Liquidation Preferences at a $965B Mark: Who Is Paid First if AI Lists Low

Anthropic's $65 billion Series H set a $965 billion mark in May 2026, and that number says nothing about who gets paid first. The liquidation preference attached to each round decides that in a sale. In a listing, an IPO ratchet can do the same job. Neither appears in a funding announcement, and together they decide whether employees are made whole.

The short answer

In a sale below the last round price, the newest preferred investors usually take their money back first, then earlier rounds, then common holders and employees. In an IPO, preferred normally converts into common and the preference disappears, unless the investor negotiated an IPO ratchet that issues extra shares when the listing price falls short.

What the $965 billion mark does and does not tell you

The $965 billion figure is a post-money valuation. It takes the price paid per share in the Series H and multiplies it by every share outstanding, common and preferred alike, as though they were identical.

Preferred shares carry contractual rights that common shares lack. The most important is the liquidation preference, a claim to be repaid before common holders in a sale or wind-down. A 1x non-participating preference means the investor takes the greater of its money back or what its shares would fetch as common.

So a post-money valuation prices the newest, most protected share and applies that price to the least protected one. Employees hold the least protected one. Our view is that the headline mark is closer to a ceiling on what common stock is worth than an estimate of it.

The vocabulary of participation, multiples and anti-dilution is covered in the primer on reading an AI cap table before a down round. This post does something narrower. It takes the round sizes actually reported for Anthropic, adds terms disclosed in a real S-1 and a Delaware court record, and works out who is protected at each exit price.

The reported sequence is short and steep. Anthropic raised a $13 billion Series F at $183 billion in September 2025. It raised a $30 billion Series G at $380 billion on 12 February 2026, led by GIC and Coatue, according to Crunchbase News. Then came the $65 billion Series H at $965 billion on 28 May 2026.

That last round was led by Altimeter Capital, Dragoneer, Greenoaks and Sequoia Capital, each putting in more than $2 billion, Bloomberg reported. It included roughly $15 billion previously committed by cloud providers, $5 billion of it from Amazon. The valuation rose more than fivefold in nine months, which is exactly the condition in which terms start doing work the price cannot.

Liquidation preference math at Anthropic's reported rounds

Anthropic's preference terms are not public. Its draft S-1 was filed confidentially on 1 June 2026, and no public registration statement had appeared as of this post. What follows is a model built on reported round sizes with stated assumptions, not a description of Anthropic's charter.

The preference stack is small next to the mark

Crunchbase put Anthropic's total raised at nearly $64 billion through the Series G. The Series H added $65 billion, but about $15 billion of that was earlier committed money. Depending on how much of it was already counted, new money was $50 billion to $65 billion.

That gives a total of roughly $115 billion to $130 billion. This is our arithmetic from reported figures, and an upper bound on the preference stack, because not every strategic dollar is necessarily preferred stock. Against a $965 billion mark, the stack is about 12% to 13% of the valuation.

Under a standard 1x non-participating structure, common holders receive nothing in a sale only if the company sells for less than about $115 billion to $130 billion. That is a fall of roughly 87% from the mark. For Anthropic employees, the preference is not the risk worth losing sleep over. The price is.

The last round is protected across the widest range of prices

The more useful question is who absorbs a smaller fall. A non-participating investor is indifferent between its preference and converting at exactly its own post-money valuation. Below that price, the newest investor always does better taking its money back, and the difference comes out of everyone ranked beneath it.

The table runs a sale at four prices. Ownership is derived from each round's size and post-money valuation, which gives Series H about 6.7%, Series G about 7.4% and Series F about 6.1% after dilution.

Assumptions, not disclosed terms: 1x non-participating preferences, no dilution after Series H, no option pool modelled. "Everyone else" lumps founders, employees and every round before Series F together. Those earlier rounds hold preferences of their own, so the real outcome for common stock is worse than this column shows.

Read the $500 billion row. A sale at roughly half the mark leaves Series H whole at $65 billion, against $33.7 billion if it had converted. The $31.3 billion difference is paid by Series G, Series F and everyone else, who collect less than their percentage ownership would suggest.

The bottom two rows show why seniority matters only at the extremes. Whether rounds rank equally or the newest ranks first changes Series F's outcome by $7 billion at a $100 billion sale. At every higher price in the table, seniority changes nothing.

