Why tech CEOs are quietly cancelling their AI plans

AI was supposed to automate away huge swaths of white-collar work. Instead, something very different is happening inside Fortune 500 boardrooms: flashy AI agents and public chatbots are being quietly shut down, even as companies spend record amounts on AI infrastructure.

The problem isn’t that AI doesn’t work. It’s that, at scale, it creates legal, financial, and technical risks that many CEOs are no longer willing—or able—to carry.

The $7.6 trillion AI bet that isn’t paying off (yet)

Big Tech is in the middle of one of the largest infrastructure buildouts in history. The U.S. Interstate Highway System cost about $630 billion in today’s dollars. By comparison, major tech companies are on track to spend around $725 billion on AI infrastructure in 2026 alone, with Goldman Sachs projecting more than $7.6 trillion in total AI capex by 2031.

For that kind of money, you’d expect a clear economic revolution. Instead, the global AI services market generated only about $25 billion in 2025—cents on the dollar compared to what’s being spent on chips, data centers, and power.

To fund this, tech giants are borrowing heavily. Alphabet’s long-term debt quadrupled in 2025 to $46.5 billion. The five biggest tech companies issued $108 billion in new bonds in a single year. Bank of America estimates AI capex now eats up roughly 94% of operating cash flows after dividends. Amazon expects a 95% decline in free cash flow due to AI buildout costs.

In other words, the AI revolution is being funded on credit, not profits. And the returns are nowhere near where investors were led to believe.

The Air Canada chatbot that changed corporate law

The turning point for many corporate lawyers came from a small case with a big message.

When his grandmother died, a passenger named Jake Moffat asked Air Canada’s website chatbot if he could get a bereavement fare discount after booking his ticket. The chatbot said yes and told him to submit his request within 90 days. He followed the instructions, then the airline denied the refund—because its real policy required bereavement fares to be requested before travel.

Moffat sued. Air Canada’s defense was that it wasn’t responsible for what the chatbot said, arguing the bot was a “separate legal entity.” The tribunal rejected this outright: if it’s on your website, you’re responsible. The airline was ordered to pay just C$812, but the precedent was huge.

The ruling effectively said: whatever your AI tells a customer, you told them. You can’t outsource liability to software. For a technology that sometimes simply makes things up, that’s a legal time bomb.

From hype to risk factor: how AI became a liability

After that ruling, AI quietly started showing up in a new place: the “Risk Factors” section of annual reports.

In 2020, only 4% of public companies mentioned AI as a material risk. By 2024, that number had jumped to 43%. Among the Fortune 500, 56% now list AI as a formal risk factor. In media and entertainment, it’s a staggering 92%. The SEC is pushing for more detailed disclosures and backed that up with $8.2 billion in financial remedies in 2024 alone.

AI went from “strategic opportunity” to “regulatory and legal risk” in about three years—faster than climate change ever did.

Why insurers are quietly excluding AI

If you want to know what really scares corporations, look at their insurance policies.

In early 2026, the Insurance Services Office rolled out new exclusions targeting generative AI. Many policies now explicitly exclude claims tied to AI-generated text, images, audio, video, and code. That covers every major large language model and most AI copilots and agents.

Insurers like W.R. Berkeley, AIG, Great American, Chubb, Travelers, and Berkshire Hathaway have all moved to tighten or exclude AI coverage across directors and officers (D&O), errors and omissions (E&O), and other key lines. State regulators are approving more than 80% of these exclusion requests.

Insurers aren’t moral arbiters—they’re risk calculators. When they can’t measure a risk, they can’t price it. When they can’t price it, they exclude it. By 2025, 42% of insurers admitted they weren’t tracking any AI risk metrics at all.

The result: courts say “whatever your AI does, you did it,” and insurers say “and we won’t be covering that.” Any company running public-facing AI is now holding the full liability themselves.

The graveyard of abandoned AI projects

While the legal and insurance risks were rising, another problem was becoming impossible to ignore: most enterprise AI projects weren’t delivering meaningful returns.

Gartner initially predicted that 30% of enterprise generative AI projects would be abandoned after proof of concept by the end of 2025. By April 2026, they had to revise that up to 40%. They now expect over 40% of agentic AI projects—the next wave of autonomous AI tools—to be cancelled by 2027.

S&P Global Market Intelligence found that organizations scrap, on average, 46% of AI proof-of-concepts before they reach production. Only 48% of projects make it to launch, and it takes about eight months to get there. That’s eight months of engineering time, infrastructure spend, and executive attention for a coin-flip chance of shipping anything.

McKinsey’s global survey in late 2025 found that while 88% of organizations use AI in at least one business function, only 39% reported any measurable impact on earnings. That means roughly six in ten companies using AI can’t point to a clear financial benefit.

Each project typically costs $5–$20 million to build and deploy. Multiply that across thousands of initiatives and you end up with tens of billions of dollars buried in a corporate AI graveyard.

For a deeper look at why many leaders misread this moment, see this breakdown of the AI delusion in the C-suite.

The hallucination problem that won’t go away

Underneath the legal and financial issues is a stubborn technical reality: large language models hallucinate, and that behavior is baked into how they work.

LLMs don’t “look things up” the way a database does. They predict the next most likely word based on patterns in their training data. When they don’t know something, they don’t go silent—they generate the most plausible-sounding continuation.

That’s how we got infamous legal cases like Mata v. Avianca, where lawyers submitted briefs full of citations to cases that simply didn’t exist. Despite public embarrassment and court sanctions, similar incidents kept happening. By 2025, there were over 1,600 documented cases of AI-generated hallucinations in legal filings, and 79% of lawyers reported using AI tools internally.

