The Foam Is Not the Beer

This piece asks what happens when the AI narrative, the egregore that has been pricing the cycle, meets the public market, when the story gets a ticker symbol and has to report its numbers every ninety days.

1. The Foam and the Glass

When you pour beer into a glass, roughly 20 to 30 percent of what fills the space is foam. The foam is real. It smells like beer. It is produced by beer. But it is not the drink you paid for. If you hand someone a full glass and tell them it contains beer, you are not lying, but you are not telling them the whole truth either.

The AI IPO cycle is a glass being handed to public investors with the foam measured as liquid.

Two days ago, the Wall Street Journal reported that OpenAI missed its first-quarter revenue and user growth targets, falling short of its goal of one billion weekly active users by approximately 100 million, and missing multiple monthly revenue targets. The company called the report “clickbait.” The market did not wait for the statement. SoftBank, which has outlined plans for up to roughly $60 billion of OpenAI exposure through staged investments, fell nearly 10 percent in Tokyo trading, losing approximately $18 billion in market value overnight. Oracle, holding a $300 billion five-year compute partnership with OpenAI, fell 4 percent. CoreWeave fell 6 percent. Nvidia fell 3 percent. AMD and Broadcom fell 4 percent each. ARM Holdings fell 6 percent.

A private company, not yet publicly traded, caused a multi-hundred-billion-dollar contagion event across global equity markets, on a single revenue miss report from the Wall Street Journal.

This is the AI egregore doing the pricing work. When the story wobbles, everything priced on the story wobbles with it. The question investors should be asking is not whether the market reaction was proportionate. It is what happens when the company is public and required to produce those numbers every ninety days, under legal obligation, for the indefinite future.

2. The Glass Before the Pour

OpenAI recently closed a $122 billion funding round at a post-money valuation of $852 billion. It has committed to spending $1.4 trillion over eight years on data center infrastructure. Its current annual revenue is approximately $13 billion. Its CFO Sarah Friar has expressed concerns, documented in the WSJ report, about whether the firm can afford the compute contracts it has already signed.

Let that structure sit. A company valued at $852 billion, generating $13 billion in revenue, committed to $1.4 trillion in spending over eight years, currently missing its own internal quarterly targets, is preparing to enter the public market.

The Federal Reserve has already responded to the broader environment these valuations exist inside. In the minutes of the October 2025 FOMC meeting, policymakers stated: “Some participants commented on stretched asset valuations in financial markets, with several of these participants highlighting the possibility of a disorderly fall in equity prices, especially in the event of an abrupt reassessment of the possibilities of AI-related technology.” Fed Chair Jerome Powell said in September 2025: “By many measures, equity prices are fairly highly valued.” Fed Governor Lisa Cook added in November: “Currently, my impression is that there is an increased likelihood of outsized asset price declines.”

These are not outside critics speculating. These are the officials responsible for financial stability putting specific warnings about AI valuations into the official record of the world’s most influential central bank.

The S&P 500’s cyclically adjusted price-to-earnings ratio has exceeded 39, a level last seen during the dot-com bubble, occurring in fewer than 3 percent of months since the index was created. The Magnificent 7, Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia, and Tesla, represent approximately 34.5 percent of the S&P 500, the greatest concentration in the index’s history. The market is not broadly overvalued. It is narrowly overvalued, in precisely the sector where the AI egregore has done the most work and where the IPO pipeline is thickest.

3. How IPOs Used to Be Priced, and What Is Broken Now

Before a company goes public, its offering price has historically been derived through three converging methodologies. These exist not as bureaucratic ritual but as the accumulated discipline of markets learning, repeatedly and painfully, what happens when enthusiasm substitutes for analysis.

Discounted cash flow analysis requires a credible path from current losses to future returns. It is unforgiving of businesses where the cost of serving each additional customer grows faster than the revenue that customer generates. As I documented in Who Gets Paid When You Talk to AI, the model company’s revenue routes immediately through a stack of infrastructure rent to entities whose contracts are secured regardless of the model’s performance. What remains after those payments at the consumer subscription tier is currently negative. Under any reasonable set of assumptions, there is no honest DCF that bridges that cost structure to an $852 billion valuation.

