The right bubble about the right technology

25 min read

The question "are we in an AI bubble in 2026?" has an arithmetic answer before it has a philosophical one. David Cahn's analysis, from Sequoia Capital, published on July 8, quantifies what until now was market intuition: a single year of investment in AI data centers requires, over the useful life of that equipment, roughly $1.5 trillion in revenue from end customers; adding up the years since ChatGPT, the cumulative bar sits around $3 trillion — against approximately $80 billion in combined annual recurring revenue from OpenAI and Anthropic, a gap of $2.9 trillion between what has been built and what is being paid for [1]. The short, quotable answer is this: yes, there is a bubble, but it isn't the bubble the word suggests. It isn't a bubble of useless technology, like tulips; it's a bubble of duration, like railways — capital with quarters-long maturity financing adoption with decades-long maturity.

The natural reflex when faced with a number with twelve zeros is to ask when it bursts. That is the wrong question, or at least the less interesting one. The right question is what this gap measures — and the answer isn't found in supply-side balance sheets, where everyone is looking, but in demand-side org charts, where no one is looking. Because the most revealing data point this month doesn't come from San Francisco or Wall Street: it comes, uncoordinated and without citing each other, from three management publications — the Harvard Business Review, twice, and the MIT Sloan Management Review — describing AI-buying companies held back not by the technology, but by themselves [5][6][7]. Supply has issued commitments against revenue the buyer doesn't yet have the internal structure to generate.

There are, therefore, two clocks running in different time zones here. One is measured in quarters: compute obligations in the hundreds of billions of dollars, credit instruments designed to lighten balance sheets, power plants reopened to feed data centers [4][2][3]. The other is measured in decades: the speed at which a real organization — with its operating model, its risk culture, its scattered pilots — can convert a general-purpose technology into bookable value [8][5]. Cahn's $2.9 trillion gap is the distance between these two clocks, expressed in dollars. And economic history suggests that when that distance is forced to close, it isn't the technology that pays — it's the capital structure that financed it.

The pattern isn't new; it is, in fact, one of the best-documented in the history of capitalism. The British railway mania of the 1840s ruined a generation of investors and left standing a network that served the country for a century; the fiber-optic overbuild of the 1990s vaporized entire market caps and left installed the capacity on which the commercial internet eventually ran. Carlota Perez formalized this cycle in the distinction between the installation phase — when financial capital overbuilds, speculatively — and the deployment phase, when the real economy slowly absorbs what was built. In both historical cases, investors were right about the technology and wrong about the timeline. This year's report on the frontier of work explicitly claims this genealogy: the transformation underway, in the image of the Industrial Revolution and the internet era, will take decades to fulfill its promise, because it involves technological, social and economic change that cannot be compressed by financial decree [8]. Decades — while the capital commitments signed this year demand revenue in years.

It is against this backdrop that the collision between the two dominant readings of the moment becomes instructive, precisely because both start from the same facts. The Economist's Schumpeter column from January describes the innovations in energy and credit surrounding data center construction — instruments designed to relieve the power grid and tech companies' balance sheets — as additional air pumped into the AI bubble [2]. The same Economist frames the moment within a broader distortion: a global economy where capitalism has become an instrument of power ("Gunboat capitalism," in its editorial's title), where political interference reshapes the world's largest companies and compresses their profits, and where even central banks are under attack [2]. In an economy this politicized, the reading suggests, AI asset prices have stopped being a reliable signal — they are, in part, artifacts of financial engineering and state sponsorship.

The Wall Street Journal Weekend of January 10 documents exactly the same phenomena and reads them with the opposite sign [3]. Three nuclear plants that were slated for closure are being reopened; demand for gas plants has surged; grid connection queues are getting longer. For this reading, none of this is air: it is load. No one revives a nuclear plant on short-term speculation — the cost of capital and the regulatory horizon of a reactor only make sense against real, measurable, growing electricity demand. Where the Economist sees instruments stretching a euphoria, the WSJ sees the physical economy responding rationally to actual consumption. The collision is pure: the facts coincide, the verdicts oppose each other. And that is precisely why the dispute isn't resolved by looking at supply — it's resolved by calling in a referee that neither side summoned.

