In Eight Hours, $42 Billion Was Gone. Why Big Companies Don't Collapse Slowly, but All at Once

The article explains why large organizations don't fade away gradually but collapse abruptly, like a phase transition in physics — after a long period of apparent solidity, they are set off by a trigger that would not harm a healthy system. Drawing on cases from Silicon Valley Bank and Lehman Brothers to Enron, Wirecard, FTX and Nokia, it identifies three recurring mechanisms (internal rigidity, external interconnectedness and criticality) grounded in complex-systems theory. It also soberly assesses the limits of predicting collapses, from Altman's Z-score to Sornette's log-periodic models, and concludes that we understand the mechanisms but cannot reliably forecast them.

The collapse of large organizations rarely looks like a gradual wasting away. It resembles more a phase transition in physics—a long period of apparent calm during which hidden tension accumulates, followed by a breakdown over the course of a few days. Here is what we know about this pattern from complex systems theory, and what we know from the autopsy of specific failures from Enron to Silicon Valley Bank.

On the morning of March 9, 2023, Silicon Valley Bank received withdrawal requests of roughly $42 billion within eight hours—nearly a quarter of all its deposits. The sixteenth-largest bank in the United States, until then known more for the dull stability of its financing of California startups, was suddenly insolvent. The next morning the California regulator closed it and handed it over to FDIC receivership. That eight-hour figure comes from the review by the California DFPI of May 8, 2023; it is also confirmed by the material loss review issued in September 2023 by the Inspector General of the Federal Reserve System.

The day before, nothing visible had happened that would explain such a panic. SVB had merely announced a $1.8 billion loss from the sale of part of its bond portfolio and an intention to raise $2.25 billion in new capital. No fraud, no concealed billions. But several prominent venture funds advised their portfolio companies to pull their money out, and the recommendation shot through closed chats and the X network at a speed the old world of brick-and-mortar branches had never known. Patrick McHenry, chairman of the House Financial Services Committee, later spoke of "the first Twitter-fueled bank run." A study by economists from the spring of 2023, later published in the Journal of Financial Economics, showed that the intensity of conversation about the bank on social networks predicted hourly drops in its share price.

And this is precisely the heart of the matter. The fall of a large organization is usually neither a slow wasting away, where the brakes can be applied in time, nor the simple consequence of one large failure. It is an emergent phase transition: a macroscopic event that arises from the interactions of many small parts and erupts nonlinearly after the system has been approaching the edge for a long time, invisibly. We understand the mechanisms of such a transition fairly well today. But we cannot reliably predict it—and the part of the literature that claims otherwise is either overstated, or it is a retrospective narrative dressed up as prediction. I leave both possibilities open here and will return to them.

Surprisingly many famous collapses have the same shape. Lehman Brothers fell on September 15, 2008, after months in which officially "everything was holding." Enron declared bankruptcy on December 2, 2001, six weeks after it had still, on October 16, announced "merely" a one-off write-down. Germany's Wirecard, a member of the prestigious DAX index, collapsed on June 25, 2020, over a single week. The crypto exchange FTX filed for bankruptcy on November 11, 2022, four days after a run on its own token had begun. In all cases, the collapse was preceded by a long period of apparent solidity, and it was triggered by a stimulus that would have done nothing at all in a healthy system.

The word "emergence" must be tamed right at the outset, because in popular texts it is waved around like a magic wand. In the strong sense, as introduced by the physicist Philip Anderson in his 1972 essay More is Different, it means a property that cannot be derived from the simple sum of the parts. In the weak sense, it is often just an ornamental synonym for "complex." This text uses a middle position: collapse is a macro-property that arises from local interactions—from the actions of employees, creditors, depositors, counterparties—but it can be causally decomposed through mechanisms at the intermediate level. Three such mechanisms recur in collapses: internal rigidity, external interconnectedness, and criticality.

The theoretical backing for the third, most abstract property was supplied by physicists. Per Bak, Chao Tang, and Kurt Wiesenfeld described in 1987 in Physical Review Letters the sandpile model: if you pour grain after grain onto a pile, it brings itself into a critical state in which one more grain triggers an avalanche of arbitrary size, and the distribution of those avalanche sizes follows a power law. Large avalanches are therefore not anomalies, but a normal property of the system. Bak elaborated his idea of "self-organized criticality" in his 1996 book How Nature Works, applying it also to economics and market crashes. Caution is in order here, however, to which I will return: the sandpile model is for a company more of a suggestive metaphor than an exact description.