Liquidation preference protects in a sale. The IPO ratchet protects in a listing.

Everything in the table assumes the company is sold. Anthropic is preparing to list, and a listing works differently.

Automatic conversion erases the preference at listing

Venture charters usually convert all preferred stock into common automatically when an IPO closes. Many apply this only to a qualified IPO, defined by a minimum price per share or minimum proceeds. Below that bar, conversion needs a vote of the preferred holders, which hands them a veto over a cheap listing.

Once converted, preferred stock is common stock. If our model company listed at $500 billion, Series H would hold about 6.7% of a $500 billion company, worth $33.7 billion. That is a 48% loss on paper, with no preference to fall back on.

This is why the listing price, not the preference, is what late-stage investors negotiate hardest when an IPO is likely. A preference that converts away on the day of listing is protection for the wrong exit.

How an IPO ratchet rebuilds the preference in share form

An IPO ratchet adjusts the conversion price of a series when the offer price falls below a threshold. The holder receives enough extra common shares to restore a target value at the offer price. Those extra shares dilute everyone else.

Put a 1x ratchet keyed to the Series H price into the model, and a $500 billion listing gives Series H shares worth $65 billion. A round carrying both a 1x preference and an IPO ratchet is protected at every exit price above the stack, by either route.

The third row is the one that surprises people. In a plain IPO with no ratchet, the newest investors share the pain evenly with employees. Our view is that anyone weighing an offer from a late-stage AI company should ask about IPO ratchets before asking about the preference, because the listing is the likelier exit.

An IPO ratchet in a real S-1: ServiceTitan's compounding clause

The clearest disclosed example in recent software listings is ServiceTitan, a vertical software company that went public in December 2024. Its Series H was a $365 million round raised in November 2022.

The clause, as filed

The S-1, filed in November 2024, set the Series H conversion to adjust if the IPO price came in below a threshold. The threshold was $84.5712 per share, accreting at 11% a year and compounding quarterly from 22 May 2024. The new conversion price was the IPO price multiplied by $84.5712, divided by a ratchet adjustment denominator.

The accretion is what made it unusual. Every quarter of delay after May 2024 raised the bar, and the dilution with it. Bill Gurley wrote on X that "a 'compounding ratchet' sounds painful (it is!)" and told founders to steer "WAY, WAY clear" of investors who ask for one. The S-1 did not disclose which investors held the term, TechCrunch reported.

What it did at a $71 offer price

ServiceTitan first marketed shares at $52 to $57, raised the range to $65 to $67, and priced at $71 on 11 December 2024. It sold 8.8 million shares for about $625 million. The stock opened and closed its first day at $101, about 42% above the offer.

That reads as a clean success, and the ratchet was still live. $71 was below the $84.5712 starting threshold before any accretion. By the formula, the Series H conversion price fell to $71 or lower, which means at least 19% more shares than the original conversion. That figure is our arithmetic from the disclosed formula, not a company disclosure.

Before pricing, Wing VC estimated that the adjustments across Series F, G and H would lift issuance from about 11 million shares to about 15 million, roughly 4% of extra dilution. That estimate assumed a lower price than the final $71, so treat it as an upper reading.

A headline of "priced above range" and a ratchet that paid out are entirely compatible, and only the prospectus tells you which happened. Meritech's Alex Clayton told TechCrunch that more of these terms would surface as companies funded in the low-rate years went public. His other line is the useful one: "You can't hide these details in an S-1".

When the preference stack exceeds the sale price: the Trados record

The sale case has its own primary document, and it is a Delaware court opinion rather than a filing. SDL bought Trados, a translation software company, in July 2005 for $60 million. The sale triggered a liquidation preference of $57.9 million for the preferred holders.

A management incentive plan took the first $7.8 million of the proceeds. The preferred holders received $52.2 million, short of their preference. Common holders received nothing. Without the incentive plan, common would have received $2.1 million, according to the Court of Chancery's 2013 opinion.

What Delaware says directors owe common holders

The common holders sued. The court applied entire fairness, its strictest standard, because a majority of the board that approved the merger was not disinterested and independent. The directors had to prove the deal was fair to common, rather than the plaintiffs proving it was not.

