Why the Biggest Liquidity Event in History Might Not Be What Anyone Expects
The AI IPO Paradox: Why the Biggest Liquidity Event in History Might Not Be What Anyone Expects
By Shruti Shah, General Partner, Symphonic Capital
Every investor I've spoken with over the last few months has said some version of the same thing: "We're waiting for the IPOs."
They mean SpaceX. OpenAI. Anthropic. Three companies heading public this year with a combined paper valuation approaching $3 trillion. The logic makes sense on its face — when these giants finally go public, capital gets unlocked, investors get their money back, and the whole machine starts moving again. The private market thaws. Everyone exhales.
I think this story is more complicated than it looks. There are four things I keep coming back to. When you hold them all at once, the picture changes pretty dramatically.
Part One: The liquidity event is real — but much smaller than it looks
When a company goes public, it doesn't sell all its shares to the world. It releases a slice — typically 15 to 25 percent — called the "float." That's what's actually available to buy on day one. The rest stays locked up with founders, early employees, and existing investors, usually for six months or more.
Facebook floated 15 percent when it went public. Google floated 19 percent.
Now do the math on these three companies. At a 15 percent float, SpaceX, OpenAI, and Anthropic would need public markets to absorb somewhere between $430 and $580 billion in a single quarter. In 2025 - which was a rebound year after multiple years of a soft IPO market - tThe entire U.S. IPO market raised about $45 billion in all of 2025. The combined demand from just these three companies exceeds the entire 2025 U.S. IPO market by roughly two to four times.
They can't float 15 percent. The market can't hold it. So they'll likely float somewhere between 3 and 8 percent. Which means the actual cash that flows back to investors — the liquidity that so many have been waiting years to receive — will be a fraction of those $3 trillion headlines.
There is a genuine bull case here worth taking seriously. The 2025 IPO market was unusually quiet — not because demand disappeared, but because supply dried up. Three years of pent-up institutional demand, sovereign wealth funds have had less access to private AI entirely, and a new generation of retail investors accessing markets through thematic ETFs creates a demand pool that is structurally larger than any prior tech IPO cycle. If SpaceX opens well in June and signals that public markets are ready, OpenAI may be able to float more aggressively than the pessimistic case suggests — perhaps 10 to 12 percent rather than 3 to 5. The pent-up demand is real. The appetite is real.
But here is what the float debate misses entirely. Whether the IPO is absorbed well or badly is almost beside the point. The moment these companies file their S-1s, three years of private narrative meets public disclosure for the first time. OpenAI's $14 billion projected loss in 2026. Anthropic's compute bill is larger than its own annual revenue. The circular relationships between investors and customers that have never had to be explained in a prospectus before. Public market investors will read these documents. Analysts will model them. Short sellers will stress-test them. The story that has been justifying valuations across the entire AI sector — public and private — gets pressure-tested against actual numbers.
The liquidity event may be larger than the pessimistic case. The reckoning will happen regardless.
There is also a quieter version of this already underway. Over the past two years, the secondary market has been releasing pressure well ahead of any IPO — OpenAI, SpaceX, and Anthropic have all run large tender offers, with secondary transaction volumes across private markets hitting record highs in 2025. The insiders with the most urgent liquidity needs — early employees, seed investors — have in many cases already sold, at peak private valuations, before the public market gets a chance to weigh in. What remains on the cap table by IPO day tends to be strategic investors who can't sell, late arrivals who need the price to hold, and retail buyers promised a slice of the float.
The headline number and the actual cash that changes hands are very different things. And much of the smartest money has already quietly left the building.
Part Two: The circular money machine nobody is talking about
Here's the part that doesn't get nearly enough attention.
The AI giants heading toward IPO aren't just raising money from venture investors. They've been spending it — in staggering quantities — directly into the revenue lines of public companies. And those public companies' stocks have been rising accordingly. What looks like organic growth in Microsoft's cloud, Google's backlog, and NVIDIA's chip sales is substantially something else: venture capital raised at inflated private valuations, converted into operating expenses, and cycled through to the earnings of publicly traded companies.
The numbers are not subtle. OpenAI spent over $12 billion with Microsoft's Azure cloud in roughly the past two years. It is projected to spend $50 billion on compute in 2026 alone — a figure that rivals the full-year capital expenditures of Google, Microsoft, and Meta individually. Anthropic just committed $200 billion to Google Cloud over five years — a single deal representing more than 40 percent of Google's entire disclosed revenue backlog. The compute bill in year one is larger than Anthropic's own annual revenue.
And then it gets circular. Google has invested over $40 billion into Anthropic — the same company it just signed a $200 billion compute contract with. NVIDIA invested $30 billion into OpenAI — the same company buying NVIDIA chips. Wall Street analysts have described this structure explicitly as fitting "squarely into the circular investment theme" — the same pattern that inflated vendor financing during the dot-com bubble. One description that has stuck: everyone is financing everyone else with money none of them have actually earned yet.
