Anthropic pre-IPO perps:
long xyz / Entropy, short OKX / Binance

Short answer: the exchanges convert so that "1 contract = 1 real share", and after listing the contract price ties directly to the Nasdaq price. A share price only applies to shares that already exist and can trade, so only shares outstanding can be used.

① Rule text. OKX says it does a one-time rebase on the "actual total share count" (text); Binance says it adjusts contract size on "the amount disclosed in S-1/A", with no adjustment under 3% (text). Across all of both venues' pre-IPO contract announcements and rules, "fully diluted" has never appeared, and options, RSUs or convertible notes have never been mentioned.

② The prospectus defines this number itself. The S-1 sentence "shares outstanding after this offering are based on X Class A and Y Class N shares" is followed directly by an "exclude" list: 1.774B options (average exercise price $2.91), unvested RSUs, RSUs granted after 9/30, shares issued for acquisitions after 9/30, and shares reserved under the incentive and matching plans. The exchanges use the number in that sentence.

③ Real precedents. SpaceX had 492M options and about 150M RSUs; OKX's 12.52B = pre-offering shares outstanding, Binance's 13.08B = shares outstanding + IPO new shares (excl. greenshoe), and neither counted options or RSUs; Oura and Quantinuum were the same (Section 6).

④ Why including them would be wrong. Options and RSUs are "shares that may exist in the future": options only become shares when employees pay to exercise, and RSUs only vest after a service period. Converting on a fully diluted count would make each contract represent less than 1 real share, the contract price would not match the share price on listing day, and the exchange would have to adjust again.

No. The announcement states "this rebase is unrelated to the actual share count of OpenAI or Anthropic" (text).

That rebase was a value-neutral unit change: contracts held × 10, price ÷ 10, with position value and account equity unchanged. The purpose was to make the single-contract price easier to trade (from about $1,700 down to about $170).

The same announcement also says: "Once the relevant companies officially file their S-1 … and disclose their actual total share count, OKX will execute a separate, supply-based rebase (so that 1 contract corresponds to 1 share)". What actually determines the conversion is that supply-based rebase, using the actual share count disclosed in the prospectus; the ratio will then be about 12.7B ÷ 10B = 1.27.

Short answer: the CEX valuation itself is not unreasonable; the problem is the denominator. The same valuation divided by different share counts can give per-share prices more than 30% apart.

① Valuations in the media and in funding reports are on a fully diluted basis. Series H post-money $965B ÷ $58.90 per share (post-split) = 16.38B shares, close to the company's fully diluted count (16.937B) and far above shares outstanding (Section 5). CEX traders are quoting on this basis too.

② But before listing, CEX contracts are converted to a per-share price using post-offering shares outstanding. $1.63T ÷ 12.7B shares ≈ $128/share; the same valuation divided by the fully diluted ~16.9B is only about $96/share. After listing there is a single real share price, so after conversion the CEX contracts are bound to be more expensive than the real price by this denominator gap, about 33% (16.937 ÷ 12.7).

③ Looking at per-share prices directly is clearer. xyz already quotes per share, currently about $106; the CEXs, after conversion, are about $126–128, 19.5–21% more. This spread has nothing to do with whether the valuation is high or low: whether the listed price is $80 or $190, a share-matched position locks in this difference (denominator chart).

Base case: post-offering (incl. IPO new shares, about 12.7B), the same as Binance.

① Precedent. SpaceX's first S-1/A with numbers gave only the pre-offering count (12.520B), with the new-share count still blank; OKX announced 2.6 hours later, executed at 12.52B the next day, and made no further adjustment. Oura's S-1/A gave pre-offering, new and post-offering counts all at once, and OKX used the post-offering 321M directly. This shows OKX used pre-offering only because the new-share count had not been disclosed, not because its basis excludes new shares (Section 6).

② This time it will most likely all come at once. In a sample of 33 large US IPOs from 2024–2026, about 30 filled in pre-offering, new and post-offering counts all at once in the S-1/A with the price range; only SpaceX and Waystar disclosed in two steps.

③ The difference. If Anthropic, like SpaceX, first gives only the pre-offering count (about 12.3B) and OKX executes within hours, the locked-in return on short OKX would be about 4 percentage points higher. This deviation favors the short and is not included in this note's base case.

Short answer: Entropy is cheaper per share than xyz and is one of the better places for the long, but it carries three risks xyz does not: rule risk, funding cost, and share-count uncertainty. It suits part of the long, not all of it.