The directors won. Vice Chancellor Laster held in August 2013 that the sale was entirely fair, because the common stock had no economic value before the transaction. The holding is summarised on the Harvard Law School Forum on Corporate Governance. The process was criticised and the outcome stood.

A Delaware court will examine how the board decided. It will not rewrite a preference the company agreed to, and it will not award common a value the price never created.

Conditional money: AI round terms that sit outside the charter

Preferences and ratchets live in the certificate of incorporation. The largest AI rounds have added a second layer of protection that lives in the investment agreements instead.

OpenAI is the cleanest case. SoftBank led a $40 billion round in March 2025. Its portion could be cut to $20 billion if OpenAI did not complete its restructuring by the end of 2025, as widely reported at the time. The investor's protection was the option not to fund, rather than a claim on proceeds.

OpenAI completed its recapitalisation on 28 October 2025, TechCrunch reported. The OpenAI Foundation holds 26% of OpenAI Group PBC plus a warrant, Microsoft holds about 27% on an as-converted basis, and employees and investors hold the remaining 47%. How that equity was priced at $852 billion in March 2026 is examined in the breakdown of OpenAI's valuation against its run rate.

Anthropic's Series H carries a milder version of the same pattern. About $15 billion of the $65 billion was money previously committed by cloud providers. Committed money that arrives in stages gives the investor information before it pays, which is protection by timing.

The listing itself may make all of this academic. Reuters reviewed Anthropic's confidential draft prospectus in late September 2026, and Fortune reported that the company is aiming for a valuation above $2 trillion, with $518 billion of future cloud, computing and infrastructure obligations. What those commitments mean for the business is covered in the reconciliation of Anthropic's reported losses and the analysis of its enterprise revenue ahead of the IPO.

At $2 trillion, every series converts gladly and no preference or ratchet matters. The terms only bite if the price disappoints. That is precisely why they are negotiated when nobody expects it to.

Employees appear to agree with the optimistic reading. Anthropic's February 2026 tender offer was set at a $350 billion valuation, and buyers hoped to buy $5 billion to $6 billion of stock, Bloomberg reported. Most employees kept their shares. Holding common through a listing is a bet that the offer price clears every threshold above you.

The wider market is not cheap enough to make that bet free. Cooley's Q2 2026 data shows the median pre-money valuation for Series D and later rounds fell from $2.4 billion to $600 million in a single quarter, while down rounds rose to 12.1% of deals. The concentration that produced these prices is mapped in the analysis of AI's share of venture capital.

Where this argument is weakest

The largest weakness is the one stated at the top of the model. Anthropic's preference terms are not public, so every number in the waterfall table rests on assumed terms. The real charter could rank Series H senior, could include a ratchet, or could be cleaner than our assumption.

Second, at Anthropic's scale the preference barely matters to employees. A wipeout needs a fall of about 87%. The exposures that genuinely move an employee's outcome are the offer price, the lock-up, and the tax treatment of options exercised at a high valuation. A post focused on preferences can make a remote risk look near.

Third, the Trados case was decided 20 years after the deal, about a small company with a stack nearly equal to its price. Its relevance to a company valued near $1 trillion is about doctrine, not likelihood.

The strongest counter-case

Investors have a fair reply. An investor asked to pay a price it considers high is entitled to protection, and a founder who wanted the higher number chose that trade. Clayton made the point directly when he pushed back on calling ServiceTitan's terms dirty: it was "an agreement between two parties".

A round that closes with a ratchet is still a round, and the alternative may be a lower price that dilutes employees more today. We would still rather see the terms disclosed next to the valuation. That disagreement is about transparency, not about whether investors are allowed to protect themselves.

Frequently asked questions

Where to start: read the document, not the press release

If you hold common stock or options in a late-stage AI company, send one email this week. Ask for four numbers: the total liquidation preference stack, the seniority order of each series, any qualified IPO price threshold, and whether any series carries an IPO ratchet.

Then order the charter yourself. In June 2024, researcher Zach Stein-Perlman obtained Anthropic's certificate of incorporation and its past versions from the State of Delaware and published his reading of its governance terms. The same filing carries the preference and conversion terms for any Delaware company, and it costs a small state fee.

With those numbers, rebuild the waterfall table above for your own company at three prices: the last round, half of it, and the preference stack itself. The middle row is the one that decides what your equity is worth.

Related reading