Arthur Hayes's AI Dilution Thesis, Tested: Unlocks, Credit and Bitcoin

Arthur Hayes argues that a wave of insider share unlocks at newly public AI giants will cap their stock prices, tighten credit for the AI build-out, slow bank balance-sheet growth and drain the dollar liquidity that bitcoin needs, until a crisis forces central banks to print again. His November 5, 2026 essay, “D is for Dilution”, calls this “the transmission channel from AI to crypto.” This piece tests every link in that chain, plus the extensions in a widely shared summary that adds power limits, open-source models, recursive self-improvement and the election cycle, against SEC filings, Federal Reserve data, academic research and the labs’ own disclosures. The short version: the first link is real but smaller than claimed, the middle of the chain is contradicted by current data, and the variable that actually governs liquidity right now is the Fed.

TLDR

Large staged unlocks are coming at SpaceX (now merged with xAI) and, once it prices, Anthropic, and the research on lockup expirations supports a modest drag on prices. But credit markets are wide open to the biggest AI spenders, US bank credit is growing about 6% a year, and M2 has risen every month of 2026, so the step from “flat AI stocks” to “less liquidity” has no support in current data. The real pressure points are narrower: debt-heavy neoclouds like CoreWeave, a Fed that raised rates in September and may raise again in December, and power grids that take years to expand. Hayes’s own essay concedes the bank evidence is “only one print,” and he sold about $22 million of crypto in October saying “we gotta go down to go up.”

Thoughts

The chain breaks where it matters most for bitcoin holders: the banks. Hayes’s claim is that weak AI equities slow bank balance-sheet growth, which slows the creation of dollars. His evidence, in his own words, is that “the banking system’s balance sheet fell WoW. I know this is only one print.” The Fed’s H.8 release shows commercial bank credit at $19,883.2 billion in September 2026 against $18,744.3 billion a year earlier, a gain of about 6%, and loans to nonbank financial firms, the channel through which private credit funds AI projects, up about 20%. The money supply tells the same story: M2 was $23,413.9 billion in September versus $22,149.2 billion a year before, up roughly 5.7% and higher in every month of 2026. The mechanism Hayes describes is real, as the Bank of England’s classic “Money creation in the modern economy” explains: bank loans create deposits. But the same paper says the amount of money created ultimately depends on monetary policy, chiefly interest rates. And policy is tightening. The Fed raised rates on September 16 to 3.75% to 4%, and on October 28 two officials voted for another hike. If liquidity-sensitive assets struggle into 2027, the culprit is far more likely to be the federal funds rate than the Anthropic lockup calendar.

The dilution itself is real, but Hayes’s numbers are his own estimates, and the research suggests the effect is a drag, not a wall. SpaceX’s IPO prospectus, filed June 12, sold 555,555,555 Class A shares at $135 and lays out a drip of releases from the 180-day lockup through December 8, 2026; shares held by Elon Musk “will not be subject to any early release provisions” and stay locked until June 2027, according to the 424B4. Anthropic’s amended S-1, which proposes listing on Nasdaq as ANT, leaves the price range and share counts blank and itself warns that post-lockup sales “could adversely affect the prevailing market price.” Hayes’s headline figure, that IPO buyers “face 83% dilution by June 2027,” comes from what he says his AI agent read in the filing, and his $1.245 trillion of buying needed to hold the price flat is arithmetic on an assumed $1.5 trillion valuation, as crypto.news notes. He even concedes it is “Not impossible, but improbable.” The academic record is more modest. Field and Hanka’s study of 1,948 lockups found a three-day abnormal return of about minus 1.5% at expiry, larger for venture-backed firms, and Ofek and Richardson found 1% to 3% drops that are well anticipated. Later work ties the drop to IPO overvaluation, not supply alone. No study covers a multi-year staged unlock at a trillion-dollar company, so nobody knows the effect, which is a reason for humility in both directions.

Credit is open to the giants and expensive for the levered builders, and that split is the most useful thing the thesis points toward. Eleven days after its IPO, SpaceX priced $25 billion of senior notes with coupons from 5.35% to 6.65%. Alphabet raised $85.2 billion of debt in the first nine months of 2026 and Amazon $115.1 billion, according to their quarterly filings, and Microsoft paid for $115.9 billion of capex out of operating cash flow. Aggregate spreads are near historic lows: the ICE BofA investment-grade spread was 0.91% and high-yield 3.32% on November 4, per FRED. The Fed’s May Financial Stability Report found corporate spreads at the low end of their range but speculative-grade technology spreads widening “more notably.” That is where the stress lives. CoreWeave’s net interest expense was $1.18 billion in the first half of 2026 against $4.65 billion of revenue, and on Thursday, as the Nasdaq fell into correction, neoclouds were among the hardest-hit stocks. The academic case that equity drives credit is also weaker than the essay implies: Collin-Dufresne, Goldstein and Martin found that “variables that should in theory determine credit spread changes have rather limited explanatory power.” The equity-to-credit link works best for weak borrowers, which is exactly the neocloud story.