More useful in practice is the work of the ecologist Marten Scheffer and his colleagues. In 2009 they summarized in Nature the universal warning signals of an approaching critical transition: as a system nears the edge, it recovers ever more slowly from small fluctuations (a phenomenon called "critical slowing down"), the variance of its key variables grows, and their autocorrelation grows. These signals are solidly documented in lake ecosystems by a whole-ecosystem experiment by Stephen Carpenter (Science, 2011), in climate data (Dakos et al., PNAS, 2008), and, remarkably, even in the onset of clinical depression in humans (van de Leemput et al., PNAS, 2014). For companies, robust longitudinal evidence is still lacking, and I will therefore return to their predictive value with healthy skepticism.

Internal rigidity is the first of the recurring mechanisms, and sociology described it long before chaos theory. Michael Hannan and John Freeman, in a study in the American Sociological Review in 1984, inverted the common intuition: an organization's structural inertia is not a managerial weakness but the result of selection. The market rewards firms that are reliable and accountable—and reliability is bought at the price of fixed routines that are hard to change. Firms that adapt quickly and often lose credibility, and the population punishes them. Inertia is thus the price of survival in calm times and a lethal burden as soon as the environment changes abruptly.

Danny Miller gave this process a name that has held up to this day. In his 1990 book The Icarus Paradox he analyzed over a hundred corporations and showed that exceptional firms often destroy themselves through the very qualities by which they excelled. The meticulous manufacturer turns into an obsessive tinkerer, fine-tuning the details of a product no one wants anymore; the builder degenerates into an imperialist who has grown through acquisitions beyond the limit of manageability. The mechanism is "deadly inertia"—the strategy, structure, and culture that brought success are amplified through positive feedback until they harden. Dorothy Leonard-Barton named the same thing in the Strategic Management Journal in 1992 as "core rigidities": the same four layers that make up a firm's core capabilities have a flip side that blocks innovation.

The best-documented collapse of this type is Nokia. Timo Vuori and Quy Huy described in the Administrative Science Quarterly in 2016, on the basis of 76 in-depth interviews, how the Finnish company lost the battle for smartphones, even though its Symbian system controlled, according to Gartner data, roughly two-thirds of the global market in mid-2007. Top management feared the competition as well as the shareholders and had a reputation as people who do not tolerate bad news. Middle management feared top management. An emergent information blockage arose: no one passed the true state of the operating system's development upward, and the leadership made decisions on the basis of knowingly embellished reports. In 2013 Nokia sold its mobile business to Microsoft. It is fitting to admit a limit: this is a single qualitative study, however exceptionally careful, and its mechanism would need to be verified on further cases before we turn it into a law.

The second mechanism comes from outside, and it was formulated most sharply by the sociologist Charles Perrow. In his 1984 book Normal Accidents, written originally over an analysis of the Three Mile Island nuclear power plant accident, he argues that systems which are simultaneously interactively complex (their components affect one another in unexpected ways) and tightly coupled (processes run fast and cannot be isolated) produce catastrophes necessarily, not by chance. He called them "normal accidents." They arise when two or three small failures unexpectedly meet and tight coupling turns them into a cascade faster than anyone can grasp what is actually happening. Perrow extended the same logic beyond industry as well: in his 2007 book The Next Catastrophe he warned against concentrated, tightly interconnected critical infrastructures, and in the preface to the 2011 paperback edition and in later articles he applied the lens of normal accidents directly to the financial crisis of 2008.

That this was no speculation was shown by the crisis of 2008. The final report of the U.S. investigative commission FCIC from January 2011 reaches the harsh conclusion that the crisis was avoidable, and it describes precisely Perrow's cascade: millions of over-the-counter derivative contracts between systemically important institutions added uncertainty and escalated the panic. The insurer AIG received a federal commitment of over $182 billion precisely out of fear that its fall would trigger a chain loss across the global system. Physicists and network scientists quantified this after the crisis. Stefano Battiston and colleagues introduced in 2012 in Scientific Reports the DebtRank metric, which they applied to $1.2 trillion of emergency Fed loans from 2008 to 2010. They found that 22 institutions formed a densely interconnected graph in which even a small dispersed shock could trigger a systemic collapse. Their conclusion reframed the entire debate: it is not so much a matter of being "too big to fail" as of being "too central to fail."

Related to this is a counterintuitive result that every risk manager should read. Diversification, which reduces the risk of an individual institution, can increase the risk of the whole system. When all firms hold similarly "safe" portfolios, their fates become correlated and in a crisis they fall together. This was shown independently by Beale et al. in PNAS in 2011 and a year later by Battiston with Stiglitz and others in the Journal of Financial Stability. A Markowitz-optimal portfolio at the level of the individual can thus be a fuse at the level of the system.

