The AI bubble is about to burst

Kind: captions Language: en AI was supposed to kick off a white collar purge, but the reality is different. In Fortune 500 boardrooms across the country, thousands of AI agents are being shut down. Not because they didn't work, but because they're a legal time bomb. The narrative is that tech giants are spending everything to win an AI arms race. But the reality is much messier. They're stripping the liability out of their systems while pouring billions into the buildout. And when the dust settles, someone still has to pay for it. The US interstate highway system is 48,000 mi long. It took 35 years to build and cost $630 billion in today's money. It is by any measure one of the most consequential infrastructure investments in human history. Meanwhile, big tech is spending $725 billion in 2026 on AI infrastructure alone. That number has more than doubled in just 2 years and is projected to keep climbing into the trillions by 2030. In fact, Goldman Sachs now estimates that the total AI infrastructure bill between now and 2031 will exceed $7.61 trillion. For that kind of money, you'd expect a revolution. What these companies are getting right now is cents on the dollar. The entire global AI services market in 2025 generated around $25 billion against hundreds of billions of infrastructure spending. When a Chinese startup called Deepseek revealed in January 2025 that it had built a competitive AI model using Nvidia's lower-end chips for under $6 million, it seemed almost impossible. But it still shook the market, wiping out around $590 billion from Nvidia's market cap in a single day. It was the largest single day loss in US stock market history. Deepseek, far from breaking AI, merely raised the question every investor had asked himself. What exactly are we paying for? The companies spending this money don't just have it lying around. They are borrowing it. While Alphabet quadrupled its long-term debt in 2025 to 46.5 billion, the five biggest tech companies issued $18 billion in new bonds last year alone. Bank of America now estimates that AI Capex now consumes 95% of operating cash flows after dividends. Amazon is projecting a 95% decline in free cash flow thanks to its AI buildout costs. They are in essence mortgaging their present against a future that hasn't shown up yet. The gamble is genuinely alarming. If it were startups chasing hype, that would be one thing. These are the most sophisticated capital allocators on the earth. When they spend like this, they either know something the rest of us don't, or they're caught up in something they can't stop without admitting they were wrong. Either way, something has broken. To understand what, you have to go back to the moment a grieving man in Vancouver asked an airline chatbot a simple question and changed corporate legal history forever. On the day his grandmother died, Jake Moffett opened his laptop and went to Air Canada's website to book a flight from Vancouver to Toronto. He had one question. Did the airline offer bereavement fairs and could he apply for one after the fact? He asked the chatbot and it said yes. It told him to submit his request within 90 days of purchase, and the discount would be retroactively applied. So, Moffett booked the ticket, flew to Toronto, came home, and then filed the request like he'd been told, and Air Canada denied it. Their actual policy, as it turns out, required bereavement fair requests to be made before travel. The AI chatbot had simply made up a policy that didn't exist, stated it with complete confidence, and sent a grieving man to buy a full price ticket based on a lie. So Moffett sued. And what happened next is the reason every corporate lawyer in America now has an opinion about chatbots. Air Canada's legal defense was interesting. They argued it could not be held liable for information provided by its chatbot because the chatbot was quote a separate legal entity responsible for its own actions. They had deployed a piece of software to represent their company. And when that software caused harm, their position was essentially don't look at us. Tribunal member Christopher Rivers was not impressed. While a chatbot has an interactive component, he wrote, "It is still just a part of Air Canada's website, it should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot." Air Canada was ordered to pay $812 Canadian, and in perspective, that amount is trivial. The precedent was not. That ruling landed in February 2024. What followed was a corporate panic that never made headlines, but could be seen in the risk sections of SEC filings. In 2020, 4% of public companies mentioned AI as a material risk in their annual filings. By 2024, that number was 43%. Among the Fortune 500 specifically, 56% now lists AI as a formal risk factor. Among media and entertainment companies, it's 92%. The SEC itself has been sending enforcement letters demanding more specific disclosures