The stocks that have been most inflated by this machine fall into a clear pattern:
Oracle is the starkest and most instructive case. OpenAI represents 58 percent of Oracle's contracted backlog — a $300 billion, five-year cloud deal announced in late 2025 that briefly made Oracle's stock soar and Larry Ellison the world's richest man. The stock has since fallen 50 percent from its September high. Investors have started pricing exactly what this piece is describing: if the circular financing structure unwinds, Oracle will have spent tens of billions building data centres for a customer that cannot afford them. Oracle is not a prediction. It is the first domino. It has already fallen.
Microsoft has received over $12 billion from OpenAI in compute fees in the past two years, with OpenAI projected to spend $50 billion in 2026 — most of it flowing through Azure.
Google/Alphabet has a revenue backlog that is more than 40 percent dependent on a single Anthropic commitment from a company that isn't yet profitable.
CoreWeave went public earlier this year with $22.4 billion in OpenAI contracts representing roughly a third of its entire order book. It is the purest expression of the machine: a company whose entire existence is predicated on the continued, accelerating spending of a money-losing private AI lab.
SoftBank owns an estimated 11 to 13 percent of OpenAI — more direct exposure than virtually any other publicly traded company. When OpenAI missed its internal revenue targets last month, SoftBank absorbed the steepest single-day stock decline of any public company in the AI complex.
NVIDIA sits in a different category — genuinely dominant with 86 percent of the AI chip market, real products, real revenues, and a moat that predates the current AI cycle. But it too is backing the very companies buying its chips, and its stock trades at a price-to-earnings ratio of 43 after a 95 percent one-year run. It is not immune to a repricing of the narrative supporting those multiples.
One important exception worth naming: Meta. Zuckerberg has been explicit that Meta's AI infrastructure spending — $125 billion in capital expenditure this year — is funded entirely from the company's own cash flow. Meta earns real profits from advertising. Its AI bet is a genuine enterprise investment. Meta's stock may face its own pressures if AI's productivity returns disappoint, but it is not part of this particular machine.
Part Three: What happens when the machine meets public disclosure
When these companies go public, the circular flow faces its first real stress test.
Public market investors will see the actual financial flows for the first time. Boards under earnings pressure will scrutinize the $50 billion compute bill. The question "do we really need all of this Azure?" becomes a shareholder question, not just an engineering one. OpenAI earned $13 billion in revenue in 2025 and spent $22 billion to do it. It is projecting $14 billion in losses in 2026 and does not expect to reach profitability until 2030. Its valuation going into IPO is roughly 65 times its annual revenue.
Public market investors are not going to shrug at that. They are going to ask hard questions. And when the most hyped company in the world starts getting marked against reality rather than narrative, it creates pressure across the entire sector — public and private alike.
The circular flow — venture capital into private AI companies, private AI spending into public cloud and chip revenue, rising public stock prices justifying higher private valuations, higher private valuations attracting more venture capital — slows. And the stocks that were built on that flow follow.
Oracle has already shown us the shape of this. A single quarter of missed OpenAI targets sent its stock down sharply and triggered a wave of analyst reassessments. Multiply that across Microsoft, Google, CoreWeave, and SoftBank, and you have the anatomy of a repricing cascade: IPO disclosures compress AI narrative multiples, proxy trades unwind, index rebalancing adds sell pressure, and private market comps follow public comps down. Each step rational on its own. Consequential in sequence.
Part Four: Three layers — and which one actually compounds
Before going further, one clarification worth making explicitly: none of this is an argument that OpenAI, Anthropic, or SpaceX disappear. They won't. The frontier model providers are the railroad infrastructure of this era — they will be used, they will matter, and they will very likely be dominant for years. But railroads, once built, become regulated utilities. They are essential and they are low-margin. The fortunes in the railroad era were not made by owning the railroad. They were made by owning the land the railroad ran through — the towns that grew up around the depots, the industries that only became possible because the infrastructure existed.
The picks-and-shovels layer — cloud compute, chips, infrastructure tooling — will always have a place. NVIDIA sells shovels to every gold miner regardless of who finds gold. That is a real and durable business. But the correction argument isn't that the shovels go away. It's that shovel valuations got priced as if every miner was going to find gold, and that assumption is now meeting reality.
What the correction actually surfaces is a third category that has been largely invisible during the hype cycle: the companies that own the land.
Think of it as three layers:
Layer one — the foundation. OpenAI, Anthropic, the frontier model providers. Essential infrastructure. Will survive and matter. But essential infrastructure compresses toward utility margins over time. This is not where the highest returns accrue.
Layer two — the shovels. Cloud compute, chips, tooling. Real businesses with real revenue. But priced, in many cases, as if the gold rush never slows. The correction reprices the shovels against reality, not against zero.
Layer three — the land. Proprietary data that the frontier models need but cannot generate themselves. Embedded workflows in markets that were ignored. Trust relationships with communities that no amount of compute can replicate. These are not infrastructure plays in the commodity sense. They are the assets that become more valuable as the infrastructure layer commoditizes — because when every company can access the same models, the only remaining question is who has the data to make those models actually work.