① Rule risk. Entropy issues its own contracts on Hyperliquid and sets its own rules (a HIP-3 deployer). Its listing conversion rule changed three times in one week, each time without an announcement: in the early hours of 10/30 it first added "convert at the fully diluted share count after listing"; at 04:37 on 10/31 it added that convertible notes count in full at their fixed conversion price regardless of ownership caps; at midday on 11/3 it changed to using the prospectus figure "Total Class A and Class N Common Stock (As-Converted, Fully Diluted)" from The Offering section directly, excluding the greenshoe and control shares with no economic rights (rule text). The 11/3 version hands the number to the prospectus, leaving Entropy little room to change the basis again.

② Share-count basis. The company's table totals assume full conversion of Amazon's notes, so the "convertible notes excluded" scenario is essentially gone. The remaining uncertainty is the figure in the table: options under the treasury stock method are only about 50M below the full count, and adjustments such as vested matching shares, Pledge issuances and acquisitions add back about 50M; on 9/30 data it is estimated at about 16.6–16.8B, essentially the same as the 16.73B working value. The biggest variable is the number of IPO new shares: every additional 100M shares adds 100M to Entropy's count, which favors the long. Once the S-1/A with a price range is filed, Entropy will publish its estimated share count.

③ Funding. Long Entropy pays ongoing funding, and the more the spread converges, the higher the funding rate; see "Is the funding cost high?" below. In a scenario where the CEXs converge toward xyz, funding may eat most of Entropy's advantage over xyz.

④ Other mechanics. Before listing, Entropy's mark price and oracle are clamped between 300 and 4,200; from 3 calendar days before the announced listing date, the oracle switches to the 1-hour average of the mark price; after the first official opening trade, the oracle switches to "share price × share count ÷ 1B". If there is still no listing by 2028-08-18, it settles under the No-IPO rule.

Because its contract is quoted on valuation from start to finish. In Entropy's pre-IPO contracts "a price of $1 represents $1B of the company's fully diluted valuation", following the private-market convention of quoting fully diluted valuation. The docs also explain that Hyperliquid does not support scaling a contract by a constant, so after listing it keeps quoting valuation, with price = share price × fixed share count ÷ 1B (Source [12]).

Since it quotes fully diluted valuation before listing, the only way to transition smoothly after listing is to convert using the fully diluted count; otherwise the price would jump at the moment of listing. OKX's and Binance's contracts also quote valuation before listing, but they rebase to "1 contract = 1 share" and then tie directly to the share price after listing, so they use shares outstanding.

Result: one real share corresponds to 1 ÷ 16.73 units on Entropy and 1 contract on the exchanges. The two share-count bases differ by about 4B shares, which is exactly why Entropy is cheap per share and the CEXs are expensive per share.

First, the conclusion: Amazon's convertible notes are not excluded entirely. The portion that can convert within the 12% ownership cap at listing (about 510M shares) becomes real stock and is already in the exchange basis of about 12.7B. What is excluded is the roughly 1.33B shares above the cap that remain as notes (the prospectus calls them "rollover notes"). There are four reasons:

① Legally it is not stock. The unconverted portion is a debt the company owes Amazon, with no voting or dividend rights, and it cannot be traded. The prospectus says the notes convert up to the cap at listing, and the part that cannot convert "will remain outstanding and will automatically convert … at the end of each fiscal quarter following the completion of this offering", i.e. it stays as notes and converts a further portion at each quarter-end after listing as the cap allows; whatever still cannot convert at maturity is repaid in cash or offset against commercial payments.

② The prospectus's own "shares outstanding after this offering" only counts the portion within the cap. The sentence "shares outstanding after this offering are based on…" lists the items it reflects, including only "the Convertible Notes Conversion", defined as "shares issuable to Amazon … without causing it to exceed the applicable ownership cap". In the "as-converted, fully diluted" table on page 21, the conversion at listing is listed under "Pro Forma Adjustments", while the rollover notes are listed separately under "Additional Adjustments", together with "RSUs granted after 9/30" and "vested matching-program shares". They are only added when computing the fully diluted figure, which shows that the company itself does not treat them as outstanding stock.

③ Exchange rules and precedents use the "outstanding" basis. Binance adjusts on "the amount disclosed in S-1/A"; OKX rebases on the "actual total share count". In the SpaceX precedent, both venues' numbers matched the prospectus's "outstanding" sentence, and options, RSUs and reserved shares were not counted (Section 6). Unconverted securities are likewise "shares that may exist in the future" and by the same logic will not be counted. Also, the exchanges rebase only once, before listing, while the rollover notes only convert gradually at quarter-ends after listing, so they do not yet exist at the time of the rebase.

④ From a contract design standpoint, counting them would be wrong. The rebase exists to make "1 contract = 1 real share", so the contract price ties directly to the share price after listing. Counting the 1.3B unconverted shares would make each contract represent less than 1 share, the price would not match after listing, and another adjustment would be needed.