backed by $ 8.2 billion in financial remedies in the year 2024 alone. Climate change took decades to achieve similar disclosure penetration across corporate America. AI did it in 3 years. What the Air Canada ruling established is that whatever your AI says to a customer, you said it. You can't outsource the liability to the software you employ. You can't just point at the machine and then look innocent. You are the machine. For a technology that can't guarantee it won't make things up, that ruling created an absolute certainty. Deploy AI at scale and eventually liability follows. The bigger question is, will there be anyone left to cover the bill? In July 2024, research and advisory firm Gartner predicted that at least 30% of enterprise gen AI projects would be abandoned after proof of concept by the end of 2025. By April 2026, they had to revise their original estimate. It turns out that it was actually more like 40% of enterprise AI projects launched in the previous 2 years that would be scrapped before they ever shipped. Separately, Gartner now predicts that over 40% of Agentic AI projects, the next generation of AI tools companies are currently betting on, will be cancelled by the end of 2027. The graveyard keeps growing. S&P Global Market Intelligence put some finer numbers on it. The average organization scrapped 46% of AI proof of concepts before reaching production. Of the projects initiated, only 48% make it to production with an average of eight months between prototype and launch. That is eight months of engineering hours, infrastructure costs, and executive attention on a coin flip chance of shipping anything at the end of it. McKenzie, another consulting firm, surveyed organizations globally in late 2025, and they found that 88% use AI in at least one business function. If that sounds like a success story, it's not. Only 39% reported any measurable impact on earnings. At most, 61% reported no meaningful financial impact on their business whatsoever. Roughly six in 10 companies have plugged AI into their operations and can't point to a measurable return it's made them. The reasons are familiar enough that Gardner has named them outright. Poor data quality, weak risk controls, rising costs, and unclear business value. Each project costs between $5 and $20 million to build and deploy. Multiply that across thousands of Fortune 500 initiatives and you're looking at tens of billions of dollars buried in the corporate graveyard. It gets even worse for big tech. A 2024 Deote survey found that 47% of enterprise AI users admitted to making at least one major business decision based on the content AI had simply invented. In other words, nearly half had already trusted a hallucination with realworld consequences. Granted, things have evolved a lot since 2024. The efficiency revolution is still an aspirational possibility as Agentic AI improves. But you wouldn't be human if you hadn't wondered why the flaws you've seen firsthand seem absent from the market's calculations. There's a particular breed of corporate fear that shows up where you would least expect it. Insurance contracts. Basically, if the people whose entire business model depends on their ability to correctly price risk deem something uninsurable, that is usually a huge red flag. In January 2026, the Insurance Services Office introduced new exclusions targeting generative AI. Essentially, if generative AI causes the damage, don't assume you're covered. New policy language began excluding claims tied to AI generated text, images, audio, video, and code. That is every large language model ever built. the chat bots, the AI co-pilots, the autonomous agents, all of them circled with black ink and excluded from the policies that cover 82% of US businesses. In 2024 and 25, insurers had been narrowing AI coverage in existing policies by tightening up language and adding exclusions. WR Berkeley introduced an absolute AI exclusion across directors and officers, errors and omissions, and fiduciary liability lines. AIG saw that and they said, "Well, we'll have more of the same." Great American Insurance followed. Chub, Travelers, Bergkshire Hathaway received state regulatory approval for their own exclusions. State regulators have been approving more than 80% of those requests. The industry has seen this movie before. A decade ago, cyber insurance went through an identical cycle. There was invisible coverage until claims mounted and then came the exclusions. Eventually, standalone products emerged at premiums most companies couldn't stomach. The AI liability market is speedr running the same cycle. Insurance companies are not moral actors. They do not exclude things because they find them distasteful. They exclude them because the numbers don't work. They tried to price in AI liability throughout 2024 and 2025. And 42% of insurers reported tracking no AI risk metrics at all. When you can't measure risk, you can't price it. And when