Hook
Two numbers shipped in the same table, and they cannot both be right.
Anthropic's pre-IPO market on Hyperliquid, deployed September 12, was reported with a valuation of $2.159 trillion. OpenAI's market on the same venue was reported at $164 million. Anthropic's last primary round put it somewhere in the $180โ200 billion range. OpenAI's mark is in the hundreds of billions. So one number is roughly eleven times too high and the other is roughly three orders of magnitude too low, and no single convention produces both.
Reversing the stack to find the original intent: these are not the same field. One is a valuation. The other is almost certainly a notional or open-interest figure wearing the wrong header. The label is wrong, and every downstream conclusion drawn from it โ "Anthropic sentiment retraced," "the market is pricing a correction" โ is noise dressed as signal.
But the mislabel is not the interesting part. The interesting part is that a table like this can be mislabeled without anyone catching it, because the product has no ground truth to check it against. A pre-IPO perpetual has no spot market. Its settlement price is declared, not discovered. A declared price cannot be arbitraged. Which means it cannot be defended.
Context: What HIP-3 Actually Ships
Strip the branding and here is the mechanism.
Hyperliquid runs its own L1 with an on-chain order book, sub-second matching, and a margin engine that liquidates positions against a mark price. HIP-3 is the permissioned extension that lets third parties โ deployers โ spin up their own perpetual contract markets on top of that engine. The deployer defines the contract: underlying, tick, leverage caps, funding parameters, oracle source, and settlement rule. Hyperliquid supplies the matching, the risk engine, and access to the HLP vault as a liquidity backstop.
Entropy is the deployer here. It did not build a chain, did not write a novel AMM, did not invent a new primitive. It created a market. The technical work is parameter selection plus a settlement specification, and the parameter selection is roughly the same work a perp listing team does at any exchange.
So what is the instrument? A cash-settled perpetual whose underlying is "the valuation of a private company." You post USDC, you take a long or short, and at some defined point the position resolves against a number. There is no delivery. You never receive a share, a SAFE, an SPV interest, or a contractual claim on anything Anthropic or OpenAI has issued. You receive or forfeit cash based on where a reference number lands.
The Anthropic market reportedly carried about $28.19 million in notional and roughly $6.74 million in traded volume. The OpenAI market showed $7.67 million in volume, with a notional figure that is almost certainly the mislabeled $164 million.
Those are the real numbers. Everything else in the launch material is narrative.
That matters because the launch framing โ "IPO approaches" โ implies proximity, causality, and a market forming around an upcoming event. None of that is in the code. There is no disclosed IPO timetable, no disclosed authorization from either company, and no disclosed settlement methodology. We have a contract that pays out on a number, and we do not know who computes the number.
I have been here before. In 2017 I spent six weeks inside the 0x v0.9.9 exchange contracts and found three unsigned integer overflow paths in fillOrder, and what made them exploitable was never the arithmetic โ it was the mismatch between what the calling code believed about a value's units and what the callee did. Unit ambiguity is the oldest bug class in finance, and it is struct-generic: it survives every rewrite.
Core: The Oracle Is the Product
A perp is an oracle with a funding rate attached
Start from first principles, because the marketing skips this.
A perpetual future has three moving parts: the mark price, the funding rate, and the settlement price. On a normal crypto perp, all three are tethered to an external spot market by arbitrage. If the perp trades above spot, someone buys spot, shorts the perp, collects funding, and waits. The trade is boring, scalable, and mechanically enforced. That arbitrage is not a feature of the venue. It is the venue's price integrity. Remove it and the mark price is just an opinion with a liquidation engine bolted on.
Now delete the spot market.
Private company equity does not trade continuously. There is no order book for Anthropic shares you can hit at 3 a.m. Secondary transfers happen in negotiated blocks, settle over weeks, require issuer consent in most structures, and are frequently capped by transfer restrictions. There is no borrow. There is no short. There is no arbitrage path between the on-chain perp and any deliverable asset, because the deliverable asset does not exist in transferable form.
This is the deterministic failure map, and it is short. Four modes:
- Oracle declaration. The settlement value is set by the deployer's chosen reference. There is no market process that can contradict it, because no competing price exists at that moment.
- Stale marking between rounds. Private valuations update every six to eighteen months. Between updates, the mark is a frozen number. Funding, liquidation, and PnL all accrue against a constant.
