The Coming AI Crash
A Familiar Euphoria
I lived through the dot-com boom and the crash that followed. The late 1990s were filled with the same euphoria we see today around artificial intelligence. Companies with no revenue were suddenly worth billions. Public markets rewarded buzzwords more than balance sheets. For a moment, it felt as if technology had suspended the laws of business. Then the bubble burst.
The parallels today are impossible to ignore. Generative AI is being marketed as the dawn of a new industrial revolution. Investors are told it will transform every industry, replace vast categories of workers, and create trillions in new value. Executives are racing to prove they have an AI strategy. The hype is relentless.
But the evidence points elsewhere. This is not the beginning of an endless golden age. It is the setup for an AI crash. The question is not if. It is when, and how deep.
Valuations Detached From Reality
The scale of today’s AI valuations is staggering. Nvidia’s market cap is trending toward $5 trillion, propelled almost entirely by AI chip demand. Microsoft, Google, Amazon, and Meta have each added hundreds of billions in market value on the strength of an AI narrative, not measurable AI revenue.
Startups with little more than access to GPUs and a flashy demo are valued at ten, twenty, even fifty billion dollars. Many of these companies have no sustainable revenue model, no proven technology moat, and no path to profitability. Yet capital continues to chase them, driven by fear of missing out.
This is not unlike 1999, when companies like Pets.com commanded enormous valuations despite fragile business models. At the time, investors convinced themselves that revenue and profit would come later. Today, many make the same argument for AI. The technology, they say, is transformative, and the business case will follow. But history suggests otherwise. Markets do not forgive inflated valuations forever.
Failure Rates and Business ROI
If the valuations are difficult to justify, the business results are even weaker. A recent MIT study, reported in Fortune, found that ninety five percent of generative AI pilots at large enterprises fail to deliver measurable return on investment.
Companies are spending millions experimenting with copilots, chatbots, and generative tools, but few can connect these projects to profit and loss. Chief financial officers may tolerate experiments for a year or two. But when experiments become a line item without a return, they are cut.
We saw this dynamic with corporate websites in the early 2000s. At first, simply having an online presence was considered valuable. Then boards asked a harder question: where is the revenue? Many internet companies disappeared overnight. AI projects that cannot show impact will face the same reckoning.
This is not the first time it has happened in artificial intelligence either. The field has already endured two winters before. In the 1970s, promises of symbolic reasoning collapsed when machines could not scale beyond toy problems. In the 1980s and 1990s, expert systems attracted enormous attention until their brittleness and maintenance costs became obvious. Each cycle ended with disillusionment, funding cuts, and retrenchment. The same forces are at work again today.
Where AI Offers Real Value and Its Limits
To be clear, AI is not useless. It does create value in specific, narrow domains. The danger is pretending that incremental productivity gains are equivalent to an industrial revolution.
Coding and Software Development
AI coding assistants like GitHub Copilot can generate boilerplate code, unit tests, and documentation. They save time, especially for junior developers. But they cannot replace the best engineers. They cannot design a distributed system, secure a protocol, or make trade-offs between cost and scalability. AI raises the baseline. It does not reach the frontier of human creativity.
Law and Contracts
In legal practice, AI accelerates discovery, contract review, and research. It can summarize documents and highlight risks. Yet it also fabricates cases, introduces errors, and cannot exercise strategy or judgment. Judges and clients rely on trust, credibility, and advocacy. These are qualities no machine can replicate.
Customer Support
Chatbots can handle simple FAQs and transactional requests. They reduce call center volume. But the moment nuance, emotion, or escalation is needed, they fail. Customers quickly lose patience with scripted AI and demand human service.
Marketing and Content Drafting
Generative AI can draft emails, posts, and campaign ideas. It scales routine communication. But it is derivative, often generic, and risks plagiarism or brand dilution. True creative direction, the spark that defines a brand, still comes from people.
Data Analysis and Reporting
AI tools quickly summarize datasets and produce dashboards. They automate repetitive reporting. Yet they cannot ensure accuracy, identify causality, or deliver genuine insight. Analysts must still validate every result.
Healthcare
AI supports doctors by drafting notes and flagging anomalies in imaging. It helps reduce administrative burden. But it does not replace diagnosis, treatment, or care. Patients require trust, empathy, and responsibility that only human doctors can provide.
Education
AI tutors can personalize quizzes and adapt lessons. For rote practice, they are helpful. But education is not just about information transfer. It is about curiosity, imagination, and mentorship. Overreliance on AI risks undermining human learning, not enhancing it.
Across these domains, the pattern is clear. AI automates repetitive, text-heavy work. That is useful, but incremental. It is not transformational.
The Autonomous Driving Illusion
Take for example self-driving cars. Over the past decade, more than one hundred billion dollars has been poured into autonomous vehicle programs. We were told cars would be driving themselves by now. Instead, the reality is stark.
Only Waymo has managed limited deployments, and even those required years of training on the same San Francisco roads. The technology struggles the moment conditions change. Waymo cars cannot seamlessly transition to a new city such as New York without enormous retraining. Despite unlimited data, fixed road rules, and over a hundred billion in investment, AI still cannot solve driving.
It is not just Waymo. General Motors’ Cruise division was forced to scale back after safety incidents. Uber abandoned its self-driving unit entirely after billions in losses. Tesla continues to market “Full Self-Driving,” but the product requires constant driver supervision and has been subject to regulatory scrutiny.