Comparable company analysis prices an offering against peers at similar stages growing at similar rates. SpaceX is being discussed at approximately 125 times sales, not earnings, sales. Google went public at 10 times sales while growing at 240 percent annually. Scott McNeely, CEO of Sun Microsystems, said in 2002 about his own company’s 10-times-sales valuation: “In order for me to return your money in ten years, I would have to pay out all of our sales as dividends, with no staff, no R&D, no manufacturing costs.” We are now discussing 125 times sales for a company analysts expect to grow at roughly 25 percent annually. When the investment banks presenting these roadshows declare that there are no valid comparables, that is not a sign of extraordinary uniqueness. It is a sign that the comparison has been set aside because it cannot be survived.

Precedent transaction analysis asks what similar companies sold for and what returns those buyers subsequently realized. The relevant precedents here are Pets.com, Webvan, and the late SPAC cycle, companies that went public at narrative valuations before their unit economics were proven, whose investors absorbed the mean reversion when the story met the balance sheet. These methodologies are the market’s memory. Abandoning them does not change what they were designed to prevent.

4. Shadow Borrowing and the New Subprime Structure

The financing architecture underneath the AI buildout is where the structural risk is most precisely located, and where it most closely resembles a mechanism investors have encountered before.

The Bank for International Settlements published a paper on March 16, 2026 documenting how hyperscalers have moved significant portions of their infrastructure financing off their balance sheets through special purpose vehicles and private credit arrangements. The BIS named this “shadow borrowing.” Meta, Oracle, xAI, and CoreWeave have used SPVs to shift over $120 billion in data center financing debt into vehicles where it is visible to neither the hyperscaler’s headline balance sheet nor the retail investor’s standard due diligence. The hyperscaler holds a minority stake and commits to long-term operating leases, capturing the infrastructure benefit while distributing the debt obligation into future accounting periods. Morgan Stanley estimates global data center spending between 2025 and 2028 at approximately $3 trillion, roughly half covered by private credit.

The BIS identified new shock transmission channels in this structure, paths through which financial stress in AI infrastructure financing could reach private credit funds, insurance companies, and pension capital. The transmission chain is specific: pension capital flows into private credit products, which fund data center SPVs, which provide AI infrastructure, which generates the compute revenue that supports the AI company valuations that the pension is also exposed to through index holdings. When AI demand disappoints, stress hits the SPV, moves through private credit, and reaches pension and insurance exposure, long before “AI risk” appears on any dashboard labeled AI risk.

Rajat Rana, a partner at Quinn Emanuel who worked on structured finance litigation after 2008, described this as “the largest peacetime investment project in human history, financed largely off balance sheet,” adding that tracking developments felt like “deja vu.” Steve Eisman, who anticipated the 2008 credit crisis, has stated explicitly that a credit cycle is coming in the software sector, noting BDC funds with high software exposure have already underperformed peers by approximately 5 percentage points since October 2025.

The parallel to 2007 is not that AI is subprime. The underlying technology is real. The parallel is structural: the mechanism of risk transfer, moving debt obligations into vehicles where they are invisible to the parties ultimately bearing the exposure, is identical. In 2007 the asset was residential mortgages. The vehicle was the CDO. The ultimately exposed parties were pension funds and insurance companies who did not fully understand what they had purchased. The BIS is describing the same transmission architecture applied to a different asset class. As in every prior cycle, the documents will enter the public record. They always do.

5. Who the IPO Really Pays: The Hidden Hands

The clearest signal in any investment cycle is not the price of the asset being promoted. It is how the people selling the shovels have structured their own exposure.

Microsoft invested $13 billion in OpenAI. OpenAI runs on Azure. Every API call is Azure compute revenue. Microsoft’s investment appreciates as OpenAI’s valuation rises, OpenAI’s valuation rises as enterprise adoption grows, enterprise adoption requires Azure infrastructure, Azure revenue funds Microsoft’s position. If OpenAI succeeds, Azure scales. If OpenAI fails, Microsoft writes down the investment and retains the enterprise relationships the partnership built. The risk profile is asymmetric by design.

