The Fault Beneath Silicon Valley

28 Aug, 2026

Silicon Valley learned to live with catastrophic risk by engineering around earthquakes and learning from failure, but AI raises a harder question because some failures may leave no chance to try again.

H. D. Chadwick, ruins near Post and Grant Avenue after the 1906 San Francisco earthquake, April 1906. National Archives and Records Administration, NARA 524396. Public domain.

There is a 72% probability that an earthquake of magnitude 6.7 or greater will strike somewhere in the San Francisco Bay region before 2043, according to the US Geological Survey. OpenAI is headquartered there. Anthropic is headquartered there, and has been expanding its San Francisco offices rather than diversifying away from them. Venture capitalists keep writing checks, engineers keep shipping code, and people keep paying extraordinary sums for houses sitting on top of some of the most famous faults on earth. Nobody considers that irrational. Californians learned a long time ago how to live beside a catastrophe they cannot prevent. I am from there, so I learned it early and without noticing I was learning anything. I have started to wonder whether that habit of mind followed Silicon Valley into artificial intelligence.

The forecast nobody flees

The numbers are not vague. The 2014 Working Group on California Earthquake Probabilities, whose findings USGS published as an outlook for the region from 2014 through 2043, put the chance of at least one magnitude 6.0 or greater event at 98%, magnitude 6.7 or greater at 72%, magnitude 7.0 or greater at 51%, and magnitude 7.5 or greater at 20%. The Hayward and Rodgers Creek fault system carries 33% on its own for a magnitude 6.7 or greater rupture, ahead of the San Andreas at 22%. Those figures cover 32 smaller faults added to the 5 major systems, plus an allowance for faults nobody has mapped.

The region has also modeled what the event would do. The USGS HayWired scenario takes a magnitude 7.0 mainshock on the Hayward Fault and works out the consequences in detail: roughly 800 deaths, 18,000 nonfatal injuries, about 450 large fires, water service lost for an average of 6 weeks and in places 6 months, and more than $82 billion in property and business interruption losses. Some downtown high-rises would be unusable for the better part of a year.

None of this is secret, contested or ignored. It is published, taught, mapped and built into code. San Francisco signed its mandatory soft-story retrofit ordinance into law on April 18, 2013, the anniversary of the 1906 earthquake, and screened thousands of wood-frame buildings for the weak ground floors that collapse first. Early warning systems exist. Emergency plans exist. Insurance exists. And then everyone goes back to work.

That response is correct, and it is worth being precise about why. Earthquake risk is exogenous. There is no negotiating with the Pacific Plate, no regulation that slows the accumulation of strain along the Hayward Fault, no version of the problem in which restraint helps. The only available moves are engineering, preparation and acceptance, so the Bay Area executes all 3 and continues living. Abandoning northern California would be a far larger catastrophe than the one being avoided.

Artificial intelligence is a different kind of problem, and the difference is the whole essay.

What the researchers say about their own field

In October 2023, Katja Grace and colleagues surveyed 2,778 researchers who had published at top AI venues, in what became the largest survey of its kind and was published in the Journal of Artificial Intelligence Research in 2025. Between 38% and 51% of respondents assigned at least a 10% probability to advanced AI producing outcomes as bad as human extinction, the range reflecting different phrasings of the question. Even among the 68% who thought good outcomes more likely than bad ones, 48% still gave at least a 5% chance to extremely bad ones. Majorities wanted more priority given to AI safety research.

Treat those numbers with the care they deserve. The response rate was 15%, which means the sample selected itself, and Melanie Mitchell has made the reasonable objection about earlier rounds of this survey that a headline built on a self-selected minority answering an unbounded question about an unspecified timeframe cannot carry much weight. The honest summary is that there is no consensus that AI will end badly, and there is a substantial professional minority that considers it possible.