Compare Entropy: Entropy uses the fully diluted basis, and its rule explicitly counts convertible notes in full "regardless of ownership caps". So the same 1.33B shares are counted by Entropy but not by the exchanges, which is one source of the roughly 4B gap between the two bases.

The prospectus says "unless Amazon elects in its sole discretion to waive the limitation", so Amazon does have that right. If it waived, 1.33B shares would become real stock at listing, and the exchange basis would jump from about 12.7B to about 14.0B. This is the largest single risk on the short side. We judge the probability to be low (under 10%), for the reasons below; the first two come from the prospectus text, the last two are inferences:

① Amazon has never waived it so far. From the prospectus: "The portion of each note that could not convert without exceeding this ownership cap was reissued as a rollover note". Every past conversion reissued the portion above the cap as a rollover note rather than converting it to stock. The existence of rollover notes itself shows that Amazon chose not to waive at every previous opportunity. The conversion terms at the IPO are identical.

② Waiving the cap gives Amazon no economic benefit. The conversion prices are fixed at $2.09 and $5.05, far below the IPO price, so converting early or late yields the same number of shares and the same upside. After listing, the remainder converts automatically within the cap at each quarter-end, and since total share capital keeps growing after listing (new shares, employee stock and so on), the 12% allowance grows with it; whatever still cannot convert at maturity is repaid in cash or offset against commercial payments. For Amazon this arrangement has no downside. In addition, large holders usually face a lock-up after the IPO, so converting early buys no liquidity; and the shares received are non-voting Class N, so it buys no control either.

③ It would invite unnecessary regulatory trouble (inference). The US FTC's 6(b) study of Big Tech AI investments in early 2025 covered the Amazon–Anthropic partnership. When the UK CMA reviewed the partnership in 2024, it found that Amazon had no material influence over Anthropic and therefore cleared it as not a merger. Voluntarily raising its stake above 12% would signal to regulators that it wants more influence and could reopen scrutiny. The 12% line was very likely designed for this reason.

④ Both sides likely prefer the status quo on governance (inference). Anthropic is a public benefit corporation whose board is controlled by the Long-Term Benefit Trust. Issuing non-voting Class N shares to strategic partners and setting ownership caps are both meant to keep large holders as financial investors. Waiving the cap is Amazon's unilateral right, but the two have a large compute partnership, and breaking the default arrangement at the IPO would do the relationship no good.

If it does happen, we will see it in advance. A waiver has to be decided at the conversion before listing, and would show up directly in the numbers in the next S-1/A: the amount on the "Convertible Notes Conversion" line would be noticeably larger, or "shares outstanding after this offering" would exceed about 12.3B plus new shares by more than 1B, or wording such as "waive" would appear. The exchanges only rebase after seeing that S-1/A, so the market would see it first. Even if it happened, with an exchange basis of about 14.0B the short-side return would be about 12 percentage points lower, still inside the break-even point (about 15.3B).

The prospectus says Google may elect to buy up to $10B of Class N shares at the IPO price, completed concurrently with the IPO; if exercised, it reduces the note capacity available under the "Google 2026 NPA" accordingly.

Whether it is counted: no source says so yet. If it closes concurrently with the IPO, the prospectus's "shares outstanding after this offering" will likely include it, since it will be real stock by then. At a price of about $100–110, that is about 90–100M shares.

Impact: about 100M more shares in the exchange basis, about 0.9 percentage points less locked-in return on the short side; whether it is counted in Entropy's basis depends on whether The Offering table includes it. The magnitude is small and does not change the conclusion. The next S-1/A will show which line it sits on.

The pre-listing outstanding portion (about 10.73B as of 9/30 + 847M Public Benefit Fund + about 510M Amazon conversion within the cap + net RSUs) is essentially fixed in the S-1/A. The remaining variables, ranked by their effect on the exchange basis (the short side):

• IPO new shares: the biggest variable. Reports say the raise may match or exceed SpaceX's $86.2B, and the WSJ reported "over $100B", including sales by existing holders (including co-founders); the split between new and secondary shares has not been disclosed. Every additional 100M new shares adds 100M to the exchange basis and lowers the short-side return by about 0.9 percentage points; this note uses 400M, with a scenario range of 400M–1B.
• Google concurrent purchase: about 90–100M shares; see the previous item.
• Whether Amazon waives the cap: low probability but the largest impact (about +1.3B shares); see two items above.
• RSU tax withholding rate: RSUs vesting at listing are counted net of tax; a withholding rate of 30% instead of 40% would add tens of millions of shares.
• Options exercised before listing: these move from options to outstanding, possibly tens of millions to 100M shares.
• Whether OKX uses pre-offering or post-offering: if pre-offering, short OKX gains about 4 percentage points; see the item above.