Amazon invested $4 billion in Anthropic. Anthropic runs on AWS and distributes through Bedrock. The investment functions as a customer acquisition cost structured as venture capital, Amazon collects infrastructure rent on Anthropic’s inference regardless of Anthropic’s profitability. Google holds stakes in Anthropic while running its own models on GCP, distributing through Search and Android. Google is hedged on every side simultaneously: it cannot lose at the infrastructure layer regardless of which model wins the narrative war. Nvidia sells GPUs to all of them, collects margins exceeding 70 percent on the hardware embedded in every inference dollar, and has no material exposure to whether any single model generates returns.

The sovereign wealth funds and late-stage investors who entered at private valuations that require the IPO narrative to hold are not making a bet on the technology. They are making a bet that the public market will provide the exit the private market cannot, that enough retail investors will buy the story at sufficient intensity to allow early positions to be liquidated at a premium.

This is a marionette market. The model companies perform on stage. The hyperscalers, infrastructure lenders, and early investors collect on every movement regardless of how the show goes. The IPO is the moment the audience is invited in, after the stage has been built, the contracts signed, and the exit structured. The hands are not hidden. They are documented in the investment disclosures and partnership announcements. They simply require reading past the narrative to see.

6. SpaceX: Dress Rehearsal for an Egregore IPO

SpaceX is the clearest current case of narrative-driven, structurally engineered valuation, and it is worth examining in detail because it shows the mechanics operating in real time before the public market opens.

The company is targeting a valuation of approximately $2 trillion at roughly 125 times sales, analysts expect revenue growth of roughly 25 percent annually. Patrick Bole, professor at King’s College London and former hedge fund manager, has described the IPO framing as a “scandal”, not because SpaceX is a bad company, but because of the specific mechanisms being used to engineer demand.

NASDAQ changed its rules to allow SpaceX fast-track index inclusion within 15 days of IPO. Standard seasoning requirements are approximately one year, designed to allow the market to form an independent price through real buyers and sellers before the company is inserted into passive investment vehicles. Reuters reported this fast-track inclusion was negotiated directly with NASDAQ. The consequence is specific: at a projected weighting above 4 percent of the NASDAQ 100, index funds and benchmark-tracking managers who cannot afford to be underweight that position are forced to buy regardless of their valuation judgment. They are not making an investment decision. They are making a tracking decision. The foam is poured into their glass whether they ordered it or not.

The valuation has been expanded sequentially by adding narratives: space launch, then satellite internet, then data centers in space, then xAI acquisition, then cursor acquisition. Each addition is described as a transformational business. Each serves the same function: extending the multiple by adding “AI inside” as a premium justifier. The question Bole asks is the right one: when you strip away the narrative additions and apply standard comparable analysis, what does the business produce at what margins growing at what rate, and what has the market historically paid for that?

The answer, at 125 times sales, does not survive the comparison.

7. Workslop and the Productivity Gap

The narrative valuation of AI companies rests partly on a specific economic promise: that AI will generate measurable productivity improvements at enterprise scale, justifying both the adoption costs and the infrastructure investment. The evidence to date does not support that promise at the scale being priced.

A National Bureau of Economic Research study published in February 2026 found that despite 90 percent of firms reporting no measurable AI impact on workplace productivity or employment, executives projected AI would increase productivity by 1.4 percent and output by 0.8 percent. The gap between reported reality and executive projection is not evidence of future productivity gains. It is evidence of the egregore operating inside the C-suite, the narrative so thoroughly internalized that projections diverge from observed results without triggering revision.

The operational costs behind that gap have been named. Research from Stanford’s Social Media Lab and BetterUp, published in September 2025 and covering 1,150 full-time US desk workers, documented a phenomenon they called workslop: AI-generated content that appears polished but is incomplete, inaccurate, or unhelpful, requiring significant human correction before it is usable. Approximately 40 percent of surveyed workers reported receiving workslop in the prior month. Each incident required an average of one hour and fifty-one minutes to resolve, twenty minutes longer than if the sender had done the work themselves. Translated to scale, the hidden tax of workslop amounts to $186 per month per employee, or over $9 million annually for a 10,000-person organization.