The same profession is considerably less impressed by the capabilities that would be required. The AAAI 2025 Presidential Panel surveyed 475 researchers, and 76% judged it unlikely or very unlikely that scaling up large language models will produce artificial general intelligence at all. Those 2 findings are not contradictory, since a person can think catastrophic outcomes are possible and also think the present technical path does not lead there. Read together they describe a field that takes the risk seriously and the timeline much less so, which is worth holding onto whenever a probability of extinction gets quoted without one attached.

Professional forecasters land lower still, and the gap between them and the researchers is the most interesting number in this essay. The Forecasting Research Institute ran an Existential Risk Persuasion Tournament putting 80 domain experts and 89 superforecasters on these questions with money at stake. Median probability of AI-caused extinction by 2100: 3% from the domain experts, 0.38% from the superforecasters. For catastrophe short of extinction, 12% against 2.13%. The Metaculus community currently sits at 2% for human extinction before 2100 across roughly 1,700 forecasters. Set that beside the AI researchers, 38 to 51% of whom assigned 10% or more, and the serious estimates span more than an order of magnitude. The tournament participants argued for months under incentives designed explicitly to make them converge, and they did not move, which the organizers called the most puzzling result they got. The disagreement is the finding.

What interests me is not the number. It is that the people producing these estimates are, in large numbers, the same people building the systems. They hold 2 ideas at once: the consequences might be catastrophic, and the work should continue.

In most professions that combination would sound like a contradiction. In the culture that produced this one, it may sound ordinary.

Two educations in risk

I want to state the hypothesis carefully, because the interesting version is weaker than the tempting version. No evidence establishes that living near earthquake faults causes anyone to tolerate catastrophic technological risk, and I am not claiming it does. The claim is cultural rather than causal. Silicon Valley grew up in a place where intelligent people routinely acknowledge a catastrophe, cannot say when it will arrive, mitigate what can be mitigated, and go on building. That habit long predates artificial intelligence.

The second education came from money. Venture capital is a machine for converting frequent failure into occasional enormous return, and its practitioners say so openly. Ilya Strebulaev and Alex Dang, writing from Stanford's business school on what they call the venture mindset, put it as home runs matter and strikeouts do not, and describe an investor discipline of rejecting more than 100 deals for every one funded. The default setting in most of the business world, they note, is caution and consensus and a low tolerance for bad bets. The Valley inverted that, and the inversion worked. Failure became experience, portfolios absorbed losses that would have ended a conventional firm, and the region produced several decades of companies that nobody would have funded under ordinary rules.

Both educations are rational. Neither is reckless. That is exactly what makes the transfer problem hard, because a mental model does not announce itself when it moves into a domain it was not built for.

Three kinds of risk

It helps to separate them.

Earthquake risk is catastrophic, uncertain in timing, largely unavoidable and external to human decisions. Mitigation is the right response. You strengthen the building because you cannot stop the ground.

Venture risk involves frequent failure, deliberately accepted, entirely human-created, usually bounded and recoverable, and diversifiable across many independent bets. Iteration is the right response. Nine companies can fail if the tenth is Google, and each failure produces information that improves the next attempt.

Catastrophic AI risk is potentially severe, uncertain, human-created, potentially correlated across an entire society, possibly irreversible and not obviously diversifiable. It borrows the uncertainty of the first category and the human authorship of the second while sharing the consoling properties of neither.

The trouble appears when models built for the first 2 migrate into the third. Earthquake logic says prepare and keep living. Startup logic says experiment and learn. Both arrive at the same instruction, and neither was designed for a failure that ends the sequence.

The other technology region, and the wider list

Silicon Valley is not the only American technology center sitting on a modeled catastrophe. The Cascadia Subduction Zone last ruptured on January 26, 1700, at an estimated magnitude between 8.7 and 9.2, on a recurrence interval near 500 years, and USGS puts the chance of a full-margin magnitude 9 event in the next 50 years at 10% to 15%, affecting more than 7 million people. Be accurate about the geography, though, because the tempting version is wrong. Mount Rainier's lahars run south and west down the Puyallup, Nisqually and White River valleys toward Orting, Auburn and Tacoma, and Seattle's own emergency managers assess a lahar reaching the city as possible but extremely unlikely, with no direct inundation zones inside city limits. Redmond, where Microsoft sits, is not in a Rainier drainage at all. The shared exposure is Cascadia, and 15% over 50 years is a good deal softer than the Bay Area's 72%. What the 2 regions have in common is not the hazard but the posture: a published, unpreventable catastrophe, and an industry that keeps hiring into it.

