actually being priced
Anthropic intends to go public as early as mid-November. Market discussion puts the IPO at roughly $1.5 trillion, which is about 55% above the $965 billion mark set during its May financing this year. Those are press numbers, not filed numbers. Treat them as an oracle feed with no attestation โ useful for direction, useless for settlement.
What is documented is the entry ladder, and it is unusually clean:
- Spark Capital put in $75 million in 2023 at an implied ~$5 billion valuation, and holds the only known board seat among VCs. At $1.5 trillion, that position is discussed at around $10.9 billion.
- Sound Ventures (Guy Oseary and Ashton Kutcher) started at $35 million and added roughly $400 million; the stake now sits in the billions.
- Menlo Ventures has committed over $1 billion cumulatively, including a $500 million Series D entered in May 2023, with a stake discussed near $21.8 billion.
- Iconiq arrived in 2025 with cumulative investment above $6 billion, entering at an implied $183 billion valuation; the position is discussed near $26.4 billion.
- Sequoia Capital began investing this year, roughly $3 billion total for about 1% of shares, discussed near $15 billion. It had previously backed OpenAI and xAI.
- GIC, Singapore's sovereign fund, first invested in 2025, accumulating over $4.5 billion for roughly 1%, also discussed near $15 billion.
- Eric Schmidt put in tens of millions in 2021 for about 1%, discussed near $15 billion.
- FTX invested $500 million in 2021 for 13.56%, diluted to 7.84% by 2024.
Lightspeed holds under 2%. Google, Amazon, Nvidia, Microsoft, Founders Fund, Coatue and Altimeter are all in the book, but public disclosure is too thin to compute precise returns.
Read that ladder again and watch the implied entry valuations climb: $5 billion, then tens of billions, then $183 billion, then a $3 billion check for one percent that implies something in the neighborhood of $300 billion. The multiple compresses with every rung. That is not a scandal; it is arithmetic.
Core: the cheapest ticket, the worst outcome
The cost-basis anomaly
Do the division nobody in the coverage did. Five hundred million dollars for 13.56% implies a post-money value near $3.7 billion in 2021. That is the lowest entry valuation of any disclosed holder โ lower than Spark's $5 billion, lower than Schmidt's 2021 round, dramatically lower than Iconiq's $183 billion. FTX held the single best-priced position on the cap table.
And it is the only one that got liquidated.
This is the part that should bother anyone who thinks of venture equity as a return-generating asset. In a private cap table, the return is a function of two variables: entry price and holding period. Everyone obsesses over the first. The second is where the money actually lives, and the second is governed by legal architecture, not conviction.
Dilution is not a rounding error โ it is a mechanical leak
FTX went from 13.56% to 7.84%. That is a 42% relative reduction in ownership. Most coverage treats that as a footnote. It is the whole story of why early-stage percentages decay.
Private companies defend their percentages through pro rata rights โ the contractual option to write additional checks in later rounds to maintain ownership. Exercising a pro rata right requires cash and requires a legal entity capable of signing a subscription agreement. A debtor in possession has neither in usable form. Every dollar it wires to defend a position is a dollar a bankruptcy court must approve as a business judgment, against a distribution timeline measured in quarters.
So the sequence is deterministic: a bankrupt holder dilutes first and sells second. Bankruptcy does not merely force an exit; it guarantees the position deteriorates before the exit happens. Every edge case is a door left unlatched, and the unlatched door here was the pro rata clause nobody modeled as a liability.
If the estate had held its undiluted 7.84% to a $1.5 trillion mark, the arithmetic lands near $117.6 billion โ before any further financing dilution, which would have pushed the real number lower. Against $1.3 billion gross. The counterfactual is not a fantasy; it is a straight multiplication of documented shares against a discussed price.
The batch sale was a price-discovery event, and the estate was the price taker
Twenty-four buyers in the first tranche is not a negotiation. It is an auction conducted by a seller with a deadline and no BATNA. In any illiquid private market, a forced seller clears below fair value because the buyer set is deliberately small โ only institutions that can underwrite a multi-year lockup and survive the diligence window.
Look at who cleared. ATIC took roughly $500 million. If that tranche maps proportionally onto the block, the implied return on that single allocation approaches two orders of magnitude in about eighteen months. Jane Street took $100 million, Fidelity $45 million. These are not gamblers. They are buyers with balance sheets, legal teams, and a clear read on the IPO calendar.