you can't price it, you exclude it. The legal system said, "Whatever your AI does, you did it." the insurance industry said, "And we won't be covering that." Which left every company still running public-f facing AI deployment, holding the entire liability themselves, all on hardware that was simultaneously losing most of its value. In the summer of 2023, if you wanted to rent a single Nvidia H100 GPU, a foundational chip in the inception of the AI arms race, you were paying between $7 and $10 an hour. To buy one outright on the secondary market, you were looking at $40,000 to $50,000 per card. Startups were putting H100s on their balance sheets like assets, and venture capitalists were asking portfolio companies how many they had on their books. Nvidia's market cap soon passed the $1 trillion threshold, and it kept rising. In February 2026, it hit 5 trillion. The H100 was, for a brief, dizzying moment the most strategically valuable piece of hardware on the Earth. Today, you can rent one for $2 an hour. In June 2025, AWS cut H100 pricing by 45% overnight. The secondary market collapsed right behind it. GPUs that had sold for as much as 50 grand during the AI boom were suddenly worth less than half that. Entire H100 server systems that once cost more than $350,000 were now trading for a fraction of their peak value. Over 300 new GPU cloud providers entered the market in 2025 alone. All of them selling capacity that nobody was buying at the rate anyone projected. Thousands of companies weighing up where to sell the cards at a loss. It is the difference between making payroll and not making payroll. This signals a deeper problem about the infrastructure thesis that justified the entire buildout. The argument was always that compute was scarce. That scarcity created value and whoever controlled the most compute would win out in the AI race. Deepseek fractured that argument by building a model competitive with GPT4 on a budget that Meta spends on catering. The telecom industry learned the lesson the hard way in the late 1990s when companies invested more than $500 billion in fiber optic cables based on projections of a 1,000% annual internet traffic growth. When monetization stalled, the debt markets closed. Miles of fiber sat dark for years owned by companies that no longer existed. The asset had been real, but the business model built on top of it had not. No amount of Google fiber buildouts by the 2010s could fix that. The H100 fire sale is the fiber glut of 2025. The hardware is real. The revenue was supposed to follow. It hasn't. Today, the trillion dollar companies watching their hardware depreciate have made a calculated decision on what to do with their cash instead. There's a move corporations make when they believe their own stock is a better investment than their future. They take their cash and they use it to buy back their own shares. This reduces the number of shares in circulation, boosting earnings per share, which then lifts the stock price. Investors love it. Boards love it. Everyone wins. Yet, it is a signal that the smartest people in the room looked at every opportunity available and they chose not to bet on the future. Instead, they bet on the thing already making money today. Over the last 5 years, the companies loudest about the AI revolution have been doing exactly this at a scale that is difficult to fully absorb. Alphabet has spent about $280 billion buying back its own stock. Meta and Microsoft have spent tens of billions. Apple, which barely talked about AI until recently, has spent $74 billion buying back its own stock over the past decade. That's more than the entire market value of 488 companies in the S&P 500. And then the data center bill started piling up. All of the money disappeared. In September 2024, Microsoft promised shareholders a $60 billion buyback program. It signaled that the company was flush, confident, and committed to returning cash to the people who owned it. A year later, $57.3 billion of that promise still sits completely untouched. Microsoft didn't change its mind, but every spare dollar was getting redirected into paying for servers, data centers, and AI infrastructure instead. In other words, they wrote the check. They just couldn't afford to cash it. This is what the AI arms race really looks like from the inside. The story being sold is one of a confident sprint into the future. The reality is a scramble to fund it. Companies are being squeezed between investor expectations and the soaring cost of AI infrastructure. For Microsoft, that's led to a $60 billion commitment sitting on the books like an IOU that nobody can collect. Alphabet's free cash flow is projected to plummet almost 90% in 2026 from 73.7 billion down to $8.2 billion. And in 2025, Alphabet, Amazon, Oracle, Meta, and Microsoft issued $121 billion in new debt via bonds to keep the lights on. That is four times what the entire tech industry borrowed on average in any year