Before the referee, though, it's worth measuring the strain on the builder's side. The Wall Street Journal of January 17 shows OpenAI committed to hundreds of billions of dollars in compute capacity while simultaneously turning to advertising — a business model its own leadership had described as a last resort — to diversify revenue that subscriptions don't cover [4]. The gesture is worth more than any projection: a company doesn't embrace the model it swore to avoid when current revenue keeps pace with the commitments made. Cahn's arithmetic — $80 billion in combined annual recurring revenue from the sector's two leaders against nearly $3 trillion in cumulative revenue that already-built infrastructure requires [1] — isn't a thesis; it's the explanation for the observed behavior. Supply is assuming, in its capacity contracts, a revenue curve the buying market hasn't yet drawn.

And this is where the referee comes in — the most significant convergence of the month, all the more credible for being unintentional. The Harvard Business Review of November-December 2025 recommends that companies stop multiplying AI pilots and concentrate effort on a single area where the technology protects or creates hard-to-copy advantage [5]. The January-February 2026 issue locates the failure in the misalignment between innovation ambitions and operating models, warning, in the words of Darrell Rigby and Zach First, that "Too much change can traumatize your organization" [6]. The MIT Sloan Management Review's Winter issue closes the triangle: AI risk management within organizations is lagging, write Öykü Işık and Ankita Goswami, not due to technical limitation but due to internal cultural and structural problems [7]. Three publications, three independent diagnoses, one single conclusion: the bottleneck on AI value within companies is organizational, not technological. Demand exists — companies are investing, piloting, reorganizing — but it advances at the speed organizations change, which is the speed it has always been.

This is why the artificial intelligence lens reveals what the financial lens, alone, cannot. Read as a market phenomenon, the 2026 story is a duel of multiples and sentiment — bulls against bears, with the same charts. Read as a technology adoption phenomenon, the story changes nature: the $2.9 trillion gap stops being a verdict on AI and becomes a verdict on the maturity of the demand financing it [1][5][7]. None of this month's sources connects the two ends — Sequoia measures the gap on the supply side; HBR and MIT SMR explain, without knowing it, where that gap lives. The juxtaposition implies what no one writes: the gap isn't a lack of demand, it's demand with the wrong maturity for the capital structure serving it. For a board, the distinction is everything — because it determines whether the risk to manage is that of betting on a failed technology, or that of being exposed, as creditor, customer or competitor, to balance sheets that issued quarters-long debt against decades-long revenue. When the correction comes, it won't pick out the companies with the worst AI. It will pick out the companies with the worst timeline.

What the AI lens makes visible: four patterns the financial reading misses

The first pattern this lens reveals is the financialization of compute — the transformation of computing capacity, a technological asset, into a financial asset with maturities, creditors and duration risk. David Cahn's analysis at Sequoia quantifies it with arithmetic that dispenses with adjectives: nearly $3 trillion in cumulative revenue required by AI infrastructure built since ChatGPT (about $1.5 trillion for a single year of investment alone) against approximately $80 billion in combined annual recurring revenue from OpenAI and Anthropic, leaving a gap of $2.9 trillion between what has been built and what is being paid for [1]. Without the AI lens, this number reads as a valuation ratio — another episode in the long history of market over-optimism. With the lens, it reads as something else: the creation of an entire asset class — data centers, chips, energy contracts, structured credit instruments — whose ultimate collateral is a corporate adoption curve that no one controls and that, as this month's sources demonstrate, moves at the speed of organizations and not at the speed of balance sheets. The Economist, in its Schumpeter column of January 17, denounces precisely the credit and energy innovations designed to relieve balance sheets and the grid as air pumped into the bubble [2]; what the column doesn't say, and the lens shows, is that these innovations are an attempt to stretch the duration of the money without admitting that the original duration was wrong.

