The AI IPO Wave Is Not About AI: Three Labs, Three Leverage Trades, One Vanishing Moat

On June 1, 2026, Anthropic's lawyers filed an S-1 under confidential cover. It was not a model release, and it did not arrive with benchmark charts. It arrived with a rumored October IPO window, a post-money valuation near $965 billion, and reports that some investors were already floating a $2 trillion scenario. The same week, OpenAI was finalizing a twenty-year power commitment for 4.25 gigawatts in Ohio, part of a $105 billion infrastructure program that will make it the anchor tenant of the American AI grid. Across the Pacific, Moonshot AI was preparing a Hong Kong A1 filing for Kimi K3, a product with $300 million in annual recurring revenue, at a reported $50 billion target valuation. Three labs. Three exit strategies. Almost two trillion dollars of collective ambition pretending to be one technology story. It is not one story. I have watched this genre of narrative before. In late 2017, my Bangkok education group was flooded with ICO whitepapers that all promised the same decentralized future. I audited fifteen projects by opening their repositories first, and I found red flags in eight. The marketing was always prettier than the code. Code doesn't lie, but narratives do — and the dominant narrative right now is that the AI IPO wave is about intelligence. It is about capital structure. The benchmark scores are just the header image on the pitch deck. Here is the framework that makes sense of the filings. OpenAI is running proof-of-work. It is spending hundreds of billions to own land, power, and compute; its confidential S-1 is being handled by Goldman Sachs and Morgan Stanley, and its listing timeline reportedly slips to 2027. Anthropic is running financial engineering: a leaner listing aimed at October 2026, supported by $71 billion in off-balance-sheet compute financing arranged through Apollo and Blackstone, a dropped acquisition of AI infrastructure startup Decart, and a quietly settled $1.5 billion copyright dispute. Moonshot AI and DeepSeek are the sovereign-adjacent plays, but they are not the same trade. Moonshot wants global institutional price discovery for a product that already generates revenue; DeepSeek wants patient industrial capital from Tencent, CATL, and NetEase on the way to a domestic listing. None of this is happening because the technology is fragile. It is happening because the technology moat is shrinking in real time. The frontier lead window has compressed from twelve to eighteen months down to six to nine months. DeepSeek-R1 reached near-frontier capability at a fraction of the usual training cost. When code becomes copyable, the market reprices everything within weeks. I saw this in 2020, when I audited SushiSwap's fork mechanics and ran workshops in Bangkok for two hundred developers learning to lend and trade on Uniswap and Aave; within months, hundreds of automated market makers were running the same open logic. Uniswap survived because its liquidity network was hard to fork, not because its code was secret. Every AI lab now faces the same question: what is the liquidity network that keeps your model from becoming a commodity? OpenAI's answer is that the durable network is physical. A 4.25-gigawatt commitment for twenty years is a scale of energy contract that makes hyperscalers look cautious. If models are on track to become a commodity, then the lasting barrier is capital intensity: land, power, substations, cooling, and the cost of running the largest inference network at the lowest marginal price. This is proof-of-work logic applied to artificial intelligence — anchor the software advantage to a physical cost of entry. The logic is internally consistent, but it quietly changes the valuation basis. At $852 billion, the market is pricing OpenAI as a software platform. A platform earns growth multiples on recurring revenue. A utility earns regulated multiples on assets and cash flow. Depreciation schedules, take-or-pay obligations, and the internal rate of return assumptions in that Ohio project will decide which story survives contact with the S-1. Floating five to ten percent of the company would raise $40 billion to $80 billion — several years of global AI venture funding in a single trade. That is why the listing matters more than any model release. Anthropic's route is more elegant and more fragile. A $965 billion valuation implies static price-to-sales multiples between 48 and 64 times, assuming annualized revenue of $15 billion to $20 billion. Mature software franchises trade near twelve to eighteen times sales. To justify the gap, Anthropic needs three to four consecutive years of high growth, a durable safety premium, and no regulatory shocks. Factor in the rumors of a $2 trillion discussion among early investors, and the valuation is no longer a number; it is a demand that the market accept a new asset class called AGI optionality. That is not an investment thesis; it is a weather forecast. The deeper problem is the $71 billion that will not appear on the balance sheet. In decentralized finance, a senior claim of that size sitting ahead of equity would be visible to anyone with a block explorer — the collateral, the liquidation thresholds, the entire waterfall. In traditional finance, it is a fundraising footnote. Anthropic's off-balance-sheet financing through Apollo and Blackstone is clever, and it will look even cleverer on a roadshow slide. But the lenders stand before the shareholders in the capital stack. If model economics deteriorate, the lenders will not care about Claude's benchmark score; they will call the collateral. Public shareholders will be buying the residual claim after the senior creditors are made whole. I developed my instincts about hidden leverage during the 2022 bear