Widen the frame past regional hazards and the more useful comparison appears. Toby Ord's The Precipice offers the best-known attempt to put humanity's catastrophic risks into a single table, assigning each a rough chance of causing an existential catastrophe within the next 100 years.

Risk	Chance within 100 years
Stellar explosion	1 in 1,000,000,000
Asteroid or comet impact	1 in 1,000,000
Supervolcanic eruption	1 in 10,000
All natural risk	~1 in 10,000
Nuclear war	1 in 1,000
Climate change	1 in 1,000
Other environmental damage	1 in 1,000
Naturally arising pandemics	1 in 10,000
Other anthropogenic risks	1 in 50
Engineered pandemics	1 in 30
Unforeseen anthropogenic risks	1 in 30
Unaligned artificial intelligence	1 in 10
Total	~1 in 6

Ord is explicit that these are subjective credences rather than measurements, rounded to the nearest order of magnitude, and his 2024 revisit reports movement among the constituents, with nuclear up and climate down and pandemics and AI mixed, without a clear shift in the total. Read them as one careful person's considered estimates rather than a forecast.

Read structurally, the table says 2 useful things. The first is Ord's own summary: humanity faces roughly 1,000 times more anthropogenic risk this century than natural risk. Earthquakes do not appear at all, because no earthquake can end a species, and once naturally arising pandemics are counted where Ord counts them, on the human-influenced side of the ledger, the natural block is purely astronomical and geological. The Big One is regional rather than civilizational, which makes the Bay Area's posture more defensible rather than less. Even Yellowstone, the hazard people reach for when they want something apocalyptic, comes in at USGS odds of 1 in 730,000 per year, roughly the same as a 1 kilometer asteroid strike.

The second is what the table does not contain. There is no line for alien invasion, and there could not be, because every entry needs either a physical base rate or a mechanism specific enough to argue about. Asteroids have a cratering record, supervolcanoes a stratigraphy, nuclear war an arsenal that can be counted. Even a Carrington-class solar storm has a historical event behind it and a contested estimate around it.

Unaligned AI has none of that, and it sits at the top of the list anyway, above all natural risk by 3 orders of magnitude and above every other human-made risk. It also sits roughly 25 times above what the superforecasters in the persuasion tournament would put there, which is worth remembering while looking at a tidy table. The number is not a measurement. It is a belief about a system that does not exist yet, produced by people reasoning about capabilities they intend to build. Which is the whole difference restated in a different register. Every other entry on that list is either a fact about the universe or an inheritance from decisions already taken. The one at the top is a live choice, being made now, by a small number of organizations, most of them within an hour's drive of a fault line.

The recovery assumption

Everything in the Valley's technical culture assumes another iteration. A startup fails and its founders start another. Software crashes and engineers patch it. A launch goes badly and the company rolls it back. A model hallucinates and gets retrained or replaced. A vulnerability appears and someone ships a fix. A fund loses money on most of its investments because a small number of them pay for everything.

That structure is an extraordinary engine, and it rests on a single unstated premise: there will be a next time. The premise is so reliable in ordinary technology work that it has become invisible, which is precisely the condition under which assumptions do damage.

Catastrophic risk is the case where the premise fails. A nuclear exchange does not permit a retrospective. A biological catastrophe cannot be rolled back to the previous version. An uncontrolled AI system, if such a scenario turns out to be possible at all, does not necessarily grant its engineers a second deployment. Build, observe, learn, fix is a superb method as long as failure stays inside the learning loop. Silicon Valley's genius is its conviction that failure is information. Catastrophic AI risk raises the possibility of a failure from which nobody remains to learn.