The structural point: the transfer of value in a bankruptcy is not from creditors to the market. It is from creditors to whoever holds liquidity during the lockup. Sovereignty-grade capital won that auction because it was the only capital class that could wait.
The entry-price ladder as a risk map
Menlo entered a $500 million Series D in May 2023 at a moment when AI valuations were still being argued about. Iconiq entered at $183 billion. Sequoia and GIC came in this year and paid for roughly one percent with multi-billion checks. Their multiples are thinner not because they were late in any pejorative sense, but because the price of admission rose faster than the underlying enterprise value could compound.
Sequoia's book โ OpenAI, xAI, Anthropic โ reads less like conviction and more like portfolio construction under radical uncertainty about which lab wins. That is a rational hedge. It is also an admission that the terminal value of any single lab is not underwritable at this stage, only the category is.
I made a version of this argument about data availability layers: the infrastructure is real, the demand sizing is not. Ninety-nine percent of rollups do not generate enough data to justify a dedicated DA layer, and the market still prices every one of them as if they will. AI lab equity is running the same drill with a different asset. The market prices hope; the auditor prices risk.
What I would actually verify in the S-1
Three things, in order.
First, the fully diluted share count. The $1.5 trillion figure is being circulated as a valuation, but nobody outside the syndicate knows whether it is pre-money, post-money, or a headline that quietly excludes an option pool and outstanding SAFEs. These distinctions move the per-share number by double digits.
Second, the lockup terms and transfer restrictions on legacy holders. If insiders are subject to a 180-day lockup with an early-release trigger tied to trading price, the first unlock window is the actual price-discovery event, not the listing. I have watched token projects engineer unlock cliffs that made the TGE look like a rounding error; equity does the same thing with more paperwork.
Third, the composition of the buyer base at IPO. If the marginal buyer is a sovereign fund with a decade-long horizon, the clearing price reflects a discount rate nobody in the retail float shares. That is not fraud. It is a mismatched duration, and it will express itself as violent volatility the first time the growth narrative stutters.
Code compiles, but does it behave? A cap table compiles too. The behavior shows up at the first lockup expiry.
The parallel I keep coming back to
In 2026 I audited an AI-agent trading protocol where autonomous agents executed on-chain transactions from off-chain LLM outputs. The vulnerability was not in the model. It was in the oracle verification layer โ adversarial prompts could shape the price feed that the contracts trusted. We built a fuzzing harness to simulate AI-driven inputs and caught a path that would have drained roughly $10 million.
The lesson transferred cleanly. In any system where an unaudited external input is treated as ground truth, the attack surface is not the input's quality โ it is everyone's assumption that the input was verified. The $1.5 trillion Anthropic number is an off-chain feed. Every return calculation in the coverage, including mine, prices against it as if it were attested. It is not. It is a rumor with a decimal point.
Contrarian: the $116 billion number is a counterfactual, and counterfactuals are not losses
Everyone is writing that FTX missed $116 billion. That framing is emotionally satisfying and analytically lazy.
An estate's fiduciary duty is to maximize recovery inside a defined window, not to express a five-year view on transformer architectures. Holding a locked, illiquid, non-controlling stake through a 2024โ2026 horizon, with distribution timelines slipping and creditor committees watching, would plausibly have been a breach of duty. The sale was defensible. The structure that produced it was not.
The sharper question is ownership. If customer assets were commingled into a venture book with multi-year lockups, the Anthropic gain was never FTX's to keep. It belonged to depositors, and every headline about a missed fortune is a headline about someone else's money being discussed as if it were the company's.
And note what the cap table actually is: a registry gated by capital, not identity. The minimum check, not the passport, determined who got the return. That is the most honest compliance document in the industry, and it says nothing about KYC.
Takeaway: watch the unlock, not the listing
The listing date is noise. The 180-day unlock is signal. Watch whether legacy holders like Spark, Menlo, and Iconiq file to sell into strength, and watch whether sovereign holders with decade horizons simply do not. That divergence tells you what the smart money thinks the terminal value is, without anyone having to say it out loud.
And a forecast: the next forced-liquidation class will not be equity. It will be token treasuries of failed DAOs and insolvent funds, dumped into a buyer set that again consists of whoever can wait. The pattern is repeatable because the asymmetry is structural. Security is not a feature, it is the foundation โ and in a cap table, the foundation is the lockup.