There is, however, also contrary evidence, and honesty demands that it be cited. Karl Weick and Kathleen Sutcliffe describe in their book Managing the Unexpected the so-called high-reliability organizations—air traffic control, nuclear submarines, dialysis units—which, in an environment just as risky as Perrow assumes, achieve an order-of-magnitude lower accident rate. They manage it through culture: a preoccupation with failure, a reluctance to simplify things, deference of decision-making to whoever has the most expertise at the given moment. The historian Scott Sagan objects to this in The Limits of Safety from 1993 that these organizations were perhaps not safe, just lucky. The dispute is not settled, but its very existence refutes the harshest reading of Perrow: tight coupling is not destiny, cultural choices carry causal weight.

That company collapses are not independent random events is suggested by the shape of statistical distributions. Robert Axtell showed in 2001 in Science, on the complete population of roughly 5.5 million American tax entities, that firm sizes follow Zipf's law with an exponent close to one. The earlier notion of a lognormal distribution was an artifact of measuring only large publicly traded firms from the Compustat database. The group around Gene Stanley, in turn, documented in Nature as early as 1996 that the distribution of firms' annual growth rates has a Laplacian shape and that its variance declines with size according to a power law, across seven orders of magnitude. Such "scaling" laws are a typical imprint of systems near criticality.

It is necessary to apply the brakes here, however—and Axtell himself does so. From the fact that the data fit a power-law curve, no specific microeconomic explanation follows; the same curve is produced by a whole family of models, and Axtell himself wrote of the simplest one that it is "more a fable about firm growth than a credible explanation." Power laws are, moreover, statistically treacherous. Aaron Clauset, Cosma Shalizi, and Mark Newman showed in a review study in SIAM Review in 2009 that a number of published "power-law" distributions are in fact lognormal or stretched exponentials and the authors never properly tested them. Statistics thus supports the thesis that this is a systemic phenomenon, not a series of independent misfortunes—but it claims no more than it can bear.

Part of the systemic reading is also the question of when organizations die. Arthur Stinchcombe formulated already in 1965 the "liability of newness": young firms perish more often than old ones. Josef Brüderl and Rudolf Schüssler corrected this picture. In a 1990 study, built on the complete set of trade-license registrations and deregistrations in Munich and Upper Bavaria from 1980 to 1989, they showed that mortality does not decline monotonically but peaks somewhere between the first and the fifteenth year after founding, depending on how large an initial endowment the firm received as a dowry. As soon as the honeymoon of initial resources runs out, the most dangerous phase arrives.

If that emergent pattern had a reliable warning system, one could make a fortune on it. There have been plenty of attempts, and it is worth knowing where their ceiling lies. The oldest and most widely used is Altman's Z-score. Edward Altman published in 1968 in the Journal of Finance a discriminant function of five accounting ratios which, on a training sample of 66 American manufacturing firms from 1946 to 1965, determined bankruptcy one year in advance with an accuracy of around 95 percent. The catch is in that word "training": on data outside the original sample, accuracy drops, according to Altman's own 2000 review, to roughly 80 to 90 percent one year before the fall, and the model had to be repeatedly recalibrated since the 1960s for other types of firms and other countries. Above all, however, the Z-score reads accounting—and at Enron, Wirecard, and FTX the accounting was partly fiction. Lehman, moreover, cosmetically improved its balance sheet through Repo 105 transactions, so that indicators based on the statements signaled imminent danger only dimly.

More ambitious is the approach of the physicist Didier Sornette. Together with Anders Johansen and Jean-Philippe Bouchaud he described in 1996 the so-called log-periodic power law: before a crash, the index supposedly grows super-exponentially, and ever-faster oscillations are superimposed on it, from which the most probable time of the break can be estimated. Sornette later called these predictable extremes "dragon kings" and distinguished them from Taleb's "black swans"—they are not random draws from the tail of a distribution but products of a special endogenous mechanism, namely a bubble feeding on itself. The academic community is, however, divided. Hyong-shik Chang and James Feigenbaum found in a Bayesian analysis from 2006 no statistical evidence that the log-periodic model outperforms a simple random walk. David Brée and Nathan Joseph showed in 2013 that the model is over-parameterized and that different local optima give dramatically different estimates of the crash time. Stories of miraculous profits on specific crashes circulate, but the most-cited ones come secondhand and you will not find them in peer-reviewed form; as evidence, therefore, they do not hold up.

