On June 1, 2026, Anthropic submitted a confidential S-1 and let slip a whispered timeline: an October listing at an implied valuation near $965 billion. Within the same 48-hour news cycle, selected investors reportedly discussed a $2 trillion mark for the company — a number, not a price. Neither figure deserves your attention. The $71 billion figure does: that is the reported size of Anthropic’s off-balance-sheet debt stack, arranged through marquee asset managers and secured against GPU infrastructure that does not yet exist. Call it what it is. It is a synthetic lease on the company’s own future, a senior claim parked ahead of every equity holder who buys the IPO. In that same window, OpenAI filed its own confidential S-1 with Goldman Sachs while pushing an expected listing into 2027; it simultaneously poured $105 billion into a 4.25-gigawatt data-center corridor in Ohio. And on the other side of the Pacific, Moonshot AI reportedly submitted an A1 filing in Hong Kong, targeting a $50 billion valuation built on $300 million in annualized revenue. One week. Three exits. Three capital philosophies. One audit trail.
The financial press keeps asking whether AI companies are overvalued, or whether the IPO wave is a bubble. That is the wrong question. Valuations are outputs. The inputs are capital structures designed to survive the long gap between model breakthroughs — and the even longer gap between a promise and a depreciation schedule. These three exits are not separate events; they are a single liquidity event wearing three different costumes. Each lab has decided how much risk to put on its own public balance sheet, how much to push into the shadows, and how much to hand to sovereign capital. I have watched this machinery before, only in markets that were smaller, faster, and more openly dangerous. During DeFi Summer, I audited lending contracts and learned that yield is often just risk wearing clean accounting. In 2022, I tracked stablecoin redemption rates against offshore NDF markets and learned that the first thing to break is not a coin — it is a covenant. Every one of those episodes had the same opening: a trail of capital commitments growing faster than the underlying assets could possibly deliver.
The lesson transferring into AI is uncomfortable, because the underlying assets are real. AI is not a meme coin. FLOPS are not vapor. But the financial structure that is being built on top of the hardware is now the same shape as the structures that broke in crypto. The audit trail does not begin at the S-1 filing. It begins upstream, at the point where compute was turned into a promise.
Context: Compute Is the New Reserve Asset
Let me lay out what actually happened over the past 90 days. OpenAI’s $852 billion post-money valuation contains a hard physical bet: a 20-year lease on Ohio land, 4.25 gigawatts of committed power, and billions in construction costs. That is not R&D spending. It is the creation of a power-and-compute utility disguised as a software company. Anthropic’s $965 billion story, meanwhile, is built around a cleaner ledger. Its $71 billion in reported infrastructure debt sits in vehicles outside the core operating entity, enabling the company to present a tidier income statement while the real liabilities accumulate behind a curtain. Moonshot AI and DeepSeek are running a completely different playbook: they attach themselves to Chinese industrial policy, sovereign-adjacent capital, and domestic listing routes where the valuation rules are set by scarcity rather than by S&P comps.
This should be understood as a divergence of not strategies but physics. OpenAI has chosen to absorb the physical cost of compute into its own equity; Anthropic has chosen to export that cost to credit markets; Chinese labs have chosen to import state-adjacent tolerance for long payback periods. Three different risk placements. The question that matters is not which model performs best on a benchmark. It is which risk placement survives the first real macro shock.
One more layer matters because it explains the timing. All three groups are racing to complete their capital formation before the next model cycle erases their current technical differentiation. The distance between frontier labs has collapsed from a 12-to-18-month advantage to a 6-to-9-month window — and sometimes shorter. DeepSeek-R1 demonstrated that a cost-efficient late follower can land within shouting distance of the frontier. If the public markets begin to price marginal model improvement, rather than absolute model capability, then the premium on any single lab shrinks dramatically. An IPO is therefore not an exit. It is an attempt to lock in capital at the peak of a narrative that has not yet been validated by the accounting.
Core Insight 1: OpenAI Bought a Utility, Not a Lab
Start with OpenAI, because its positioning is the most deceptive. Everyone sees the $852 billion label and assumes the market is pricing artificial general intelligence. But consider the Ohio project’s financial logic: on a $105 billion investment, with a 20-year lease and 4.25 gigawatts of committed power, the returns will not come from model subscriptions. They will come from the spread between the fixed cost of electricity and the variable price of inference. At that scale, you are no longer a technology platform; you are a regulated-equivalent utility with a very optimistic depreciation curve.