The second pattern is the physical materialization of demand — the moment a technology thesis leaves the Nasdaq and enters the power grid. The Wall Street Journal Weekend of January 10 documents three nuclear plants that were slated for closure and are now reopening, surging demand for natural gas plants, and long queues for grid connection [3]. These are the same facts the Economist interprets as bubble symptoms [2] — and the collision between the two readings is, in itself, the month's most valuable data point. Because a purely financial bubble doesn't revive nuclear plants: a plant is reopened to serve real, contracted load, with decades-long horizons. What the AI lens reveals is that electricity demand is the one point in the chain where adoption is already measurable in megawatts rather than promises — and that, paradoxically, it is also the point where the clocks least coincide, because a nuclear plant depreciates over decades while the capital justifying it demands returns in quarters. Where the Economist sees air and the WSJ sees load, the lens sees a duration contract signed by two parties who measure time differently.

The third pattern lives on the supply side, and is called revenue improvisation. The Wall Street Journal of January 17 shows OpenAI committed to hundreds of billions of dollars in compute capacity while simultaneously turning to advertising — a business model its own leadership had described as a last resort — to diversify revenue that subscriptions don't cover [4]. Seen through the financial lens, it's prudent diversification; seen through the AI lens, it's a structural confession. When the dominant supplier of a frontier technology embraces the revenue model it had sworn to avoid, the signal is not one of demand strength — it's that contracted revenue isn't keeping pace with the commitment made, and that the company is buying time with the one asset it has in abundance: attention. Advertising monetizes users; enterprise subscriptions monetize transformation. The migration from the second to the first, even partial, indicates where the curve is failing — and it isn't on the users' side.

The fourth pattern is the most important, because it explains the previous three: the organizational bottleneck of demand, now documented in triplicate. The Harvard Business Review of November-December 2025 says to stop multiplying AI pilots and concentrate effort on a single area where the technology protects or creates hard-to-copy advantage — the diagnosis by Goutam Challagalla, Mahwesh Khan and Fabrice Beaulieu is that companies scatter capital across dozens of experiments that never reach scale [5]. The following issue, written by Johannes Berndt and co-authors, adds that generative AI experimentation demands testing at the organizational level, not just at the tool level — and that too much change, warn Darrell Rigby and Zach First, traumatizes organizations [6]. The MIT Sloan Management Review's Winter issue closes the loop: Öykü Işık and Ankita Goswami show AI risk management lagging due to internal cultural and structural problems, while Chris K. Anderson and Fredrik Ødegaard document collusion lawsuits against pricing algorithm providers and the companies that use them — some already settled, others dismissed, others active [7]. This last data point deserves any board's attention: the first large-scale lawsuits over algorithms don't concern the technology failing, but the technology working too well — coordinating prices without human coordination. AI's legal risk is no longer hypothetical; it has a case number.

Together, these four patterns compose the picture no single source, alone, offers. Corporate demand for AI is real — companies are investing, piloting, reorganizing — but it's held back by three bottlenecks no amount of capital resolves: scattered pilots without strategic concentration [5], misalignment between innovation ambitions and operating models [6], and a risk culture that hasn't matured at the pace of the technology [7]. The report on the frontier of work is explicit about the timeline: this transformation, in the image of the Industrial Revolution and the internet era, will take decades to fulfill its promise, because it involves slow technological, social and economic change [8]. Against that clock, capital was raised with quarters-long maturity [1][4]. None of the sources contradicts another on the facts; it's the juxtaposition that exposes the mismatch — and it's the mismatch, not the technology, that the correction will punish.

There is an implicit international comparison in this material worth making explicit. Infrastructure is concentrated in the United States — the reopened plants, the grid queues, the compute commitments are American phenomena [3][4] — but the demand that must pay for it is global, and globally slow. A European company reading HBR and concentrating its pilots on a single area of advantage [5] contributes to American supply revenue at the pace of its own internal transformation, not at the pace of the debt calendar of whoever built the data center. Supply geography and demand geography don't coincide — and the AI lens shows that duration risk also has a territorial dimension: the balance sheets that issued the commitment are in one time zone; the organizations that must honor it are in another, in every sense of the word.