of the previous decade. The AI revolution is being funded on credit, not profits. And the org charts that were supposed to prove this was working have started to change. At some point in 2023, someone in a corporate communications department typed the words generative AI into a press release and felt genuinely excited about their job. The phrase had some weight. It had mystery. It implied that the machine was creating something, thinking, almost generating. The word did a lot of heavy lifting for a lot of companies that had no idea what they were actually going to do with the technology. By 2026, the impetus for generative AI has turned into agentic workflows. People are talking about autonomous process orchestration and intelligent automation. It's the exact same technology, but the promises for what it can deliver are dramatically smaller. Today, the gap between the language of 2023 and the language of 2026 tells you almost everything you need to know about what happened in between. The conversation has shifted from possibility to proof, from what could this do to what did it actually deliver. The companies that survived the graveyard of AI project rollouts did so by making their AI projects smaller, narrower, and harder to sue over. Not by funneling all their cash into chat bots that could answer any question, or AI that writes company strategy. Well, what this means is a lot of routing software that sends customer service tickets to the right department and AI that scans contracts for non-standard clauses. Things that are dull, specific, and measurable. Things that are wrong in ways that humans can still catch before they get expensive. And the job titles reflect that. In 2025, 26% of organizations had a chief AI officer or CIO. By 2026, that number jumped to 76%, tripling in a single year. According to an IBM study of 2,000 CEOs across 33 countries, but read the fine print on what those CIOS are actually being hired to do, and it is obvious the role has fundamentally changed. It used to be AI evangelists promoting the technology. Now, its executives responsible for things like risk management, regulatory compliance, and governance. The EU AI act kicks in fully in August 2026, requiring companies to identify exactly who is accountable when their AI causes harm. The CIO is increasingly the person whose name goes on that form. Probably not the job description anyone imagined when they coined the title, but here we are. Meanwhile, the outcome of the earnings calls has quietly shifted, too. AI powered is out. outcome focused workflows is in which is if you read it slowly enough a very expensive way of saying that the AI does a specific task and you can measure whether it worked or not. The dream that produced the $725 billion buildout was always something closer to science fiction. AI that could run an entire department, write strategy, replace an entire team or what have you. C AIOS with compliance checklists and the EU AI act obligations are now doing work that would have been unrecognizable to the people who wrote the original press releases. The revolution got a job in middle management. And the reason it ended up there is something the industry has been trying to explain away since the beginning. Back in 2023, when ChatGpt was still a twinkle in many people's eyes, attorneys Steven Schwarz and Peter Luduka filed a legal brief in the Southern District of New York. The brief cited very real sounding cases with very real sounding names and very real sounding outcomes as president. Vargasi v. China Southern Airlines, Martinez v. Delta Airlines and Shaboon v. Egypt Air and then Judge P. Kevin Castell looked them up and none of them existed. Chad GPT had invented every single one with absolute confidence. Schwarz and Luca were sanctioned, publicly humiliated, and their names are now permanently attached to one of the most cautionary legal filings in American judicial history, Mata v. Aviana. That was in 2023. Things didn't improve. In August 2025, US District Judge Allison Bacas sanctioned a lawyer whose brief contained 12 fabricated or unsupported citations out of 19. A California court ordered two law firms to pay just over $31,000 in fees for filing what the judge described as bogus AI generated research. In February 2025, three lawyers from the national firm Morgan and Morgan were sanctioned for the same thing. There are now over 1,600 documented cases involving AI generated hallucinations with 79% of lawyers reporting that they use AI tools internally in their practice. Why does this keep happening? Why hasn't it been fixed? Well, because it can't be. Large language models are actually very poor at looking things up. They're not attuned to retrieving facts from specific databases and handing them to you on a platter. Rather, their entire mechanism, start to finish, is to predict the next most statistically likely word based on patterns learned from training data. That's it. When