The physical layer is the strongest part of the extended essay, and it points in a direction Hayes does not take. Power really is a multi-year ceiling: Lawrence Berkeley National Laboratory’s Queued Up report found the median project built in 2025 took 61 months from interconnection request to operation, against 36 months in 2015. PJM’s 2027/2028 capacity auction cleared at its cap and 6,516.6 MW short of its reliability target. And the frontier is getting more expensive, not cheaper: OpenAI lists GPT-5.5 at $5 and $30 per million tokens, double its predecessor; Google’s Gemini 4 Argon doubles to $4 and $20 on December 1; Anthropic’s Claude Fable 5.1 costs $10 and $50, two and a half times Opus 5.5. Below the frontier, prices fall fast, and open-weight models now carry most tokens on OpenRouter while closed models still take 73% of spend. That is a two-tier market: scarce, rising-price frontier intelligence and abundant, falling-price commodity intelligence. Hayes himself predicts “compute will flood the spot markets as more data centers begin operations.” Both can be true at once, and the frontier tier is the one the IPO valuations are pricing.

The last two layers, recursive self-improvement and the election cycle, are where the extended essay outruns its evidence. Efficiency is real and fast: Epoch AI estimates the compute needed for a fixed level of language-model performance halves about every eight months. But total demand keeps rising anyway, because training compute is still scaling, as Epoch puts it, “at around 4-5x/year, despite recent advances in compute efficiency.” Anthropic’s own research note on recursive self-improvement says “We are not there yet, and recursive self-improvement is not inevitable,” and lists energy and compute as possible binding constraints. On politics, Hayes never ties the bailout to elections. The research on independent central banks finds no political monetary cycles, and the Fed’s big rescues of 2008, 2020 and 2023 followed funding-market breakdowns, not falling stock prices; in 2022 the Fed tightened into a falling market. The essay’s closing claim, that liquidity expansion is the high-probability end state, has history behind it. But the trigger to watch is a seizure in funding markets, not a lockup expiry, and today the Fed’s balance sheet shows primary credit borrowing of just $9.5 billion. The honest bet is not “dilution, then bailout.” It is “watch the Fed, the levered builders and the grid.”

Key Takeaways

Hayes’s core claim is that AI IPO dilution will keep lab stock prices flat, tighten AI credit, slow bank balance sheets and squeeze dollar liquidity until central banks bail markets out.
SpaceX sold 555,555,555 shares at $135 in June 2026, with staged lockup releases through December 8 and Musk’s shares locked until June 2027 with no early release.
Anthropic’s S-1/A proposes listing on Nasdaq as ANT, has not priced, includes selling stockholders, staggers releases across many dated tranches, and warns post-lockup sales could hurt its price. OpenAI has no public IPO filing.
Hayes’s 83% dilution and $1.245 trillion figures are his own estimates, not disclosures, and he calls a flat price “Not impossible, but improbable.”
Academic studies find lockup expirations cause roughly 1% to 3% price drops and a lasting 40% volume increase, concentrated in overvalued venture-backed firms.
Credit stayed open: SpaceX sold $25 billion of notes after its IPO, Alphabet and Amazon raised record debt, and investment-grade spreads were 0.91% on November 4.
The stress is in levered builders: speculative-grade tech spreads widened, and CoreWeave’s half-year interest cost was about a quarter of its revenue.
US bank credit grew about 6% in the year to September 2026 and M2 rose every month of 2026, contradicting the bank-balance-sheet link.
The Fed raised rates in September and held in October with two votes for a hike, so tighter liquidity is coming from monetary policy, not dilution.
Power is a genuine multi-year constraint: 61 months median from interconnection request to operation, and PJM short of its reliability target.
Frontier token prices are rising while commodity AI prices fall, and open-weight models carry most OpenRouter tokens but a minority of spend.
Efficiency gains have not cut total compute demand, and Anthropic says recursive self-improvement has not arrived.
Researchers find no political monetary cycles at independent central banks, and past Fed rescues followed funding breakdowns, not stock declines.

Detailed Summary

Link 1: Unlocks to Flat Prices (Moderate Support)

Hayes argues “you don’t trade the market cap; you trade the price per token,” comparing AI stocks to high-valuation, low-float crypto tokens. He estimates about 8.6 billion more SpaceX shares unlock by the end of June 2027, mostly Musk’s, and says “it is almost impossible mathematically for the stock price to go up between listing and June 2027.” The filings confirm large staged releases, and Anthropic’s own risk factors flag the danger. The research supports a measurable but modest effect at each release, so the sign is right and the size is uncertain. Demand can grow too: Krishnamurti and colleagues find returns after unlocks track liquidity changes and can be positive.

