The Real Reason Silicon Valley 'Moved Fast and Broke Things' (It Wasn't a Philosophy)

The mythology goes like this: Silicon Valley succeeded because it had a culture of boldness. Move fast. Accept failure. Disrupt incumbents. The philosophy of velocity and irreverence toward existing institutions produced the companies that define the modern economy.

This is almost completely backwards. Silicon Valley didn’t succeed because of a philosophy. It succeeded because of three specific structural conditions that made rapid, unconstrained deployment the rational strategy for companies operating in those conditions. Those conditions are gone. The strategy that followed from them is now wrong for the environment AI companies are actually operating in. And every AI startup that still performs the “move fast” ethos — as if it were a timeless truth rather than a context-specific tactic — is confusing the costume for the body wearing it.

Let me explain what was actually happening underneath the mythology.

The first condition was a regulatory vacuum of a specific kind.

From roughly 1995 through 2012, the internet existed in a legal space that had almost no settled precedent and several explicit regulatory protections for digital platforms. Section 230 of the Communications Decency Act, passed in 1996, granted online platforms immunity from liability for user-generated content. This was not a minor carve-out — it was an enormous structural subsidy to platform business models. It meant Facebook, YouTube, Twitter, and every other major platform could operate at scale without the legal exposure that would have made their business models impossible for a newspaper or a broadcast network operating under established liability frameworks.

The absence of data protection law meant companies could collect anything they wanted from users without consent frameworks, retention limits, or audit requirements. The absence of competition enforcement in digital markets meant winner-take-all dynamics could play out without antitrust interference — a company could acquire competitors and achieve market dominance in ways that would have triggered regulatory response in older industries. The Federal Trade Commission spent the 2000s largely catching up to problems that had already happened.

In this environment, moving fast before competitors could react was rational. There were no legal speed bumps that slowed you down relative to incumbents. The regulatory risk of deploying quickly was minimal, because the regulatory structures simply weren’t there. The only question was whether you could capture network effects before someone else did.

The second condition was cheap capital with unusually patient time horizons.

The venture capital structure of the late 1990s and especially the 2010s produced an unusual situation: enormous amounts of money available at low cost, with fund structures that allowed patient holding periods (10-year fund cycles) and investors who had learned — from Amazon, Google, and Facebook — that the biggest returns came from letting winners grow through extended periods of unprofitability. Zero interest rate policy, which persisted from 2008 to 2022, made patient capital even cheaper by lowering the opportunity cost of not receiving returns today.

The result was that startups could operate at losses for years — sometimes a decade — while building user bases, network effects, and market positions that would be worth something eventually. Uber lost billions every year for years and was treated as a success story because its gross booking growth was impressive. WeWork was valued at $47 billion by SoftBank despite losing $1.9 billion in 2018. These numbers look insane in retrospect, but they reflected a coherent theory: that the companies capturing user attention and network effects in winner-take-all markets were worth subsidizing through enormous losses because the eventual returns in those markets would be proportionally enormous.

That theory worked for a handful of companies. It didn’t work for most. But the capital was there to run the experiment, at minimal cost, for a long time.

Interest rates changed everything. After March 2022, cheap capital disappeared rapidly. The 10-year Treasury yield moved from near zero to 4.5% within 18 months. Suddenly, companies that were burning $50 million a quarter needed a path to profitability that wasn’t theoretical. The 2022-2024 tech layoffs — more than 200,000 jobs cut across major tech companies over those two years — were the direct consequence of the capital conditions changing and the “grow at all costs” strategy suddenly being irrational.

The third condition was that the markets being disrupted were not high-stakes by existing regulatory standards, and the failure modes were visible and reversible.

Social networks, search, e-commerce, ride-sharing — these are not industries where the government has a strong pre-existing regulatory structure built around preventing serious harm. You can argue that they caused harms; there’s substantial evidence they did. But the regulatory system in 2005 wasn’t treating social media the way it treated aviation, pharmaceuticals, or nuclear power, because those industries had histories of catastrophic, visible, attributable failures that created political pressure for regulation. Consumer internet companies were given the benefit of the doubt by default.

This mattered because it meant the downside risk of moving fast was borne primarily by users, not by the companies. If Facebook’s News Feed algorithm turned out to have terrible effects on mental health — and there is substantial evidence it did — the legal liability exposure in 2012 was minimal. The company could run the experiment and, if it went badly, settle some civil suits, update some policies, and continue operating. The externalization of risk is the key. You got the upside of speed and you didn’t pay for the downside.

Now look at AI and tell me those conditions apply.