- Discontinuity at resolution. When the reference does move โ a new round, a tender, an IPO price โ the move is a step function, not a drift. Step functions are where liquidation engines eat their own tail.
- Book fragility. With $6.74 million in volume, the book is thin enough that a single participant can be the top of book. Thin books do not dampen moves. They amplify them and then report the amplification as price discovery.
Mode one is the one that matters.
The arithmetic of the contradiction
Back to the table, because it is diagnostic.
Take the reported Anthropic figure at face value: $2.159 trillion. Against a primary valuation of roughly $183 billion, that is a factor of about 11.8. Now take OpenAI at $164 million against a plausibly reported valuation in the hundreds of billions โ call it a factor of a thousand or more in the other direction. Two fields, one header row, opposite error directions.
That pattern is not a typo. It is a schema collision. The Anthropic field is a valuation expressed in the wrong unit scale, or inflated by a notional-per-contract multiplication that was never divided back out. The OpenAI field is a notional or open interest that got promoted into a valuation column. Somewhere a spreadsheet consumed two different structs and rendered them under one label.
Truth is not consensus; truth is verifiable code. And here the code โ or at least the reporting layer โ is not verifiable in either direction. So the honest response is to discard both fields and use only the internally consistent ones: notional and volume.
Do that, and the market size becomes legible. $28.19 million of notional on a company whose last private round was north of $180 billion is approximately 0.015% of the underlying. The traded volume is smaller still. This is not a market forming around an IPO. It is a rounding error with a ticker.
That is the information gain the launch material did not provide, and it is the number that should anchor every subsequent read of this venue.
Who takes the other side
Every trade needs a counterparty. On a Hyperliquid perp the counterparty is either another user or the HLP vault acting as market maker of last resort.
Now do the hedging arithmetic. A market maker quoting a normal perp hedges delta on spot, or on a correlated perp, or on a basket. What hedges a pre-IPO valuation exposure?
You cannot short a private company at size. You cannot buy the underlying to cover a short. You cannot construct a delta-neutral position because there is no delta to neutralise โ there is a step function with an unknown step date and an admin-set size. So the market maker's position is not a hedged spread. It is an outright directional position in a number, held by a vault whose depositors were told they were providing liquidity.
Abstraction layers hide complexity, but not error. The abstraction here is "market maker." The reality is "underwriter of an unfalsifiable number." Those are different risk objects and they belong in different risk buckets.
And note the conflict geometry. The same operator that defines the reference valuation also operates the venue on which that valuation is traded and, through the vault, may be the effective counterparty to the flow. That is not decentralised finance. That is a bookmaker who also owns the scoresheet. The only thing preventing extraction is the deployer's reputation and its staked HYPE โ which is a bond, not a proof.
The abstraction leaks at the validator set
HIP-3 markets inherit their performance and their security from Hyperliquid's L1. That inheritance is real: the matching engine is fast, confirmation is sub-second, and the order book is a genuine on-chain primitive rather than a batch-settled derivative of one.
It is also where the L1 analogy breaks. Hyperliquid's throughput advantage comes from a validator set that is deliberately small and geographically concentrated โ that is what makes sub-second finality affordable. The decentralisation claim lives at the interface layer; the operational reality lives at a handful of nodes. This is the same structural pattern I documented in 2021 when I traced roughly 40% of popular ERC-721 collections back to a small number of centralised IPFS gateways and watched the market keep calling it trustless ownership.
The deployer's control surface compounds it. In a builder-deployed market, the deployer sets leverage caps, funding parameters, the oracle reference, and the delisting path. Governance over that market is, functionally, a single key. Hyperliquid's own risk engine may cap losses, but it does not adjudicate what the reference valuation is.
There is one honest mitigant: HIP-3 deployment generally requires the deployer to stake HYPE, often through an auction mechanism. That creates a real capital commitment and a real exit cost. It is a bond against bad behaviour. It is not a substitute for disclosure, and it does not bind when the operator's payoff from mispricing exceeds the stake.
The registration question
This is the largest exposure in the structure, and it is not technical.
Anthropic and OpenAI equity are securities. A cash-settled contract whose value derives from the price or valuation of a security, and which is not itself the security, sits on a fork: either it is a security-based swap, or it is an event contract. Both forks are regulated. Neither fork is registered here, so far as any public disclosure shows.
