A lockup expiry is a door, not a printing press. Shares that already existed become free to trade. The company does not have to issue a single new one for the float to jump, and the price then hangs on a blunt question: who shows up to buy what insiders are finally allowed to sell?
Arthur Hayes, co-founder of BitMEX, builds a much larger story on that ordinary mechanic in an essay called “D is for Dilution, ” as reported by crypto.news and restated by KuCoin. He is not focused on the models stalling. His near-term worry is that a lot of AI equity could become eligible to trade, lenders could flinch, and Bitcoin could feel the draft. If you are funding AI inside a real company, the crypto trade is the least useful part of that chain. Your inference bill, your automation contract, and your capacity date all sit downstream of someone else’s cost of capital.
Eligible supply is not a new issue
After a listing, employees and early backers are often barred from selling for a set window. When the window ends, the tradable float can expand even though the share count does not. Hayes reaches for the word “dilution.” Eligible supply is the cleaner description. Permission to sell is not a sale, and selling stock that already existed does not create a new issue.
On Anthropic, he is working from what both outlets call the company’s updated registration statement. Some restrictions, as they describe the filing, run 180 days after the offering. Others run to 365. From that schedule he estimates that early public investors could face what he calls 83 percent “dilution” by June 2027. KuCoin’s recap of the same essay softens the claim to “up to 83 percent.” Either way, the percentage is his reading of potential selling supply. The company has not reported that rate.
The trillion-dollar figure is the same math with a valuation plugged in. Eighty-three percent of a hypothetical $1.5 trillion starting value is about $1.245 trillion. Hayes says that is the extra buying required if the entire eligible slug were sold and the price did not move. The figure assumes the valuation and the sale. Anthropic has not said it must raise that capital.
I have not read that registration statement. The 180- and 365-day terms, and an infrastructure figure covered below, stay attributed to crypto.news and KuCoin. They are not line items checked here, and the coverage does not even pin the filing to a date.
Three names, three clocks
This is not one event landing on one Friday. Treat them as three stories with three levels of certainty.
Anthropic
This is the firmest of the three, and firm is a relative word. The outlets describe an updated registration and a share-release schedule measured from the offering, with Hayes pointing at June 2027 as the point by which early public holders could be free to sell in size. That is more than a rumor. It is still his interpretation of how much stock could become tradable, not a record of shares that will be sold.
SpaceX
In Hayes’s telling, SpaceX is already public, and he says the tradable count has grown considerably since its IPO. He estimates another 8.6 billion shares could become eligible by June 2027, on the order of half the existing share count. The estimate and the premise arrive together. The write-ups supply no listing date, exchange, or share-count source, so the 8.6 billion stays his. He also raises further pressure from a rumored combination with Tesla. No completed deal establishes that. He treats the company’s ability to generate enough cash for shareholder returns as a condition for avoiding pressure. He does not treat it as a result already shown.
OpenAI
OpenAI is the softest link, and it is not close. It has not publicly filed an IPO registration statement. Hayes sketches an initial offer of about 5 percent, a January listing, and a valuation approaching $1.5 trillion, with employees and early investors able to sell more later. None of that is an announced plan. A path that needs an unfiled listing, a chosen month, and a chosen valuation is a scenario. Do not lay it beside Anthropic’s 180- and 365-day windows and call them the same clock.
The credit link has to actually hold
Hayes draws a long route from those calendars to Bitcoin. Newly tradable shares weigh on prices. Lenders ask for more yield. The infrastructure bill gets harder to carry. Bank credit slows. Dollar liquidity thins. Crypto trades worse. Every link has to hold, and he has not shown that any of them has. He cites one weekly decline in U.S. commercial bank balance sheets, then notes himself that one reading is not a trend. The write-ups give neither the size nor the week. A print with the date torn off is not a credit cycle.
If the strain became a broader crisis, he expects central banks to ease, and he thinks the added money would later support Bitcoin. He names no date and does not establish that a rescue will happen. His firm, Maelstrom, plans to stay sidelined until liquidity looks clearer. Maelstrom’s plan is a positioning choice. It is not evidence the pipes are already dry.