For a board, what this lens makes visible — and the financial lens hides — is the exact nature of exposure. The wrong question is "should we invest in AI before the bubble bursts?"; the right question is "what is the duration of the capital we are exposed to, directly or indirectly, and what is the actual maturity of our internal demand?" A company with twelve scattered pilots and no area of concentration isn't behind on AI — it is financing, with its experimentation budget, the illusion of revenue for whoever sells compute [5][1]. A company that concentrates, aligns its operating model and matures risk governance [6][7] converts the same budget into advantage — and positions itself to buy cheap capacity when the correction reprices infrastructure. In April, this outlet wrote that infrastructure speed and regime speed were in compression ("Sam Altman vs Mustafa Suleyman: The Compression Between Infrastructure and Regime"); what was then a qualitative tension now has a price: $2.9 trillion is the cost, measured by Sequoia, of two clocks that failed to sync [1].

BOX — Key concept: duration mismatch applied to AI

In finance, a duration mismatch occurs when an agent finances long-term assets with short-term liabilities — the classic mechanism of banking crises, where deposits redeemable on demand sustain thirty-year loans. Applied to AI, the concept describes the situation in which a single year's worth of AI infrastructure — data centers, chips, energy contracts — was financed against an expectation of $1.5 trillion in revenue that corporate adoption, for organizational rather than technological reasons, will only deliver over decades [1][8].

The mismatch's signature isn't found in asset prices but in the parties' behaviors. On the supply side, it manifests in revenue improvisation: OpenAI embracing the advertising its leadership had treated as a last resort, to cover commitments of hundreds of billions in compute [4]. On the financial side, it manifests in the credit and energy innovations the Economist denounces — instruments designed to relieve balance sheets and the grid, that is, to stretch the duration of the liability without acknowledging it [2]. On the demand side, it manifests in the convergent diagnosis of three management publications that don't cite each other: too many pilots [5], misaligned operating models [6], immature risk governance [7].

The critical distinction is between a thesis correction and a duration correction. A thesis correction invalidates the technology — that's what happened to much of e-commerce in 2000. A duration correction invalidates the capital structure but preserves the asset: the fiber optics overbuilt in the telecom bubble changed hands for pennies on the dollar and then sustained the entire internet. This month's data — nuclear plants reopened to serve real load [3], documented but slow corporate demand [5][6][7] — point to the second type. What's at risk isn't AI; it's whoever financed it on the wrong timeline.

For the decision-maker, the concept translates into a three-question discipline: first, what is the balance sheet's indirect exposure — as creditor, as a customer with long-term contracts, or as a competitor of leveraged companies — to capital structures with this mismatch; second, whether one's own portfolio of AI initiatives is generating defensible value or merely feeding supply's short-term revenue [5]; third, whether the organization will be positioned to acquire capacity — talent, compute, assets — when repricing arrives. Duration mismatch isn't avoided; it's arbitraged. And whoever understands it before the correction buys, afterward, what others built.

Implications for decision-makers

For the CEO of an AI-buying company, the correct reading of the $2.9 trillion gap quantified by David Cahn [1] is counterintuitive: the supply-side hole is, in part, a mirror of one's own pilot portfolio. When HBR's November-December issue recommends stopping the multiplication of scattered experiments and concentrating investment in an area where AI protects or creates hard-to-copy advantage [5], it is describing the aggregate behavior that keeps OpenAI's and Anthropic's combined annual recurring revenue at about $80 billion — real, growing, but nearly two orders of magnitude short of the revenue already-built infrastructure requires [1]. Every pilot without a path to production feeds supply's short-term revenue without building defensible value on the buyer's side. The first board-level decision is, therefore, not "how much to invest in AI"; it's where AI connects to the value chain with accumulable effect — and where it's merely rent paid to whoever financed an infrastructure requiring $1.5 trillion in revenue for every year of investment [1][5].