This is one large-scale study, not the final word on AI and productivity, but it captures the shape of a problem many teams are now reporting anecdotally and it puts a dollar figure on a cost that does not appear in any AI adoption business case. AI-themed productivity gains that look like progress in the headline metric, time to first draft, queries handled per hour, generate hidden costs in the correction cycle that do not appear on the dashboard most executives see. The productivity the egregore promises and the productivity the balance sheet eventually reflects are two different numbers. For companies going public on the promise of the first, the reckoning arrives when the second must be reported.

8. Policymakers Behind the Business Model

The regulatory gap in AI is not primarily a technology gap. Technology moves faster than legislation in every cycle. The specific gap that matters here is the business model gap, the distance between the financing structures being used and the oversight frameworks designed to govern them.

The BIS paper is a regulatory institution documenting risk it cannot yet fully measure. The Bank of England has flagged data gaps and systemic risk in private credit markets through its financial stability work. The IMF has drawn explicit comparisons to the dot-com bubble, warning that a market correction could stunt global growth and weaken developing economies. The FOMC minutes contain the specific phrase “disorderly fall in equity prices.”

These are not alarmist voices. They are the institutions whose function is to see the full picture before the market does. When multiple global financial regulators are publishing coordinated warnings about the same asset class simultaneously, the appropriate response for an investor is not to dismiss the warnings as excessive caution. It is to ask what they are seeing that the narrative does not accommodate.

The policymakers are not behind the technology. They are behind the business model, behind the SPV structures, the off-balance-sheet commitments, the private credit transmission channels, the index inclusion mechanics, and the valuation methodologies that have been set aside to allow a story to be priced as an asset. That is where the systemic risk lives, and that is what quarterly earnings will begin to illuminate the moment these companies are required to report in public.

The glass is being handed to you. The BIS has named the transmission channels. The Fed has named the valuation concern. Eisman has named the credit cycle. The NBER has named the productivity gap. The workslop research has named the hidden correction cost. The NASDAQ rule changes have documented the index engineering. The SPV structures have been reported.

None of this requires a prediction about when. Markets reprice on their own schedule. The egregore can hold its intensity longer than any short position can remain solvent, as the aphorism goes.

The only question that matters for an investor reading this is whether you are buying the beer or the foam, and whether, having read both the S-1 and the BIS paper, you are comfortable with the answer.

Resources:

• https://www.insurancejournal.com/news/international/2026/03/17/862128.htm
• https://www.insurancejournal.com/news/international/2025/12/11/850724.htm
• https://www.bis.org/publ/qtrpdf/r_qt2603u.htm
• https://finance.coin-turk.com/ai-infrastructure-investment-alters-financial-dynamics/
• https://www.primointeractive.com/blog/openai-misses-targets-microsoft-exclusivity https://www.reuters.com/business/openai-falls-short-revenue-user-targets-it-races-toward-ipo-wsj-reports-2026-04-28/ https://europeanbusinessmagazine.com/softbank-has-bet-64-billion-on-one-company-and-the-credit-agencies-are-circling/
• https://www.perplexity.ai/search/907f464c-49a2-4b3a-b362-733195320276
• https://www.betterup.com/workslop
• https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity
• https://www.nber.org/system/files/working_papers/w34836/w34836.pdf
• https://news.ycombinator.com/item?id=47388640

Originally published at https://fafi25.substack.com.

A message from our Founder

Hey, Sunil here. I wanted to take a moment to thank you for reading until the end and for being a part of this community. Did you know that our team run these publications as a volunteer effort to over 3.5m monthly readers? We don’t receive any funding, we do this to support the community.

If you want to show some love, please take a moment to follow me on LinkedIn, TikTok, Instagram. You can also subscribe to our weekly newsletter. And before you go, don’t forget to clap and follow the writer️!