Can AI’s sky-high valuations survive disclosure?

Within a single year, artificial intelligence (AI) is completing a transition that took the internet economy nearly a decade: from a venture-funded experiment to a publicly traded asset class. 

SpaceX, with xAI (now rebranded as SpaceXAI) folded in, listed on Nasdaq in June, raising US$75 billion in the largest initial public offering (IPO) in history, and with its market value surging past US$2 trillion at one point. Anthropic confidentially filed its prospectus with the Securities and Exchange Commission (SEC) on 1 June, days after closing a private round at a US$965 billion valuation; OpenAI, last valued at US$852 billion, followed a week later. 

In Asia, MiniMax and Z.ai debuted in Hong Kong in January; Moonshot AI has circulated a shareholder resolution for a Hong Kong listing that could price within six months, on the heels of a US$3.5 billion round valuing it at US$35 billion; and DeepSeek is reportedly eyeing Shanghai’s STAR Market in 2027. Meanwhile, memory chipmaker CXMT surged 466% on its Shanghai debut on 27 July, vaulting past ICBC to become China’s most valuable onshore-listed company — even as more than US$1 trillion was wiped off global chip stocks in the same week, with SK Hynix falling double digits despite posting record quarterly profits.

Euphoria and panic, coexisting in the same news cycle. Is this a bubble, or a market doing its job? Both dynamics are at work, and distinguishing them requires looking at what an IPO actually does — informationally, institutionally and economically.

Disclosure: turning on the lights

Much of my research concerns how markets produce and aggregate information, and from that vantage point, the significance of this listing wave is not primarily the capital raised. Moonshot does not need public money to survive, and neither does OpenAI, which sits atop tens of billions from SoftBank and, reportedly, a proposed US$250 billion financing backstop from Nvidia tied to a massive data-centre buildout.

What going public does is subject AI’s economics to mandatory disclosure and continuous, adversarial price discovery. Private AI valuations have been set in negotiated rounds among a small club of strategic investors — chipmakers, cloud providers, sovereign funds — many of whom are simultaneously customers, suppliers or competitors of the companies they fund. This “circular financing” is precisely what unnerved markets last week: when a chipmaker invests in a model developer, which buys the chipmaker’s processors, which books the revenue that supports the chipmaker’s valuation, the informational content of any single price in that loop is degraded.

Securities disclosure requirements break part of this circularity, and they are our best instrument for probing whether these companies are bubbles. A registration statement and subsequent quarterly filings compel what private markets never did: audited financials, management’s discussion of margins and cash burn, risk factors, segment reporting, customer-concentration disclosures, related-party transaction details, and the fine print on purchase obligations and vendor-financing arrangements. 

Anthropic’s public prospectus, when it emerges from SEC review, will show how much of its reported US$47 billion revenue run rate — up from US$10 billion a year earlier — is durable enterprise demand versus credits recycled among strategic partners. Moonshot’s Hong Kong prospectus will do the same for its roughly US$300 million in annual recurring revenue, and CXMT’s filings already reveal a 7.67% global DRAM share against which its US$488 billion market value can be judged. 

Bubbles thrive on opacity and narrative; disclosure regimes exist to replace narrative with verifiable numbers, and short sellers and analysts will do the rest. In this sense, the sell-off and the IPO wave are not contradictory. They are the same phenomenon: the market beginning to demand fundamentals where it previously accepted stories.

A crucial caveat, however: disclosure is necessary for price discovery but not sufficient. Public equities have hosted some of history’s greatest bubbles — the dot-com mania unfolded entirely under the SEC’s disclosure regime — because prices reflect not only information but sentiment, and the two can decouple for uncomfortably long stretches.

Spillovers, comparables and the transmission of (mis)pricing

How will Moonshot’s IPO affect other AI listings? IPO research offers a clear answer: pioneering listings in a new sector generate powerful informational spillovers. First movers resolve uncertainty about investor demand, regulatory treatment and appropriate valuation levels that followers then free-ride upon. This is one reason IPOs cluster in waves, and why sequencing is a strategic decision.

The mechanism runs through the workhorse of practical valuation: comparables and relative pricing. When no listed pure-play AI model company exists, bankers and investors have no market-based anchor; once Moonshot trades, its revenue multiple, its growth-adjusted premium or discount, and its aftermarket volatility become the reference grid on which Zhipu, MiniMax’s follow-on offerings, and eventually DeepSeek are priced.

Relative valuation is informationally efficient — it economises on the impossible task of forecasting decades of cash flows for a technology whose applications are still being discovered. But it has a well-known dark side: comparables transmit errors as readily as information. If the pathfinder is mispriced, the entire comp set inherits the mispricing, producing the correlated booms and busts characteristic of sector waves. 

A Moonshot that prices well and trades stably would validate the “AI plus Hong Kong” thesis and re-rate the pipeline upward; a broken deal — pricing below the last private round, or a swift descent through the issue price — would chill listings for a year or more, as high-profile disappointments did for Chinese tech in 2021-2022.

Moonshot is an unusually informative test case for two further reasons. It is dismantling its variable interest entity (VIE) structure in favour of a joint-venture model aligned with regulatory preference, a template that, if successful, lowers the legal-uncertainty discount that has weighed on Chinese tech equities for two decades. And its Kimi K3 model is open-weight, forcing investors to confront a question markets have deferred: how do you value a frontier model whose weights anyone can download? The answer — that value resides in distribution, data flywheels and serving infrastructure rather than in the model artefact — must then be reflected in the multiples applied across the sector.