Catastrophe is survivable, until it is not

The Bay Area's history with disaster is not abstract. On April 18, 1906, an earthquake the USGS catalogs at magnitude 7.9 ruptured roughly 300 miles of the San Andreas Fault. The shaking was severe and the fires that followed were worse, burning for 3 days across 4.7 square miles and destroying more than 28,000 buildings. Estimates of the dead run past 3,000, and about 225,000 people out of a population near 400,000 were left homeless.

The city rebuilt. It happened again on October 17, 1989, when the magnitude 6.9 Loma Prieta earthquake injured more than 3,700 people and displaced 12,000, with direct damage around $6.8 billion. The region rebuilt a second time, and the retrofits that followed are part of why the next one will be less lethal than it would otherwise have been.

The lesson taught by that history is true and useful: catastrophe is survivable. San Francisco survived 1906. The Bay Area survived 1989. Buildings improve, bridges improve, warning systems improve, and disaster becomes part of institutional memory and eventually part of regional identity. There is even a proper noun for the one still coming. People talk about the Big One in the way you can only talk about something you have decided to live with, which is to say casually, with a joke attached and an earthquake kit in the closet.

I grew up with that, and I was there for the second one. Loma Prieta hit at 5:04 on an October afternoon with the third game of the World Series about to start at Candlestick, which is why a national television audience watched it arrive live. The recovery is the part worth reporting. The Bay Bridge lost a span, the Cypress Structure came down on the cars beneath it, 63 people were dead, and the World Series resumed at Candlestick 10 days later. That is not denial, and I am not describing a region in denial. It is what living somewhere requires, because you cannot hold a 72% probability in your foreground for 30 years and still get anything done, so the mind does the sensible thing and puts it away.

That is exactly the move I now watch people make about AI, and it is why I distrust my own comfort with it. The reflex is native to me. Somebody says a number that should stop a conversation, and the conversation continues, and nothing in me objects.

The distinction that matters is between catastrophe and irreversible catastrophe. The Valley has deep, hard-won, genuinely admirable experience with the first. Nobody has experience with the second.

What a probability does to a catastrophe

Numbers are how modern institutions make catastrophe thinkable, and they do something peculiar in the process.

A 72% chance of a damaging earthquake before 2043 is not remote, but distributed across 30 years it does not feel like an emergency, and nobody wakes up treating it as a 72% chance before dinner. Paul Slovic's foundational work established that people respond to hazards according to familiarity, controllability and novelty rather than expected value, which is why a familiar large hazard can feel smaller than an unfamiliar small one.

There is a directly relevant experiment. Kazuya Nakayachi, Branden Johnson and Kazuki Koketsu surveyed 750 San Francisco Bay Area residents in work published in Risk Analysis in 2018, varying both the stated probability of a major earthquake, at 20%, 70% and an implicit 100%, and the time horizon, at 10 or 30 years, while also varying whether experts explicitly acknowledged scientific uncertainty. Acknowledging uncertainty made the experts seem more honest. It did not change judged risk, preparedness intentions or support for mitigation policy, and neither did the probability itself.

That finding is not evidence that Bay Area residents ignore earthquakes. It is evidence of something subtler, which is that knowing a probability does not automatically produce behavior proportional to it. Preparedness research has long identified normalization bias, optimistic bias and fatalism among the beliefs that mediate between knowing a number and doing anything about it.

Apply the same observation to a professional community and you get the phenomenon behind the shorthand. Someone can say their P(doom) is 5% and then go back to improving the technology, and those 2 acts are psychologically compatible even when they look contradictory from outside. The remarkable thing about P(doom) as a term is not that people disagree about the value. It is that the probability of civilizational catastrophe has become an ordinary professional quantity, small talk at a conference, a number you can put in a slide.

Quantification is valuable. It creates discipline, permits comparison and forces people to say what they believe. It also creates distance. Human extinction is emotionally unmanageable; P(doom) = 0.05 is a variable.