For the investor, the operating distinction is between exposure to the technology and exposure to the capital structure financing it. The Economist documents how credit and energy innovations designed to relieve balance sheets are shifting duration risk into increasingly illegible instruments [2]; the WSJ Weekend shows the other side of the same coin — three nuclear plants slated for closure now reopening, surging demand for gas plants, long grid connection queues [3]. A balance sheet may hold not a single AI stock and still be exposed: as a creditor of data center financing vehicles, as a counterparty to long-term energy contracts, as a shareholder in utilities whose customer base has concentrated around a handful of hyperscalers. The due-diligence question has changed: it's no longer "does this company benefit from AI?", it's "what maturity of revenue supports this company's liability?" [1][2].

For those leading transformation within organizations, the convergence between HBR and MIT SMR supplies the map of the bottleneck. HBR's January-February issue locates the failure in the misalignment between innovation ambitions and operating models, warning — in the words of Darrell Rigby and Zach First — that "too much change can traumatize your organization" [6]; MIT SMR, through Öykü Işık and Ankita Goswami, shows AI risk management lagging due to internal cultural and structural problems, not technical limitations [7]. The practical implication is that accelerating adoption isn't buying more compute — it's redesigning risk governance, the operating model, and the alignment between stack and strategy before scaling. Whoever reverses this order will produce exactly the pattern the three publications diagnose: high investment, intangible value [5][6][7].

There is also a talent implication rarely part of this conversation. Heidrick & Struggles' 2026 High End Independent Talent Report records 151% growth since 2021 in demand for interim C-suite leaders, with interim CFOs accounting for 51% of all interim leadership requests and requests for interim COOs up 250% year over year [9]. Sunny Ackerman writes that independent talent is "moving to the center of how critical work gets done" [9] — and the reading through this lens is precise: organizations are assembling flexible execution capacity for transformations whose duration no one can predict. In a mismatch of clocks, leadership elasticity is a risk hedge as relevant as capital elasticity. The board that treats AI as a technology project hires engineers; the one that treats it as a duration restructuring hires reconfiguration capacity [8][9].

Finally, the coldest implication: prepare the balance sheet to buy after repricing. If the correction that comes is one of duration and not of thesis — and this month's data point in that direction [3][5][6][7] — then there will be compute, energy contracts, talent and physical assets changing hands at a discount, as the fiber optics of 2001 did. Availability of liquidity at others' wrong moment is the oldest of competitive advantages.

The paradox: real demand as an alibi for the wrong capital

The tension this lens reveals is this: the more tangible physical demand becomes, the more legitimate the capital structure financing it appears — and the more dangerous it is. The reopened nuclear plants and the grid queues the WSJ Weekend documents [3] are the favorite argument of those dismissing the bubble hypothesis: electric load isn't speculation, it's physics. But the physics of demand says nothing about the timeline of revenue. A plant reopened to serve data centers whose monetization model is still under construction — OpenAI turning to advertising its own leadership had described as a last resort [4] — is real demand in the service of a liability with the wrong maturity. The concrete is real; the timeline is fiction.

The paradox's second floor is more uncomfortable: buyers' prudence, which protects them individually, is what worsens the gap collectively. When HBR says to concentrate instead of multiplying pilots [5], and when Rigby and First warn against too much change [6], they are giving organizationally correct advice — and every company that follows it stretches the interval between the capital invested and the revenue meant to service it. Buyers' micro-rationality produces financiers' macro-fragility. Neither party is wrong; it's the clocks that don't coincide [1][6][8].

And there is a third floor: the credit innovations the Economist denounces as air pumped into the bubble [2] are, seen up close, the system's most honest attempt to solve the real problem — stretching the duration of the liability to bring it closer to the duration of adoption. The paradox is that it does so without admitting it, packaging decades-long risk into instruments priced as if they were quarters-long. The remedy exists; it's just being administered in secret, which guarantees no one doses it correctly.