The reason prediction is so stubbornly weak is, after all, contained in the very nature of the thing. Jean Carlson and John Doyle showed at the turn of the millennium, most recently in PNAS in 2002, that power-law tails in engineered systems—networks, technologies, but also firms—do not arise from spontaneous criticality as in a sandpile, but from human optimization. They called it "highly optimized tolerance." Such a system is "robust yet fragile": it superbly withstands the failures it was designed for, and it fails catastrophically at those the designer did not foresee. The more you tune a system against known risks, the harder a blow you take from the unknown. We can predict only what we put into the model—and the phase transition comes precisely through the door that is not in the model.

Now comes the strongest objection to everything I claim here: "emergent collapse" is often just a narrative we assemble after the fall. Phil Rosenzweig, in his 2007 book The Halo Effect, showed how stories about the rise and fall of firms are saturated with knowledge of the outcome—the same behavior is interpreted as visionary when the firm succeeded, and as hubris when it fell. A control group is missing: how many firms had the same "warning signals" and did not collapse? Without it, talk of prediction has no support, only survivorship bias. Added to this is the epistemic critique of transferring physics models onto society—Philip Mirowski developed it already in More Heat Than Light in 1989, and a group of economists around Mauro Gallegati summarized it in Physica A in 2006 into a warning against "worrying trends in econophysics." People, unlike grains of sand, have preferences and react reflexively to predictions, thereby breaking them.

And there is the hardest objection: many famous collapses were simply fraud and governance failure, not emergence. Wirecard invented €1.9 billion in Philippine accounts that did not exist, while the auditor EY saw nothing for over ten years and the German supervisor BaFin investigated, instead of the bank, the Financial Times journalists who had been pointing out the fraud since 2015. FTX commingled client money with its trading firm Alameda, and its new crisis chief John Ray III, the man who wound down Enron, wrote that in his entire career he had never seen such a complete failure of controls and such a total absence of trustworthy financial information. If emergence explains this too, it explains nothing.

This objection is correct and must not be evaded—which is why I let it stand. The answer is not that fraud is not fraud, but that the emergent framework explains something else: not why people lied, but why they got away with it for so long. An audit that fails, a regulator captured by the industry, a culture in which deviation is gradually normalized until the exception becomes the standard—all of these are systemic, emergent properties, which the sociologist Diane Vaughan described in her 1996 analysis of the Challenger shuttle disaster as the "normalization of deviance." Enron, after all, shows that the boundary between fraud and complexity is porous: its roughly three thousand special-purpose entities were so intertwined with the price of its own stock that when the stock fell, so did the "hedging" Raptor structures, which held that very falling stock as their only collateral. The board's investigative report of February 2002 described it soberly: these were entities in which only Enron itself had a real economic stake. Fraud and positive feedback cannot be separated here.

What follows from this for someone who runs, finances, or regulates a company is less flashy than they would wish. Stop looking for one trigger and one indicator that will predict the crash. Instead of performance metrics, track fragility metrics: concentration of liabilities—at SVB roughly 94 percent of deposits were uninsured, a textbook red flag; maturity mismatch between what you owe short-term and what gets repaid to you late; hidden leverage in derivatives and off-balance-sheet structures; and yes, also rising variance and autocorrelation of key series according to Scheffer. Build redundancy where tight coupling has no strategic value. And when a firm holds an asset whose price hangs on its own balance sheet—whether it is the FTT token or stock hedging itself—treat it as a top-level alarm: it is precisely the loop that can turn a slow problem into an eight-hour fall.

Silicon Valley Bank disclosed the loss figure and the capital-raising plan on Wednesday evening. By Friday morning it no longer existed. In between, no slow decline elapsed, only the flipping of a system from one state into another, as soon as enough depositors believed the others would do the same. The question worth asking after reading this is not which firm is loss-making today. It is: which one looks solid precisely because it is tuned to a world that is about to change the rules?

Citations are given in a form that allows them to be traced in the primary source. Where claims rest on a secondary source or remain unverified, this is explicitly stated.

Complex systems theory and phase transitions

Organizational theory and structural inertia

Tight coupling, normal accidents, high reliability

Network cascades and systemic risk

Statistical regularities

Prediction and its limits

Antifragility and critique of the narrative

Case studies — primary and institutional sources

Transparency of authorship:

The concept, structure, and editorial line of the article are the work of the author, who prepared the content outline, established the key theses, and directed the entire writing process. Generative AI (Claude, Anthropic) was used as a tool for research, locating primary sources, and the stylistic elaboration of the author's content outline.

The author continuously edited the outputs, verified the key findings, and approved the final wording. No part of the text was published without human review. All factual data were verified against the publicly available sources cited in the text.

This procedure complies with the requirements of Article 50 of EU Regulation 2024/1689 (AI Act) on the transparency of AI-generated content. #poweredByAI