Lou Mannheim would not be surprised by today’s AI boom. But he would have much to say about it.
Mannheim was the senior partner at Jackson Steinem & Co., occasional mentor to an ambitious young broker named Bud Fox, and one of the few people willing to describe the excesses of the 1980s market honestly. “Quick-buck artists come and go with every bull market,” he told Fox, “but the steady players make it through the bear markets.”
Mannheim was fictional, of course, a character played by Hal Holbrook in Wall Street. His judgment was not. The same pattern appeared during the dot-com boom and again before the financial crisis. Technical and human innovation create the opportunity. Financial innovation arrives later to extend the cycle, preserve easy credit, and defer the moment when customer economics must justify the investment.
Have we reached that point with AI? Who’s to say. Even Anthropic CEO Dario Amodei concedes that AI companies “haven’t yet delivered on our big promises to benefit the world.” The investment cycle is now sustained by more than demonstrated results. Nearly every participant bears greater immediate risk from questioning the consensus than from continuing to finance it. That does not mean AI will fail or the boom will end soon. It means the traditional mechanisms that discipline credit and investment are weakening as commitments grow larger and harder to see.
What should we conclude when financing the purchase becomes more important than the product itself? Since 2022, the AI story has been about what models can do. It has quietly become a story about credit. In today’s Dispatch, I want to examine what sustains this boom and what an institution should protect before the correction comes.
The big picture
On August 10, six of the largest investment firms agreed to help Nvidia mobilize more than $500 billion for AI infrastructure through intermediaries that keep the debt off its balance sheet. Bloomberg had already identified roughly $70 billion in guarantees that could still come due. The exposure now appears much larger. A Wall Street Journal analysis found roughly $3 trillion in off-balance-sheet commitments across nine technology companies, mostly tied to AI, including leases that have not begun and contracts for future equipment and services. Many cannot easily be canceled. Underneath the jargon, the arrangement is straightforward: a special-purpose entity borrows to buy the chips and repays the debt from a customer’s lease payments; if those payments stop and new customers cannot cover the shortfall, a chipmaker or hyperscaler may provide the backstop. But the firms supporting these arrangements have substantial commitments of their own, some extending decades into the future. Risk is moving into larger, less visible pools of institutional credit while the capacity of model labs and hyperscalers to honor their leases and guarantees still depends on the same optimistic assumptions about AI productivity and future revenue.
Potential buyers of this debt are no longer uniformly convinced. Bill Eigen of J.P. Morgan told CNBC that the cycle was beginning to resemble real estate finance more than technology investment. The mismatch is difficult to ignore. Investors are being asked to buy thirty-year debt secured by data centers filled with equipment that may depreciate within five years. The debt is made more attractive by long leases and guarantees from hyperscalers such as Meta, but those firms already carry substantial commitments. It is not clear how much protection a backstop provides if falling AI demand weakens the customer, the collateral, and the guarantor at the same time. Off balance sheet, Eigen observed, does not mean the obligation has disappeared. He heard similarly confident claims about limitless fiber demand in the late 1990s.
These new financing structures do not remove credit risk. They make it harder to see. Supporters make a plausible case: AI demand will continue growing, the debt will amortize as the hardware depreciates, and cash-rich guarantors can absorb any losses. Most of these deals appear rational. The systemic concern lies in the assumption they share. Financial engineering cannot produce the productivity gains needed to justify the buildout or ensure that model labs will generate enough profit to service their debts. Adding more creditors spreads the exposure, but it does not diversify the underlying wager on AI demand. So why do credit markets remain so accommodating?
Why the consensus holds
Timur Kuran’s work helps explain why. He studied how systems maintain an aura of stability even when the private judgments within them have begun to diverge from the public consensus. He called the mechanism preference falsification: people adjust what they say publicly as conformity brings reputational benefits, while dissent imposes an immediate cost. Private doubts can grow quietly without weakening the visible consensus. The first person to challenge it bears the greatest risk.
The financing market can reinforce that logic without requiring dishonesty. Credit analysts who turn cautious about AI risk losing customers. A CIO who urges restraint about AI’s likely effects can be dismissed as behind the times. Eigen encounters the same pressure when he questions the assumptions behind the AI buildout. “And all I’m saying, and the funny thing is, when I bring any of this stuff up, people get angry with me, and that tells me something also,” he told CNBC. The anger matters because it reveals the reputational cost of challenging the consensus, even for an investor speaking within his own expertise. Being wrong with the consensus is institutionally safer than being right alone and too soon. The doubt is kept off the public ledger.