When the air clears, the question that was always the right question becomes audible again: what actually has a moat?
The answer is not the company with the best model. Models are increasingly commoditized. The frontier models are converging in capability and the cost to access them is dropping fast. You cannot build a durable company by being slightly better at prompting a model someone else built and maintains.
Data is not a commodity.
Take Croux, a Symphonic portfolio company. Croux connects hospitality businesses — hotels, stadiums, country clubs, catering operations across the Heartland — with on-demand workers. Ninety percent shift fill rates. A sub-two percent no-show rate. Thirty-one thousand active workers, each carrying 63 quality signals built from real work history: trust scores, certifications, proximity, and reliability across hundreds of actual shifts.
The data moat isn't the matching algorithm. Any well-funded competitor can build a matching algorithm. The moat is the 31,000 worker profiles, built shift by shift, in Birmingham and Green Bay and Huntsville — cities and workers that have never appeared in a venture-backed labor market dataset before. There is no LinkedIn for a line cook in the Heartland. There is no Glassdoor profile for a server at a regional country club. Croux has spent years earning the trust of both sides of that market, and the data that trust generates cannot be purchased or replicated on a short timeline regardless of how much compute a competitor rents from Microsoft.
Here is the deeper irony: the communities that traditional capital ignored created the scarcest data set in AI, because nobody has the data. The companies that did go there — that built real products for real customers in those markets — are sitting on something that the OpenAIs of the world would struggle to replicate regardless of their valuation. You cannot buy your way into years of trust built with workers and employers in markets that mainstream venture has overlooked.
The moat deepens as AI regulation tightens. Data provenance and consent will matter enormously. Companies that built in underserved communities often did so with explicit trust and consent frameworks from day one, because they had to earn access rather than assume it. That is an ethical and regulatory advantage that companies built on scraped internet data are about to discover they don't have.
Real customers validate the data. Marriott, Hyatt, Wyndham, and dozens of independent operators across the Heartland keep coming back to Croux not because of a clever pitch deck — because the data produces results.
Part Five: What happens to the liquidity when it arrives
What happens to the liquidity when it finally arrives matters as much as when it arrives.
The mechanics are less straightforward than they appear. Much of what VC funds distribute will arrive as stock, not cash — newly public shares in thin markets with lockup overhangs still clearing. Rebalancing requirements absorb a significant portion before new commitments are even on the table. Large institutional investors who have been over their private market allocation targets for three years face a 6 to 18 month policy and governance process before capital can be redeployed into new funds. And the secondary market has already done much of the work quietly — the most sophisticated sellers took money off the table over the past two years at peak private valuations, which means the pool of genuinely fresh capital waiting to deploy is smaller than the headline liquidity numbers imply.
The capital that does move quickly will, in all likelihood, chase the most visible names: IPO allocations, post-lockup purchases, the same AI companies whose narratives are loudest at exactly the moment the financials are becoming public. That is how liquidity cycles have always worked. The crowd moves toward the thing that just became liquid.
The question the correction surfaces — gradually, then suddenly — is which companies were building something real underneath the narrative. That question does not get easier to answer once the crowd arrives.
What this means for where capital goes next
I run an early-stage venture fund called Symphonic Capital with my co-GP Sydney Thomas. We invest in AI as infrastructure — specifically for communities that traditional capital has chronically overlooked, in health, wealth, and climate resilience. Our thesis has never been about riding the AI narrative wave. It has been about finding the companies with data nobody else has, customers who depend on them, and revenue that doesn't require a $50 billion compute bill to generate.
Yes, when these companies go public it will create liquidity that is much needed for this ecosystem to thrive. However, the liquidity, when it arrives, will be smaller than the headlines suggest, slower to redeploy than the mechanics allow, and largely directed toward the names that are loudest rather than the ones that are soundest. The correction will not be random — it will separate companies with real infrastructure value from those riding narrative multiples. Oracle's 50 percent decline is the preview. And the window to get into the companies that survive that separation does not stay open while the crowd waits for permission.
More importantly, the S-1 filings do something the last three years of private market narrative could not: they make the argument legible. Right now, the case for data-defensible AI infrastructure in overlooked markets requires an LP to imagine a world where the narrative compresses. OpenAI's prospectus makes that world real. The circular financing structures that have been propping up public stocks become public record. The contrast between companies built on compute spend and companies built on proprietary data becomes visible in a way it simply wasn't before.
The AI valuation bubble and the AI infrastructure opportunity are two different things running in parallel, often confused for each other. The first is about to face a reckoning it has been deferring for years. The second — built on real data, real customers, and communities that the circular money machine never bothered to reach — is just getting started.
The window to get into the infrastructure layer — the companies with real data, real customers, and real moats in markets that mainstream capital has ignored — is the period before the crowd gets its liquidity and figures out where to put it. That period is now.
Shruti Shah is General Partner at Symphonic Capital, an early-stage venture fund investing in AI as infrastructure across health, wealth, and climate resilience.