Taken together, the exchange basis realistically lands around 12.5–13.5B; break-even is about 15.3B (long xyz) and about 15.6B (long Entropy), leaving a wide margin. The only way to approach that line is "Amazon waives the cap + far more new shares than expected" happening together.

Low on the short side and for long xyz; not low for long Entropy, and it rises as the spread converges.

① Short side and xyz. Binance is about +0.005%/8h (received by shorts, about 5.5% annualized), OKX about 0. xyz's pre-IPO funding rate multiplier is only 0.005 (1% of normal), so the long's cost is close to 0.

② How Entropy's funding rate is computed. It settles hourly, = 0.125 × the standard Hyperliquid formula (hourly average premium, plus an interest term clamped at ±0.05%). The premium is measured on the "impact price" relative to the oracle, and the oracle = 95% × internal price (5-minute average of the order-book mid) + 5% × external price. So the premium has two components:

• Internal price lag: when io rises quickly, the book price is clearly above the 5-minute average and the premium can spike to about 1%; on the way down it goes negative. This part is symmetric and averages close to 0 over time, but it makes individual hourly funding very volatile (annualized rates from 14% to 60% have appeared within the same day).
• External price anchor: about 5% × (io ÷ external price − 1). This is the persistent part and determines the long-run cost.

③ What the external price is. Entropy's docs only say it is a weighted aggregate of private-market data, with undisclosed sources. Backing it out from 8 days / 183 hours of data and tick data, the external price is almost exactly Binance's valuation quote (deviation about 0.1%, volatility about 0.06%), roughly 90% Binance + 10% OKX.

④ Key inference: the closer to convergence, the higher the funding. After listing io converts at about 16.73B shares, while Binance's and OKX's valuation quotes convert at 12.7B shares, a structural gap of about 32%. Right now the CEXs are about 20% more expensive per share than io, offsetting most of that gap. For every 1 percentage point the CEXs converge toward the real share price, the external price falls about 1%, io's premium over the external price rises, and annualized funding rises by about 6 percentage points:

⑤ As a cost. About two weeks remain until the oracle switch (3 days before the announced listing date). On notional: if the CEX spread holds the whole time, about 1.4%; if it converges all at once at the next S-1/A, about 4.4%; with linear convergence, about 3.9%. Meanwhile, Entropy's per-share advantage over xyz moves between about 2–4% with the market. So on the base-case convergence path, the extra advantage of long Entropy may be eaten by funding. A more conservative approach: when Entropy becomes more expensive relative to xyz (its discount narrows), switch the Entropy long back to xyz.

⑥ When it ends. From 3 calendar days before the announced listing date, the oracle switches to the 1-hour average of the mark price, the external price drops out, and funding falls back to near 0. If the listing date is announced late, the high funding lasts longer.

① Volatility. Daily moves of 5–10% are common in pre-IPO contracts. The two legs are on different venues, so one side shows an unrealized gain and the other an unrealized loss; the losing side needs enough margin on its own and cannot rely on the gain to offset it. We suggest under 3× leverage on both sides, with margin to absorb an adverse move of 30% or more.

② An exchange rebase changes the tier. OKX tiers by number of contracts held, and the maintenance margin rate is set by the tier of the whole position. At the rebase the contract count is multiplied by about 1.27, so the same position may be pushed into a higher tier and the maintenance margin rate rises with it. For larger positions, split across several sub-accounts, each tiered separately. Binance uses progressive tiers with a deduction per tier, so the effect is smaller.

③ Rebase-day details. Per OKX's rebase announcement: take-profit and stop-loss orders are cancelled and not kept; unrealized P&L is settled into realized; after execution there is a post-only period, with its length set by OKX depending on liquidity. Do not rely on stop orders around the rebase.

④ One-sided risk. With a share-matched hedge the position is broadly neutral to the share price, but if the two legs fill out of sync there is brief one-sided exposure; building in slices and executing both sides together keeps it within one slice.

This is the risk with the longest holding period. Each venue handles it differently:

• trade.xyz: if there is still no listing by 2027-01-19, the grace period ends and it cash-settles at the time-weighted average price over the whole period (timeline).
• Entropy: if the listing is delayed or withdrawn, it keeps running on the pre-listing method; if there is still no listing by 2028-08-18, it settles under the No-IPO rule.
• OKX, Binance: no explicit long-term rule has been published; the contracts are expected to keep trading, but the share-count conversion would remain unresolved.

A delay does not make the spread disappear, but it lengthens the holding period, during which you bear funding, margin usage and the risk of rule changes on each venue. If an IPO withdrawal is announced, CEX prices could swing sharply, the conversion logic would lose its basis, and the position would no longer be a locked-in return. With the S-1/A already publicly filed and the roadshow reportedly starting the week of 11/9, this risk is low, but RW (withdrawal) filings on EDGAR are worth watching.