Driving should, in theory, be one of the easier AI challenges. The rules are clear. The environment is structured. The task is repetitive. If AI cannot fully master driving, then the promises of general-purpose intelligence in far more complex domains look unrealistic at best.
Unsustainable Costs
The economics of generative AI are another red flag. Training a frontier model today costs hundreds of millions of dollars. Running those models at scale requires massive energy and data-center buildouts.
Most startups are simply wrappers around existing foundation models. They face high costs, thin margins, and no durable competitive advantage. This is the same as dot-com retailers paying exorbitant fees to buy ads on portals like Yahoo in the 1990s, hoping to survive on wafer-thin margins. When capital dries up, the economics break.
The environmental costs also cannot be ignored. Training and serving large models consumes vast amounts of electricity and water. As regulators and the public become more aware of this footprint, pressure will grow. Expensive and resource-hungry systems that fail to deliver returns are unlikely to survive long-term scrutiny.
The Winner Takes It All Dynamic
Another reason the current AI boom will not end well for most participants is that this is a winner takes it all market. In practice, a handful of large players with access to capital, data, and distribution will dominate. Microsoft, Google, Amazon, and Meta already control the infrastructure. Nvidia controls the chips. A few model providers such as OpenAI, Anthropic, and Mistral are building at scale, but even they depend on Big Tech for funding and cloud power.
This leaves thousands of startups competing with no moat. Most are thin wrappers around the same foundation models, offering features that can be copied overnight by a larger competitor. The economics do not allow dozens of winners. Just as the dot-com boom left us with Amazon, eBay, and a few survivors while thousands of others became dot-gones, the AI boom will leave only a few dominant platforms standing.
The rest will vanish, no matter how much capital they raise today.
Built on Thin Air
The brutal truth is that ninety nine percent of AI startups are built on thin air. They do not own infrastructure, they do not train foundation models, and they do not control data. What they offer is usually a simple interface that calls ChatGPT or another large language model through an API. The value is not in the software they build, but in the marketing spin around it.
This explains why so many of them attract sky-high valuations despite having no moat. If OpenAI or Anthropic changes its pricing, or if Microsoft or Google rolls out the same feature, the entire startup collapses overnight. These are not technology companies in any durable sense. They are feature demos packaged as businesses.
It is the same pattern we saw in the dot-com era, when thousands of companies claimed to be building the future but were really just front-ends on someone else’s infrastructure. When the market corrected, only those with true scale, true customers, and true intellectual property survived. The same will happen again.
The Data Mirage
Another myth propping up today’s AI valuations is the illusion of unique data. Investors are told that companies have proprietary datasets that give them an advantage. In reality, most are training on the same scraped internet text, images, and video. The foundational models already consumed it all.
Without proprietary data, there is no moat. The companies that truly own unique datasets, such as those in finance, healthcare, or defense, are rare. They will capture the value. Everyone else is building on an empty promise.
This mirrors the dot-com era when startups measured success in “eyeballs.” Everyone claimed attention was a valuable asset, but very few could monetize it. Today’s AI companies counting on “data advantage” will learn the same hard lesson.
The Productivity Trap
Supporters argue that even with flaws, AI boosts productivity. It helps write memos, summarize meetings, or produce drafts faster. That is true. But productivity gains at the margin are not enough to justify multi-trillion valuations.
The printing press was transformational because it created entire industries. The internet was transformational because it unlocked global commerce. Generative AI mostly speeds up tasks that were already being done. It saves time, but it does not create new categories of economic activity.
This is the productivity trap. Companies will adopt AI to reduce costs, not to drive new growth. Valuations built on the assumption of entirely new markets are bound to collapse when cost savings are the only measurable impact.
The Trust and Regulation Wall
Finally, AI faces an approaching wall of regulation. Governments are waking up to issues of copyright, privacy, bias, and liability. Europe has already passed comprehensive AI laws. China is imposing strict controls on content. Compliance costs will be enormous. Many startups will not survive them. Even the largest players will face limits on what they can deploy. Investors counting on explosive, unregulated growth will be disappointed.
This is another echo of the dot-com era. After the crash, only companies that could handle regulation around payments, securities, and logistics survived. The same filter is coming for AI.
The Psychology of Bubbles
Technology bubbles are fueled not just by money, but by psychology. In 1999, investors believed every website was revolutionary. Today, the same psychology drives AI. Venture capital firms fear missing the next Google or Amazon. Boards push management teams to adopt AI whether or not it fits the business.
But bubbles always end the same way. When revenue fails to match valuations, when projects fail to deliver, the market corrects. The correction is not gentle. It is brutal.
What Will Survive
This does not mean AI itself will vanish. Applied AI, the kind that has been working for decades in fraud detection, cybersecurity, and logistics, will endure. These systems generate real return on investment and are embedded in critical industries.
Generative AI will not disappear, but the gold rush around it will collapse. Thousands of startups will vanish, valuations will contract, and funding will freeze. The survivors will be the few with domain-specific applications, access to proprietary data, or deep integration with Big Tech infrastructure.
The dot-com bust did not kill the internet. It cleared away excess so real businesses could grow. The same will happen with AI. The technology will remain, but the fantasy of limitless AI transforming every aspect of society will not.
Conclusion
I witnessed one bubble burst firsthand. I see the same warning signs today. Valuations are divorced from fundamentals. Projects fail to scale. Costs are spiraling out of control. Investors are chasing hype instead of business models. This is how crashes unfold. They are not mysterious events but predictable corrections.
An AI crash is certain. The real question is how deep the fall will be, and how much capital will disappear before reality sets in.