Public markets manufacture sentiment too

Only a small fraction of shares actually floated; mainland households sit on vast savings with an impaired property channel and few attractive alternatives, and the stock offered the first meaningful onshore vehicle for owning both the memory supercycle and the national self-sufficiency drive. Add short-sale constraints — newly listed shares are difficult or impossible to short in mainland and Hong Kong markets, and expensive to short anywhere — and the classic ingredients of speculative overpricing are present: when optimists can buy freely but pessimists cannot easily bet against them, prices are set by the most enthusiastic marginal buyer, and part of what buyers pay for is the option to resell to someone even more enthusiastic.

The US mega-listings carry their own structural accelerants. Tiny initial floats relative to headline market capitalisations mean first-day prices are set on slivers of supply; lockup expirations then test those prices months later. Index inclusion mechanically conscripts passive capital — retirement funds must buy a newly indexed trillion-dollar company at whatever price momentum has set, regardless of any view on fundamentals. And in the AI sector specifically, cross-holdings among listed companies mean sentiment in one name transmits to others through balance sheets, not just through comparables. 

So the honest statement is this: the listing wave improves the production of information about AI companies enormously, while simultaneously exposing their pricing to retail flows, index mechanics and momentum-chasing that private markets, for all their opacity, largely excluded. Disclosure gives sober investors the tools to identify mispricing; it does not guarantee they set the price, particularly early in a stock’s life.

Regulation as an input to innovation, not just a constraint

The role of regulators in this wave deserves more attention than it typically receives. In research with Murillo Campello and Diemo Dietrich on regulatory uncertainty and financial-technology innovation, we show that when regulators must expend resources to understand a new technology, the pace of innovation depends not only on entrepreneurs but on the budget, skill and preparedness of the regulator itself. Regulatory competence and private innovation are complements: firms invest more when they expect regulators to evaluate novel business models quickly and predictably, and under-resourced regulators can trap an industry in a low-innovation equilibrium even without ever issuing a prohibition.

The current listing geography illustrates the point. Hong Kong’s listing reforms for specialist technology companies, the securities regulator’s clarified guidance on offshore structures, and Shanghai’s STAR board have together converted regulatory ambiguity into a legible pathway — and Chinese AI firms are responding exactly as the theory predicts, with a queue of listings. 

In the US, a more accommodating securities-review posture has similarly opened the window for OpenAI and Anthropic. But the same framework counsels caution: valuations that partially capitalise a “regulatory blessing” are exposed to shifts in that blessing, whether from export-control escalation abroad or from a change in supervisory posture at home. Investors should treat regulatory capacity and predictability as a priced fundamental of emerging-technology firms, not as background noise.

Fundamentals: how foundation models actually earn money

Ultimately, bubble questions resolve into a question about business models: how do these companies price what they sell? Foundation-model providers sell intelligence in tokens — charging dollars per million tokens across input, output, cached and reasoning types — making the token a unit of information, computation and account, much as the kilowatt-hour is for electricity.

In my recent work on AI tokenomics, I show that these token pricing structures are not mere billing conventions: they are instruments through which providers approximate application-specific pricing across thousands of downstream uses they cannot directly observe. They also shape which applications are developed in the first place, because the same tokens are used for both experimentation and real-world deployment.

This framework clarifies several facts that superficially feed bubble fears. Headline token prices have fallen relentlessly, which sceptics see as evidence that AI is becoming a commodity. But because experimentation consumes the same priced input as production, cheaper tokens expand the set of viable applications: total token demand can grow faster than unit prices fall when many latent applications sit near the profitability threshold — a Jevons-type dynamic, and a legitimate reason revenue at leading labs has compounded even as prices dropped. 

At the same time, the framework identifies where pricing power is fragile: competition squeezes margins most on portable workloads that customers can redirect across providers and routers, while ecosystem-bound workflows — through deep integrations, accumulated context or enterprise trust — remain more defensible. A model with lower token prices can still cost more to complete a task, so competition is increasingly centred on the quality-adjusted cost of completing the task, rather than headline token prices.

For public investors, this yields a concrete due-diligence agenda that disclosure now makes feasible: What share of revenue comes from portable, easily re-routed API traffic versus locked-in enterprise workflows? How are margins trending per task rather than per token? How much revenue is vendor-financed or recycled from strategic partners? Companies that can answer well deserve premium multiples; those that cannot are renting a moment of benchmark glory. 

The demand underlying AI is real — memory capacity is sold out industry-wide through 2026, and record chip profits attest to it. What the recent sell-off punished was not absent demand but prices that assumed a permanent plateau at a cyclical peak, financed through loops that obscured where the money originated.

The right conclusion is not to call the top, which no one can do reliably, but to understand what this transition does and does not deliver. The listings of 2026 and 2027 will generate the disclosures, the comparables, and yes, the disappointments that allow capital to discriminate between AI companies with defensible data, distribution and pricing economics, and those without. But that discipline arrives on a lag — through earnings seasons, lockup expirations, and the slow accumulation of short interest — while sentiment, scarcity and flow can dominate pricing in the interim, as CXMT’s debut just demonstrated. 

Bubbles form readily in the dark; they can also form in full daylight, when enough investors decide the light is telling them what they want to hear. The listing wave is turning on the lights. Whether we choose to read by them is up to us.

Related: How China split Silicon Valley over AI | What SpaceX’s IPO means for China