Link 2: Equity to Credit (Weak Support)

Research by Norden and Weber finds stock returns lead credit default swap and bond spread changes, more strongly for weaker credits. But the Bank for International Settlements reported in March 2026 that tech equities underperformed on valuation worries while credit spreads stayed compressed, even as hyperscaler bond issuance topped $100 billion in 2025. Equity weakness has not yet translated into broad credit tightening.

Link 3: Credit to a Slower Build-Out (Weak Support)

Capex accelerated in every major 2026 filing. Alphabet’s nine-month capex was $132 billion against $63.6 billion a year earlier, Amazon’s $175.7 billion, Meta’s purchases of property and equipment $83.1 billion. Oracle filled a funding gap with $19.9 billion of at-the-market stock sales rather than new debt, which is dilution of a different kind. SpaceX signed a commitment for about $40 billion of GPU lease financing with an Apollo-managed entity in October. A reinforcing loop is plausible for neoclouds, but no data yet show financing slowing the build.

Link 4: Banks and Liquidity (Contradicted)

Bank credit, loans to nonbank financials and M2 are all growing. Bank capital is historically high, per the Fed’s stability report, and FINRA margin debt was about $1.48 trillion in September, near its highs. The Fed ended balance-sheet runoff on December 1, 2025, and has been buying short-term Treasuries to keep reserves ample. Hayes’s view that the Fed is “on strike” refers to rate policy and to officials wanting banks rather than the Fed to create credit, not to a shrinking money supply.

Link 5: Power as an Independent Ceiling (Moderate Support)

NERC’s long-term assessment finds 13 of 23 regions face resource adequacy challenges as data center demand surges, and GE Vernova’s gas turbine backlog and reservations reached 122 gigawatts. Against that, Epoch estimates a gigawatt-scale data center can be built in about two years, and Anthropic eased its own capacity strain within weeks by signing for SpaceX’s Colossus 1 and expanding with Amazon. No source links power scarcity directly to token prices.

Link 6: Sticky Token Prices and Open Models (Weak to Moderate)

New flagship models launched at higher prices across OpenAI, Google and Anthropic, while prices within tiers fell. Epoch finds the best open-weight models lag the closed frontier by about four months on average since January 2026, while Artificial Analysis’s newer index shows a wider gap. Open models plausibly cap prices for most tasks, but there is no evidence they are diverting demand from the big labs’ capital raises.

Link 7: Efficiency and Recursive Improvement (Contradicted as Stated)

The extended essay itself admits that cheaper capability can unlock more demand, which is the fair reading. Epoch measures inference prices for a fixed capability falling by a median of about 50x a year, and frontier training power growing about 2.2x a year. Cheaper intelligence has so far meant more total spending on it, not less, so efficiency does not reduce the capital the build-out needs.

Link 8: Elections, Bailouts and Bitcoin (Weak Support)

In March 2020 the Fed replaced a fixed purchase pledge with open-ended buying within days, the “insufficient first response, then bigger” pattern the essay describes. But that followed a breakdown in Treasury and funding markets. On bitcoin, BIT’s analysis finds a lagged correlation with global M2 of about 0.5, a same-month correlation of about 0.04, and a decoupling since mid-2025. Hayes’s own record is mixed, and on October 24 he sold about $22 million of ETH and altcoins, citing macro and geopolitical risk, as reported by CryptoSlate.

Notable Quotes

“This is the transmission channel from AI to crypto.”
Arthur Hayes, D is for Dilution

“I know this is only one print, but set against the backdrop of negative AI credit market vibes, you can see why I’m worried.”
Arthur Hayes, on his evidence that bank balance sheets are shrinking

“While we wait, compute will flood the spot markets as more data centers begin operations.”
Arthur Hayes, in a line that cuts against the sticky-token-price thesis

“Variables that should in theory determine credit spread changes have rather limited explanatory power.”
Collin-Dufresne, Goldstein and Martin, Journal of Finance

“Training compute scaling will likely continue at around 4-5x/year, despite recent advances in compute efficiency and the shift towards reasoning models.”
Epoch AI, on why efficiency has not cut demand

“We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for.”
Anthropic, on recursive self-improvement

“I think we gotta go down to go up.”
Arthur Hayes, after closing Maelstrom’s directional crypto book in October

Hayes says the timing of the bailout will be “answered in subsequent essays.” Read his full essay here, and for the political backdrop to Thursday’s tech selloff, see our report on the midterm data center backlash.

Arthur Hayes’s AI Dilution Thesis, Tested: Unlocks, Credit and Bitcoin