AI systems are being deployed in healthcare, financial services, legal services, and criminal justice — all sectors with extensive existing regulation and substantial potential liability. The EU’s AI Act, which entered into force in 2024, creates a tiered liability structure that makes high-risk AI deployments subject to significant requirements before they’re allowed into the market. In the US, state-level regulation of AI in employment, credit, and healthcare is moving faster than federal frameworks. Litigation is already happening — over AI-generated defamatory content, over AI hiring tools that failed disparate impact tests, over AI medical recommendations that conflicted with clinical standards.

Capital is expensive. The compute costs for frontier AI training runs are enormous: GPT-4 cost an estimated $100 million to train, and costs have escalated since. The companies spending that money are not operating in the patient-loss-is-fine VC environment of 2015. They are under intensive pressure to show revenue, because the capital markets funding them have gotten expensive and their investors have alternative uses for money that now have meaningful yield.

The failure modes are not social-media-style visible and slow. An AI medical diagnostic tool that misses diagnoses at higher rates for certain patient populations causes harm that’s diffuse, hard to attribute, and doesn’t show up as bad press until years later. An AI credit scoring system that discriminates creates class action exposure without generating a visible incident. An AI content moderation system that’s manipulated by adversarial actors causes political harm that’s genuinely hard to price. The asymmetry between upside and downside risk, which made moving fast rational in the consumer internet era, is absent or reversed in the high-stakes domains AI is now being sold into.

So why does the “move fast” mythology persist in the AI era?

Part of the answer is cultural momentum. The mythology has been told so many times by so many successful people that it has acquired the status of received wisdom. Y Combinator has structured its entire brand around it for twenty years. Paul Graham’s essays, which shaped a generation of founders, are full of “do things that don’t scale,” “launch before you’re ready,” and “move fast to capture the market.” These ideas were grounded in specific conditions that prevailed when Graham was active in the ecosystem. The conditions changed; the essays are still being read.

Part of the answer is that the mythology is economically useful to specific people. VCs benefit from the “move fast to capture winner-take-all markets” narrative because it justifies the portfolio model of investing — spray capital broadly, expect most investments to fail, wait for the massive winner. Founders benefit from the credibility that comes with performing the Silicon Valley ethos. It signals boldness. It signals ambition. It attracts certain kinds of investors and employees.

And part of the answer is that the people building AI companies today were trained in an era when the mythology was more closely connected to the mechanism. The 35-year-old founder who cut their teeth at Stripe or Airbnb or Lyft has strong, visceral intuitions from environments where moving fast was genuinely rational. Those intuitions are being imported into a context where they’re wrong, but intuitions formed in successful environments are hard to update.

The actual strategy for building AI companies in 2026 is almost the opposite of “move fast and break things.” It involves identifying specific, narrow tasks where AI adds reliable, measurable value. It involves deploying in those tasks only, with extensive instrumentation to catch failures before they become liability events. It involves building regulatory relationships and audit capabilities proactively, so that when the regulations arrive — and they are arriving — you have compliance infrastructure rather than a crisis response. It involves being honest with clients about what the system can and cannot do, which is commercially harder in the short term and catastrophically better in the long term.

That strategy is slow. It doesn’t generate the kind of growth curves that justify billion-dollar valuations on the timelines most VCs need. It doesn’t make for great conference keynotes. The companies executing it mostly don’t get written about in TechCrunch.

There are companies doing something close to this approach. They’re building AI tools for specific regulated applications — clinical documentation, legal discovery, financial compliance — with mandatory human review requirements, clear liability frameworks, and deployment practices that would be recognizable to a regulated industry veteran. They’re not the companies generating the most press. They’re also not the companies accumulating the most liability exposure.

The irony is that the “move fast” mythology, which was developed by people who were genuinely successful, is now being applied by a generation of founders who are taking the costume without understanding the conditions that made the underlying strategy rational. Zuckerberg’s “move fast and break things” was right for Facebook in 2007 — building a social network in a regulatory vacuum, in a winner-take-all consumer market, with cheap capital and no high-stakes liability exposure. Applied to healthcare AI or financial AI in 2026, it’s a recipe for the kind of expensive, regulatory, and reputational disaster that ends companies.

The companies performing the Silicon Valley ethos of 2012 in 2026 are going to collide with the legal and financial reality that’s arriving. The collision will not be pretty. There will be a Watson-style reckoning — expensive, documented, irreversible — and everyone will be surprised, even though they shouldn’t be.

The mythology will survive the collision anyway. Myths are more durable than evidence. But the companies won’t.