What this creates is a mismatch between asset longevity and model obsolescence. A power purchase agreement lasts 20 years. A frontier model may be marginal in 12 months. The $105 billion Ohio bet only makes sense if OpenAI believes that model-level differentiation will keep disappearing — and that owning the physical layer is the only remaining moat worth financing. That is a bold bet, but it is not a technology bet. It is a commodity bet. If inference costs eventually approach the marginal cost of electricity plus hardware amortization, then OpenAI’s vast infrastructure could become a low-margin energy business — stable, but not deserving of 20 percent software margins.
There is also an underappreciated signal hiding in the S-1 timeline. OpenAI reportedly filed confidentially but pushed its expected listing into 2027, later than Anthropic’s. This is not just pacing. It signals an intention to let the market absorb Anthropic as the reference point first. If Anthropic’s public listing succeeds, OpenAI can argue for a higher multiple by pointing to a richer revenue mix. If Anthropic stumbles, OpenAI gets to watch the market learn the risks before pricing its own shares. That is the behavior of a company that knows the asset base underneath its valuation is untested.
Core Insight 2: Anthropic’s Clean Ledger Is a Structured Product
Anthropic has chosen the exact opposite financial posture: keep the operating entity light, keep the debt off the reported balance sheet, and let the optics of a clean income statement support a narrative about safety, quality, and enterprise-grade reliability. The theatrical part is not the $1.5 billion copyright settlement Anthropic reportedly paid to clear its training-data risk; that was a necessary insurance premium before going public. The theatrical part is the structure around the debt, effectively a way of packaging future compute costs into vehicles that sit in a gray zone between sponsorship and ownership. If the model market slows and those GPUs are worth less than the debt secured against them, the lender is protected but the ordinary shareholder is not. The valuation may say $965 billion; the equity that the public will own is a residual claim on whatever is left after the senior claims of this hidden ledger are paid.
This is the point where the audit trail of a broken liquidity trap begins. Not with a collapse. Not with a bad earnings print. It begins with a tidy balance sheet that obscures the rough beast prowling just offstage. In crypto, I have seen this play out in miniature: teams borrow against volatile assets, keep the loans in separate legal vehicles, and present the core treasury as pristine. It works until the asset price drops by 40 percent — then the off-balance-sheet structure becomes the first thing that management tries to defend, and the equity becomes the last thing to get paid. Anthropic’s decision to abandon the reported acquisition of Decart makes sense in this context: buying an AI infrastructure startup would inject messy goodwill and integration risk into the public company’s first financial statements. That discipline is rational, until you realize that it prioritizes financial optics over technological synergy. The market will eventually demand clarity on how the $71 billion is secured, what happens at refinancing, and who holds the first claim on future revenue. That answer will not arrive in the prospectus — it will arrive during the first crisis, which is exactly the wrong time to learn the structure.
Core Insight 3: China’s Labs Trade Performance for Permission
The Chinese labs represent a third philosophy that is not really comparable. Moonshot AI’s $300 million in annualized revenue and a reported $50 billion target valuation implies a price-to-sales multiple north of 160x. That figure is mathematically sustainable only if revenue compounds faster than any enterprise software business has ever grown. The more relevant number is that a Chinese frontier lab can access domestic capital markets on terms that prioritize national AI strategy over financial efficiency — the listing is not just a liquidity event for shareholders; it is a signal that the company enjoys official permission to exist as a strategic asset. The advantage of sovereign alignment is patient, policy-protected capital. Its disadvantage is that your most important customer is the state, and your most important supplier is blocked by export controls.
DeepSeek’s reported investor base — Tencent, CATL, NetEase — is equally revealing. Tencent provides cloud distribution; CATL provides batteries and energy storage; NetEase provides content and consumer channels. None of them, however, can deliver the one input that matters most under U.S. export controls: a reliable, cutting-edge supply of advanced silicon. Chinese labs have solved commercialization. They have not yet solved the physics of restrictive trade policy. The market will eventually price the long-term training bottleneck that this creates — not as a one-time squeeze, but as a permanent structural tax on their iteration speed. The fact that their stock listings are nonetheless likely to be heavily oversubscribed tells you something about the global capital cycle. When sovereign priorities align with liquidity, valuation discipline is the first casualty.