market, when I pivoted from retail crypto education to compliance training in Thailand and watched leveraged projects die while cleaner ones survived. Alpha hidden in the noise: whenever a capital-intensive company markets a clean balance sheet, look for the debt they moved elsewhere. The copyright settlement shows where Anthropic believes the real line of competition is. Fifteen billion dollars to settle training-data litigation is not a legal cost; it is a purchase of provenance. Clean training data and auditable behavior are not compliance overhead in an era when the EU AI Act is a baseline entry tax on every serious model provider — they are a pricing premium. Anthropic is selling trust as a feature, with an eight-to-ten-point reported lead over OpenAI on SWE-bench as the proof point. OpenAI, meanwhile, is still litigating its data questions in multiple jurisdictions. Trust is the new currency, and Anthropic is spending aggressively to be the lab with the cleanest reserve. The Chinese labs complicate the neat three-way story. Moonshot's $300 million in revenue against a $50 billion target implies a multiple above 166 times sales. That is extreme by any standard, even for Hong Kong. But it is at least tied to an invoice; the company has moved from research narrative into an actual revenue story. That is more than can be said for valuations that embed an unexamined option on artificial general intelligence. DeepSeek is a different vehicle: backed by Tencent's cloud and CATL's energy balance sheet, constrained by export controls on advanced GPUs, and aimed at Shanghai's STAR Market. Lumping these two into a single Asian basket is the kind of analytical shortcut that produces bad trades. Moonshot is courting global funds in Hong Kong. DeepSeek is nested inside domestic industrial policy. Their capital sources will drive their model strategies apart faster than any benchmark comparison. The uncomfortable part that few commentaries will name is that this entire wave is a hedge. The labs are trying to refinance technical uncertainty with patient public-market capital before the benchmark race settles. If the pioneer prices too high and breaks — and the market has a long history of making an example of the pioneer — the repricing will not stay contained. It will sweep across the sector with twenty-to-fifty-percent discounts, and latecomers in the queue will suffer the most. Snowflake spent four years growing into the multiple it wore in 2020. That is also why OpenAI lets Anthropic go first. Setting the comp sheet for a new asset class is not an advantage; it is a risk. The side that goes second gets to watch the reception before committing to a price. The contrarian position is uncomfortable for both AI bears and maximalists. The most mocked valuation here — Moonshot at 166 times sales — may be the most honest, because it is tied to product revenue. Meanwhile, the lab presenting the cleanest balance sheet may carry the most hidden leverage, while the lab with the grandest infrastructure vision may be repriced as a utility when the depreciation schedule is published. The S-1s will also expose an option overhang that nobody is pricing yet; when employee equity vests after listing, real compensation costs hit the income statement. Software analysts call it compensation shock. Crypto natives call it a token unlock. Same chart, different asset class. In other words, stop betting on the stories and start reading the filings. The S-1 is the source code. The roadshow is the whitepaper. I learned that ordering in 2017, when opening repositories before reading pitch decks saved me from eight of the fifteen projects. That habit has not aged a day. For those inside the crypto-AI convergence, the wave redefines the opportunity set. I spent 2025 building an Autonomous Ethics Lab in Bangkok as AI agents began transacting on-chain. The insight from that work is simple: autonomous systems need transparent and auditable rails, and now the largest AI companies are going to spend the next decade discovering that public markets demand exactly that. Model weights are not going to be the most valuable infrastructure in AI. Verifiable training data, auditable capital structures, and regulator-accepted governance tooling will be. The EU AI Act is already creating a compliance industry the way MiCA created one for crypto: model audits, risk assessments, explainability reporting, provenance tracking. Open networks and blockchain tooling can occupy that ground. Trying to out-compute OpenAI with a token is not a strategy; it is a memorial. The next twelve months will answer whether a market built on quarterly honesty can price a technology that keeps promising to change everything in the next breakthrough. Watch the quality of the institutional order book in Anthropic's October listing. Watch the depreciation assumptions and power contracts inside OpenAI's S-1. Watch how much of Moonshot's revenue comes from outside its own ecosystem. Do not watch the benchmark leaderboard; by the time a model score is public, the capital has already moved. The same discipline applies to this article: treat every leaked valuation and off-balance-sheet number here as a rumor until the filings land. The reports of $2 trillion discussions, the $71 billion financing, the terms of the Decart withdrawal, and the ownership details behind DeepSeek all need confirmation from primary documents. But the direction of travel is already visible: the AI wave is securitizing before it is decentralized. The first lab that lets the public audit its infrastructure debt, its data provenance, and its governance will be the one that deserves a premium. Until then, this is a market buying promises. Alpha hidden in the noise goes to the people who read the actual source code. That is what the S-1 is.