Markets hit a harder version of the same wall. A prediction market is a good instrument because being wrong costs money and being right pays, and that mechanism switches off precisely at the end of the scale that matters here. A contract that settles only in a world with nobody left to collect is worth nothing to buy and free to sell, so betting against catastrophe is underpriced by construction and any market touching extinction reads low for structural rather than evidential reasons. The instrument goes silent on exactly the outcome we would most want it to price. That is the recovery assumption again, wearing a different institution: the Valley's iteration loop needs a next round to learn from, and the market needs a settlement date somebody survives to collect on.

I should say plainly what this parallel does not license. The 72% earthquake figure and any P(doom) estimate are not comparable quantities. One is a frequentist forecast built from geology, paleoseismic records and rupture models over a defined 30-year window. The other is an elicited subjective credence over an unspecified horizon with no track record to calibrate against. The comparison here is between how communities behave around a number, not between the numbers.

Engineering the danger downward

California's earthquake culture does not express itself mainly as denial. It expresses itself as engineering. Retrofit the soft story, strengthen the bridge, build flexible pipelines, install early warning, write the evacuation plan, improve the next code cycle. USGS frames its own hazard work in exactly those terms, on the reasoning that understanding a hazard is what allows a community to reduce injuries, damage and disruption.

Frontier AI labs run comparable logic. Build more capable systems, measure the dangerous behaviors, construct evaluations, improve alignment, add safeguards, monitor deployment, invest in interpretability, engineer the danger downward. That is a serious program staffed by serious people, and it may well be the correct response. A technologically sophisticated society should absolutely try to engineer safer AI, and the institutions most publicly committed to that engineering are themselves concentrated in the Bay Area, which is the part of the puzzle that makes it interesting rather than damning.

The asymmetry is philosophical rather than technical. Earthquake engineering mitigates a hazard that human beings did not create and cannot increase. AI safety engineering typically accompanies the continued creation of the hazard it is mitigating, and does so inside organizations whose capability research and safety research advance together. Studying the Hayward Fault does not make it stronger. Studying frontier models generally does make them more capable. That single difference is where the analogy stops carrying weight, and it is worth stating rather than smoothing over, because an analogy is most dangerous at the point where it almost still holds.

Bounded losses and unbounded ones

The venture half of the culture has an equivalent seam. A portfolio works because losses are bounded at the capital invested and independent enough to diversify. Both properties are doing real work in that sentence, and neither obviously survives the move to systemic technological risk. Civilization has no second portfolio, no hedge and no uncorrelated position. If an AI failure were genuinely severe and widely correlated, the downside would not stop at the equity of the company that caused it.

That is a familiar governance problem rather than an exotic one. Banks can be individually prudent and collectively dangerous, which is why macroprudential regulation sits on top of firm-level supervision. A factory can be profitable while externalizing its pollution, which is why environmental regulation sits on top of accounting. The insight in both cases was that private incentives and systemic consequences come apart, and that no individual actor, however well intentioned, can price a risk that exists only in aggregate.

I keep arriving at this seam from different directions. I have argued that context engineering is governance at inference time, that the SEC's silence on agentic trading leaves deployment responsibility unassigned, that governance arrived too late for agentic AI, and in the monoculture pieces that correlated judgment turns individual decisions into systemic ones. Risk culture sits upstream of all of them. The question is not whether the firms building frontier systems understand the risk, because they demonstrably do and say so publicly. The question is whether a society should rely exclusively on organizations under competitive pressure to decide how much systemic risk is acceptable on its behalf.

The case for building it here

The strongest objection to everything above is that Silicon Valley's risk tolerance is a reason to want AI built there rather than a reason to worry.

An excessively precautionary culture has its own catastrophic failure modes, most of them invisible, consisting of the medicine not developed and the discovery not made. A culture capable of acting before certainty arrives is what converts possibilities into products, and waiting for proof that a technology will work means surrendering the chance to build it. The people closest to these systems also understand their failure modes better than most regulators, and the same experimental culture that worries me may be what eventually solves alignment.