Looking ahead: three repricing scenarios

The most likely scenario, in light of this month's data, is a phased duration correction: the most leveraged compute commitments are restructured, physical assets — plants, data centers, energy contracts [3] — change hands at a discount, and the application layer keeps growing on top of repriced infrastructure, just as the internet grew on top of the cheap fiber of 2002. In this scenario, the $2.9 trillion gap [1] doesn't disappear; it transfers from the original financiers to buyers with patient balance sheets. The winners will be companies that, following HBR's playbook, concentrated adoption in one defensible area [5] and arrive at repricing with operating discipline and liquidity.

The second scenario is forced monetization: supply, unable to wait for buyers' organizational maturity, accelerates consumer revenue models — OpenAI's advertising is the first signal [4]. This path shortens the duration mismatch, but at the cost of turning frontier labs into media companies, with everything that implies for incentives, product and trust. It's the scenario in which capital wins and the thesis changes nature.

The third scenario — the least discussed — is successful slow absorption: buying organizations, supported by flexible leadership whose demand is already growing at double digits [9] and by matured risk governance [7], convert pilots into production at the pace the frontier-of-work report anticipates for transformations of this scale — decades, not quarters, in the image of the Industrial Revolution [8]. In this scenario, the technology fully delivers on its promise; it simply delivers it too late for the generation of capital that financed it. It's the scenario where everyone is right and someone still loses money.

In all three, one constant: the correction, when it comes, will no longer correct only on the Nasdaq. The convergence between the Economist and the WSJ Weekend established that the AI bubble has become a phenomenon of both physical and financial markets — grid, nuclear, gas, credit [2][3]. Repricing will be felt by balance sheets that never bought a single token.

Conclusion

In April, this outlet wrote that infrastructure speed and regime speed did not coincide — it was, then, a qualitative tension. July brought the number: $2.9 trillion is the price, in David Cahn's arithmetic, of that compression [1]. What this month's sources add, read together, is the anatomy of the gap: it doesn't live in the technology, which works; nor in demand, which exists and is already reopening nuclear plants [3]; it lives in the interval between capital raised with quarters-long maturity and an adoption that management publications themselves, without citing each other, describe as structurally slow [5][6][7][8].

Financial history knows this pattern well: short-term debt was issued against long-term revenue, and it was called vision. When the difference comes due, it will say little about artificial intelligence and much about financial intelligence. The AI bubble, if it bursts in 2026, won't burst from an excess of future — it will burst from a lack of present in the calendar of whoever paid for it.

Frequently asked questions

Is there really an AI bubble in 2026? It depends on what's being measured. David Cahn's (Sequoia) analysis quantifies a gap of $2.9 trillion between the nearly $3 trillion in cumulative revenue that already-built infrastructure requires and about $80 billion in combined annual recurring revenue from OpenAI and Anthropic [1]. Physical demand data — reopened nuclear plants, grid queues [3] — and the organizational diagnosis from HBR and MIT SMR [5][7] suggest the imbalance is one of capital duration, not technology validity.

What does Sequoia's $2.9 trillion gap mean? It measures the difference between the cumulative revenue that AI infrastructure built since ChatGPT requires over its useful life (nearly $3 trillion, by Cahn's count) and the revenue that infrastructure generates today [1]. It doesn't measure a lack of demand: it measures demand with the wrong maturity — buying companies are held back by scattered pilots, misaligned operating models and immature risk governance [5][6][7], so revenue arrives at the pace of decades while the liability comes due at the pace of quarters.

If the bubble bursts, does AI fail? Probably not. A duration correction invalidates the capital structure and preserves the asset — like the fiber optics overbuilt in 2000, which changed hands at a discount and later sustained the entire internet. The correction would punish whoever financed on the wrong timeline, not the technology [1][8].

How should companies prepare? Concentrate AI adoption in one defensible area instead of multiplying pilots [5], align operating model and risk governance before scaling [6][7], map indirect balance-sheet exposure to leveraged capital structures [2], and prepare liquidity and flexible leadership [9] to acquire capacity when repricing arrives.

Related reading: «Sam Altman vs Mustafa Suleyman: The Compression Between Infrastructure and Regime» · «The Invisible Infrastructure of Decision-Making»