One step in his chain is easy to over-read. He argues that a public listing would give these companies a more direct path into corporate debt markets to fund computing needs. That is his claim. Large private labs already sign huge compute commitments and already borrow. A listing changes who can hold the paper and how it is priced. It does not switch credit on.
He also folds three different pipes into one liquidity story: bank loans to AI businesses, financing of stock purchases, and banks’ own technology books. A data-center credit, a margin loan, and a securities portfolio do not fail in unison because a lockup expires. A weaker equity tape can raise the yield a lender wants. The argument is coherent. A single bank print does not prove it.
The number that makes any of this matter to an operator is the infrastructure claim, and it needs the same label. Both outlets attribute approximately $895.4 billion of AI infrastructure commitments to Anthropic’s filing, not to Hayes’s model. Without the document, there is no way to say whether that sum is contracted, contingent, or aspirational. If it is even in the neighborhood, the compute buildout has outgrown a venture round. Equity calendars, bond desks, and lender appetite are now part of the AI supply chain, whether or not Bitcoin notices.
The bill shows up in the product you buy
Most operating companies will never trade an AI-lab lockup. They will budget on the hope that intelligence keeps getting cheaper. That hope is a financing assumption wearing a product roadmap.
Extra supply pressures a price only if the bid does not rise to meet it. Plenty of lockups pass without a drama. The awkward ones are the ones where a large new float and a huge compute bill arrive in the same market, and the story cannot recruit buyers fast enough. When that happens, the cost of capital leaves the treasury footnote and shows up as a feature of the thing you are buying. Wider credit spreads never appear in a model card. They show up later, in three places a buyer can actually watch.
Published inference prices and committed-use rates are the first. If funding the buildout gets dearer, the list price and the discount for locking in volume are where it lands. A rate card that had been falling on schedule, and then stalls, is a cleaner signal than any essay about Bitcoin.
The second is the contract already on your desk. Vendors selling AI automation, AI agents, or an AI for sales rollout have been willing to discount implementation and promise a tight go-live while capital was easy. Watch the slippage: a quieter discount, a longer statement of work, a phase two that used to be phase one. That is someone else’s cost of money, renamed as your timeline.
The third is delivery of the metal and the building. GPU allocations and data-center capacity promised under cheap capital slip when the bill has to clear a wider spread. If a capacity date moves and the price concession moves with it, you are no longer looking at a narrative. You are looking at a supply chain. AI for business sits downstream of someone else’s capex.
Cheaper computing, if it arrives, accrues first to the operator putting that capacity into daily work. It is not a reason to treat a token story as a macro call.
What a buyer should ask instead
-
What would I need to see before treating lockup supply as real selling?
Eligibility is a permission, not a transaction. You would want volume after the restriction lifts, a price that actually gives way, and some sign that holders are distributing rather than sitting on the stock.
-
Is the 83 percent figure company-reported dilution?
No. It is Hayes’s label for eligible supply by June 2027, and $1.245 trillion is simply 83 percent of a hypothetical $1.5 trillion valuation if that entire slug were sold at an unchanged price.
-
What in a vendor contract depends on someone else’s cheap capital?
Committed-use pricing, the discount on an automation or sales deployment, and the date attached to GPU or data-center capacity. Those are the clauses that move if the buildout gets harder to finance.
-
Which single print is not enough to change a budget?
One weekly bank balance-sheet reading, one unlock headline, or one softer vendor discount. Hayes himself says a single bank print is not a trend, and the write-ups do not even give its size or date. Reprice a program when pricing, contract slippage, and delivery dates move together.
-
Does a listing create these labs’ access to debt?
Not from nothing. Hayes argues a public listing opens a more direct path into corporate debt markets. Private labs already commit to compute and already borrow. A listing changes the buyer base and the price of money.
Lockups are ordinary. The prophecy laid over them is not. Do not build an AI program that needs a central-bank rescue nobody has promised, or a forever-falling inference price that assumes yesterday’s yield still finances the buildout. The labs can keep chasing superintelligence and still hand markets a very ordinary problem: too much paper, not enough bid, and a compute bill that does not care about the narrative.