Outsiders see the commitments, not the reservations or time horizons behind them. At a private gathering of veteran investors in Maine, the doubt was explicit: no one could say whether AI earnings would justify the spending, yet no one was ready to leave the trade. The Wall Street Journal described them as holding their doubts in one hand and pressing “buy” with the other. Confidence is not judgment, and a good deal of the confidence on offer right now will prove wrong. Each new investment becomes public evidence for a thesis its participants are privately questioning. Kuran called the deeper effect knowledge falsification. Preference falsification first hides doubt; over time, its absence from public discourse makes the underlying assumptions harder to challenge. Once AI transformation becomes a planning premise rather than a claim requiring evidence, institutions stop asking whether the projection is sound and ask only how quickly to join. The responsible alternative is to participate while making commitments that protect the institution if rosy projections prove wrong.
A model for participating without ceding control
Wendell Weeks offers a different model, and not because he doubts AI. Corning supplies the optical fiber and silicon photonics used to connect the servers inside AI data centers. It has signed multibillion-dollar agreements with Meta, Nvidia, and Amazon. But Weeks, who ran Corning’s fiber-optics business through the telecom collapse after 2000, has negotiated terms that sometimes require customers to provide upfront capital for the factories and workers needed to fulfill their orders. Nvidia, for example, is investing in Corning and prepaying billions to expand production. Weeks accepts responsibility for research, innovation, and production. The uncertainty created by a customer forecasting extraordinary demand belongs with that customer and its shareholders. If a customer orders a million units and ultimately needs one hundred, Corning should not bear the full consequence. Weeks is willing to serve the boom. He will not “bet the family farm” on someone else’s overly optimistic forecast.
Siemens is taking a similar posture as it expands production of electrical and power equipment for data centers. The company is seeking longer-term supply agreements to support new capacity and aims to recover its factory investments within three or four years. It is also using third-party manufacturers for some orders, giving it room to reduce outside production before demand weakness reaches its own factories, and it is maintaining business in other industrial markets rather than becoming wholly dependent on data centers. These are different mechanisms serving the same principle. The company making the demand forecast should retain meaningful exposure if the forecast proves wrong. That is skeptical optimism. It does not require smaller ambition. It requires commitments that allow the enterprise to benefit if demand holds while protecting its people, capital, and capacity if it does not.
What this means for higher education
Higher education cannot negotiate from Corning’s market position or manage factory capacity as Siemens does. But we can apply the same discipline. Corning requires customers to help fund the capacity their forecasts demand. Siemens seeks longer commitments, uses outside production for flexibility, and protects lines of business that do not depend on data centers. Both distinguish between the capabilities they control and the demand risk their customers create. Universities should do the same. We should invest most confidently in capabilities that retain value across vendors and market corrections: governed institutional data, process knowledge, integration capacity, and the judgment of faculty and staff. Contracts should place appropriate obligations on vendors. Architecture should limit the consequences if a forecast proves wrong and preserve a practical way out. The commitments deserving the most scrutiny are those tied to one vendor’s adoption forecast, pricing model, or technical design. The question for university leaders is simple: if this product, its pricing, or this vendor changes materially in three years, what obligations, disruptions, or stranded investments will the university be left to absorb?
The same principle applies to our students. Universities must prepare them for an economy shaped by AI, but we should not confuse proficiency with today’s tools for preparation across a career. AI is beginning to absorb some of the entry-level work through which people once learned a domain, encountered exceptions, and acquired professional judgment. Stanford researchers have reported a 19 percent relative decline in employment among workers aged 22 to 25 in occupations most exposed to AI. The first rung is where professional judgment is formed. Our responsibility is therefore larger than teaching students to operate a particular model. We must give them disciplinary depth, the ability to frame consequential problems, and enough judgment to evaluate uncertain outputs and stand behind the decisions that follow. Those capabilities will remain valuable whether the current investment boom continues, contracts, or eventually gives way to a more durable AI economy.
The final word
AI will matter. That does not make every forecast sound or every investment prudent. Kuran explains why public confidence can outrun the private judgment beneath it. Weeks and Siemens show how to participate in a boom while keeping its risks with those best positioned to bear them. For universities, the work is to invest in data, process knowledge, judgment, and infrastructure that outlives the hype cycle. Good governance is not a wall against new tools. It produces the judgment to use them and keeps others’ optimism from becoming an obligation the institution cannot carry.
Good article, you make a lot of good points. A lot is being made of the lease and purchase commitments being made off the books. The insinuation is that these commitments have to be paid. But isn't the reality that all such deals have exit terms, and that there will be some type of settlement for significantly less than the contracted amounts if the business goes south? The holder of the commitment is incentivized to settle, especially if they don't have to deliver anything, just take the settlement money.
Additionally, many of these deals are driven by sales to end customers that have already been made; won't those customers have the same type of commitment to fund even if they end up not using the services? Again, settlements will be reached.
So while if the AI industry goes down the tubes it will suck for everyone, that risk is being spread out across a huge chunk of the economy. The pain will not be felt only by the AI vendors, but by our whole economy.
Rewiring our economy around a new tech (which is what we are doing & why it requires such gigantic investment) is inherently a very risky proposition.