Run the elements. Money is invested. It is a common enterprise in the sense that all long holders share the same reference. Expected profit is the entire point. And the value depends predominantly on the efforts of others โ the company's operating performance and, more acutely, the operator's declaration of a reference number.
The closer precedent is not the tokenised-equity experiments of 2020, which at least attempted to hold a claim. It is the event-contract category that has been litigated and re-litigated over the past several years. A contract that resolves on a declared real-world outcome, cash-settled, without a deliverable, is a synthetic event exposure. Venues offering these have spent years in jurisdictional fights about whether they need to be designated contract markets.
Two further obligations were never addressed in the launch material. First, whether Anthropic or OpenAI have consented to the use of their valuation as a settlement reference โ the trademark and endorsement surface is non-trivial. Second, the entity structure. The front end typically geo-blocks the United States. The protocol layer does not. That gap is the standard evasion architecture, and it is the reason "decentralised" and "outside the regulator's reach" are two different claims.
In my Terra/Luna post-mortem I spent four weeks locating the exact iteration where the peg-breaking loop became numerically irreversible. The lesson I carried out of that work was not about algorithmic stablecoins. It was that the point of irreversibility is always structural and usually disclosed in advance, in the design, by someone who assumed nobody would read it. A settlement rule that does not exist in public is the same class of thing. It is not a missing document. It is an unhedgeable design choice.
Contrarian: The Consensus Read Is Wrong Twice
The consensus read is that this is an early RWA wedge โ a beachhead for tokenised private equity, currently tiny, therefore low risk, therefore worth watching.
Both halves of that are wrong.
First, the size does not reduce risk. It concentrates it. A thin market with a discovered price is inefficient. A thin market with a declared price is a lever. With $28.19 million in notional, the capital required to move the book is trivial โ and more importantly, no capital is required at all, because the number that resolves the trade is set by policy, not by flow. Small size limits contagion to the venue. It does not limit it to the trader. Anyone reading "tiny market" as "contained risk" has inverted the threat model.
Second, the systemic risk is not the perp. It is what the perp gets quoted as. If a $2.159 trillion figure ever escapes a launch table and lands in a journalist's column, a secondary desk's internal memo, or a research note, it acquires a kind of reality. It becomes a shadow reference: a number that is wrong, that everyone privately knows is wrong, and that nonetheless circulates because it is the only continuously updating price attached to that name. Pre-IPO valuations are already a soft consensus of round marks, tender prices, and analyst opinions. Adding a thin on-chain print does not improve that consensus. It adds a fifth, noisier source that is structurally easier to move than any of the other four.
That is the actual failure mode: not a trader losing money on a bad settlement, but a fabricated quote becoming citable. The abstraction leaks outward.
And here is the pre-mortem, since 2022 taught me to write them before the fact rather than after. Anthropic closes a round at a valuation that differs from the market's stale reference. Funding has been accruing for months against a constant. One side of the book is deeply underwater on a mark that was never real. The settlement rule is invoked for the first time under adversarial conditions, by an operator who has never published it. A dispute follows. Hyperliquid, weighing regulatory optics against a market generating a rounding error in fees, delists. The deployer's staked HYPE takes the hit. The narrative rotates to the next unlisted name within a quarter.
Nothing in that sequence requires malice. It requires only that the settlement rule stay unpublished until the moment it matters.
Takeaway: What to Verify Before the Next Round
Three disclosures would change my assessment, and none of them are expensive to produce.
One: the settlement specification, published, with the oracle's identity, the data sources, the averaging window, and the dispute path. If that document does not exist before Anthropic's next primary round, the venue is not a market. It is a quote with a liquidation engine.
Two: a named counterparty for the other side, with a stated hedge. If the HLP vault is taking pre-IPO valuation risk unhedged, that is a depositor-facing disclosure that belongs in the vault docs, not in a launch thread.
Three: a jurisdictional answer. Which regulator, which registration, which exemption. "Front-end geo-block" is not an answer; it is a delay.
The interesting variable is not Anthropic's eventual IPO price. It is whether a declared number can survive contact with the first party that has an incentive to litigate it. Until then, treat every print from this venue as an opinion โ and mark it, as the reporting already has, with the units unspecified.