Core Insight 4: Tokenized Compute Is the Canary in the Calculation
The most useful data point for investors tracking these three exits may not come from any S-1. It will come from decentralized compute markets — platforms that tokenize GPU rental or trade access to future compute capacity. I have spent the past two years modeling decentralized compute as an emerging liquidity layer, and the pattern is becoming clearer. When public AI companies raise billions for infrastructure whose cost assumptions are locked into 20-year depreciation schedules, the marginal clearing price of GPU capacity becomes a direct market referendum on those assumptions. If decentralized marketplaces show GPU rental prices sagging relative to the unit economics implied by the Ohio deal, while tokenized compute derivative assets quietly reprice downward, the public equity market has just been given a leading indicator — and it will almost certainly ignore it.
OpenAI’s infrastructure bet is essentially a very large, very illiquid token: it prices in the expectation that frontier models will keep requiring exponentially more compute. If that expectation breaks — if model efficiency improves faster than training scaling laws — then the marginal value of the Ohio corridor declines faster than any utility-style depreciation schedule can absorb. The large physical investment is in some sense one of AI’s more honest, transparent bets. Decentralized compute markets, by contrast, are less honest but more dynamic; they price GPU capacity continuously and have no narratives to protect. An attentive macro analyst will watch those decentralized pricing signals before trusting the fixed narrative of a public-market IPO roadshow.
The Contrarian Blink: The Decoupling Everyone Refuses to See
The market is now accepting a very convenient story: AI is decoupled from the rest of the liquidity cycle. Interest rates could be volatile, crypto could have its own bear cycles, but AI infrastructure is treated as an evergreen allocation that transcends capital-market conditions. That belief is contradicted by the very structure of the AI IPO wave. This is not a technology cycle. This is a debt cycle — the same kind of cycle that ravaged crypto valuations in 2022 when projects loaded up on cheap capital and committed to long-term infrastructure at fixed costs. The prevailing fiction is that AI compute is such a strategic asset that its financing cannot fail. The quiet reality is that every single one of these capital structures requires either model capability gains, revenue growth, or refinancing access to continue — and all three of those variables die simultaneously in a liquidity crisis.
The clearest contrarian signal is the accepted valuation of intangible optionality. The reported $2 trillion valuation discussions for Anthropic, a company that has yet to prove its revenue is genuinely recurring, should be treated as an indication that the market is pricing the option value of safe AGI — an asset that may take twenty years, or may never arrive. Public markets have historically struggled to price options that have no expiration anchor. When a narrative depends on distant optionality, even a small delay in horizon causes a disproportionate contraction.
Moreover, the AI Act quietly functions as an instrument of consolidation. For small AI startups, compliance costs are a fixed cost that consumes real runway. For frontier labs, they are a toll that separates oligopolists from challengers. This is the same dynamic I saw when the EU’s MiCA framework emerged for stablecoins: clear regulation favored incumbents who could afford continuous compliance, while it became a de facto wealth transfer from entrepreneurs to capital-heavy operators. The AI IPO wave will accelerate that pattern. The regulated will be rewarded; the fragile will be absorbed; the gap between global AI giants and regional players will widen — all while the crowd celebrates the supposed clarity of rules that hardened into moats.
The Takeaway That Matters
So what do you actually do with this information, in a world where capital flows are moving faster than attention spans? The lesson is not to avoid every AI IPO. The lesson is to audit the structure before touching the story. Ask where the liabilities sit. Look for off-balance-sheet vehicles; they will not be described clearly in the risk factors but they will shape the actual recovery order in a downturn. Ask whether the revenue is customer-driven or subsidy-driven, then calculate how many of those customers are strategic investors or sovereign-adjacent buyers. In a bear market — and make no mistake, this current market is a bear market for leverage, even if equity indices pretend otherwise — survival is not a function of hype. It is a function of balance-sheet cleanliness, and cleanliness is precisely what all three of these structures are designed to avoid. Watch the audit trail of a broken liquidity trap; it will not begin with a hack. It will begin with a footnote in a filing that no one read before the price was set. Watch the tokenized compute data. Watch the quality of the book that actually converts at the IPO. And remember that every one of these firms is a startup selling stock before the interest expense catches up to the story — the only question is which one sells first.
When the public markets finally absorb the true gap between AI’s $2 trillion dreams and its dependency on $0.03-per-kilowatt-hour electricity, the correction will not be a critique of the technology. It will be recognition that the capital structure and the physical reality were on different clocks all along. The dead giveaway — even before the first post-IPO lockup expiry — will be subtle. When insiders at every one of these firms finally have a paper price, watch whether they hold or sell. That will be the clearest signal of what they think about their own balance sheet’s depth. The audit trail doesn’t need to scream to be read. It just needs one more sentence than the headlines.