Earthquake engineering supports that reading. San Francisco did not become safer by leaving. It became safer through better structures, better science and better preparation, all of it produced by people who stayed. Retrofit rather than retreat is a defensible model for a society facing a hazard it cannot wish away.

There is also a serious scholarly position holding that the entire frame of this essay is wrong. Arvind Narayanan and Sayash Kapoor argued in April 2025, in a long essay for the Knight First Amendment Institute, that AI should be understood as normal technology rather than as a species of existential hazard. Their case does not rest on optimism. It rests on the gap between what a method can do and how fast it actually reaches consequential use, since invention, application and adoption run on different clocks, and diffusion into safety-critical domains has historically taken decades. They treat catastrophic misalignment as a speculative risk built on dubious assumptions about how technology gets deployed, note that a system would have to demonstrate reliability in minor contexts long before anyone handed it major ones, and point out that poorly controlled AI will usually be too error prone to make business sense. Their conclusion is that keeping AI under human control does not require drastic intervention or a technical breakthrough. If they are right, the anxiety this essay describes is a category error, and the correct response to the Valley's risk culture is roughly none at all.

Grant all of that fully, and one difference still refuses to dissolve. Earthquakes do not get stronger because engineers study them.

What I am actually claiming

Competition probably explains more of this than psychology or geography do. A researcher who assigns real probability to catastrophic outcomes can still conclude that unilateral restraint accomplishes nothing, because the systems will be built by someone, and the AAAI panel describes an environment in which companies and countries compete fiercely and globally to lead the race. So the earthquake hypothesis should stay one layer among several rather than a master explanation. Geography, entrepreneurial culture, venture economics, engineering mentality and competitive dynamics all point toward action under uncertainty, which is precisely why the resulting disposition is so hard to examine from inside.

I cannot prove that earthquakes shaped how Silicon Valley approaches artificial intelligence, and I would distrust anyone who said they could. What strikes me is the similarity of the 2 conversations. In both, well-informed people acknowledge a catastrophic possibility, attach a probability to it, argue about the number, design mitigations and continue with ordinary life, having concluded that acknowledging a risk is not the same as being ordered to stop. For earthquakes that conclusion is correct and there is no alternative. For artificial intelligence it is a choice being made, and it is worth noticing that it is a choice.

A fair reader will have noticed that I quoted estimates spanning more than an order of magnitude and then carried on as though the worry survived all of them. It does, and the reason matters more than any of the numbers, because the essay was never an argument about the value. If the superforecasters are right at 0.38%, the Valley's composure is vindicated and I have described a culture handling a modest risk sensibly. If Ord is nearer the mark at 1 in 10, that same composure is among the most consequential habits of mind in the world. What does not change anywhere across that range is the mechanism: a probability gets named, mitigation gets designed, and the naming does not alter the trajectory. A culture that would behave identically at 0.38% and at 10% is not responding to the number. It is running a reflex, and that is worth knowing whichever end of the range turns out to be right.

No single laboratory can slow this down alone, and a firm that tried would most likely trade its own influence for nothing. Grant that entirely and the conclusion does not follow, because the fact that no individual can make a choice is the classic signature of a choice that has to be made collectively. That is what capital requirements are, and emissions rules, and every arms control treaty ever ratified by governments who each preferred the weapons to unilateral disarmament. A risk nobody can price alone is the standard argument for pricing it together, and it is a peculiar move to treat the same structure as proof that nothing can be decided at all.

California learned to live with earthquakes because it could not stop them, and Silicon Valley learned to tolerate failure because failed companies can be replaced. Both habits helped produce one of the most inventive cultures in human history, and neither is a mistake. If the people building this technology genuinely believe there is some chance of an irreversible outcome, mitigation cannot function simply as another reason to keep moving.

The question is not whether Silicon Valley understands risk. It understands risk unusually well, better than most industries and considerably better than most regulators. The question is whether it has drawn the wrong lesson from having survived it. The San Andreas Fault will move whether anyone wants it to or not. Artificial intelligence is different. We are the ones moving the plates.

My blog