OpenAI researchers have described hallucinations as “mathematically inevitable” in current LLM architectures. A 2026 benchmark across 37 leading models found hallucination rates between 15% and 52%. Even the best models get things wrong roughly one in seven times. In legal queries, Stanford found hallucination rates as high as 88%. In medical case summaries, 64% hallucinated without mitigation.

A human employee who was confidently wrong 20% of the time wouldn’t last a week. Yet these systems have been deployed to millions of users at once—and companies are expected to own the consequences.

The GPU fire sale: when the infrastructure story breaks

For a while, the AI boom was built on a simple thesis: compute is scarce, and whoever owns the most GPUs wins.

In 2023, renting a single Nvidia H100 GPU could cost $7–$10 an hour. Buying one on the secondary market could run $40,000–$50,000. Startups bragged about how many H100s they had on their balance sheets. Nvidia’s market cap passed $1 trillion and kept climbing, eventually hitting $5 trillion in early 2026.

Then the economics flipped. By mid-2025, AWS slashed H100 rental prices by about 45% overnight. Suddenly, you could rent one for around $2 an hour. Secondary market prices collapsed. Entire H100 server systems that once cost more than $350,000 were selling for a fraction of their peak value.

More than 300 new GPU cloud providers launched in 2025, all selling capacity that demand hasn’t caught up to. Many are now deciding whether to sell their cards at a loss just to make payroll.

At the same time, Chinese labs like DeepSeek showed they could build competitive models on cheaper hardware for under $6 million—wiping nearly $600 billion off Nvidia’s market cap in a single day. The idea that “only hyperscalers with massive capex can compete” suddenly looked shaky.

This echoes the telecom boom of the late 1990s, when companies poured over $500 billion into fiber optic cable based on wildly optimistic traffic projections. When the revenue didn’t materialize, miles of fiber sat dark, owned by companies that no longer existed. The infrastructure was real; the business model wasn’t.

The H100 fire sale is AI’s fiber glut moment. The hardware is real. The revenue hasn’t shown up.

Share buybacks vs. AI: where CEOs are really betting

There’s a move companies make when they believe their existing business is a better bet than new investments: they buy back their own stock.

Over the last decade, the loudest AI cheerleaders have been some of the biggest buyers of their own shares. Alphabet has spent roughly $280 billion on buybacks. Meta and Microsoft have spent tens of billions. Apple, which was relatively quiet on AI until recently, has spent about $704 billion buying back its stock—more than the entire market value of 488 companies in the S&P 500.

But as AI data center bills piled up, that behavior started to change. In September 2024, Microsoft announced a $60 billion buyback program. A year later, $57.3 billion of that authorization was still untouched—not because the company changed its mind, but because every spare dollar was being redirected into AI infrastructure.

Alphabet’s free cash flow is projected to fall almost 90% in 2026, from $73.3 billion to $8.2 billion. In 2025, Alphabet, Amazon, Oracle, Meta, and Microsoft issued $121 billion in new debt—four times the industry’s average annual borrowing over the previous decade—just to keep funding AI buildouts.

The story sold to the market is a confident sprint into the future. The reality looks more like a scramble to keep the lights on in data centers.

From “generative AI” to “agentic workflows”

Listen closely to how companies talk about AI now compared to 2023, and you can hear the comedown.

Back then, “generative AI” sounded magical—machines that could create, think, and maybe even replace entire teams. By 2026, the buzzwords have shifted to “agentic workflows,” “autonomous process orchestration,” and “outcome-focused workflows.” It’s the same underlying tech, but the promises are much smaller and more grounded.

The conversation has moved from “what could AI do?” to “what did it actually deliver?” The projects that survived the first wave of cancellations are narrower, duller, and much harder to sue over: routing customer tickets to the right department, scanning contracts for non-standard clauses, automating specific back-office tasks.

These are use cases where errors are easier to catch, impact is measurable, and legal exposure is lower. It’s a far cry from the early dream of AI running entire departments by itself.

If you’re interested in how some leaders are quietly making this more modest approach work, see this case study of a CEO who rebuilt his company with targeted AI.

The new role of the Chief AI Officer

As the hype cooled, a new executive role exploded: the Chief AI Officer (CAIO).

In 2025, only about 26% of organizations had a CAIO. By 2026, that number had jumped to 76%, according to an IBM survey of 2,000 CEOs across 33 countries.

At first, CAIOs were mostly AI evangelists—people tasked with selling the vision internally and driving adoption. Today, their job descriptions look very different. They’re increasingly responsible for risk management, regulatory compliance, governance, and making sure AI deployments don’t land the company in court.

The EU AI Act, which fully kicks in from August 2026, forces companies to name exactly who is accountable when their AI causes harm. More and more, that name is the CAIO. The AI revolution, in other words, has landed squarely in middle management and compliance.

Where AI is actually making money

While many enterprise AI projects struggle, there is one sector where AI is printing money: quantitative trading.

Firms like Jane Street, Citadel Securities, and Hudson River Trading are quietly running some of the most profitable AI operations on earth. Jane Street alone generated $39.6 billion in net trading revenue in 2025—more than Goldman Sachs or JPMorgan’s entire trading arms—with around 3,000 employees across four offices.

What’s different? Not the compute—the GPUs are similar. It’s the feedback loop. In high-frequency trading, data is clean, outcomes are clear, and mistakes are punished instantly. If an AI model is wrong, the market takes money away in microseconds. That brutal, real-time feedback forces constant retraining and improvement.

By contrast, a customer service chatbot might hallucinate a policy, mislead a user, and only face consequences months later via a complaint or lawsuit. The feedback is slow, noisy, and often ambiguous. That makes it much harder to improve models and control risk.

The winners of the current AI wave are those who’ve found problems where wrong answers are caught fast, cost real money, and can be corrected quickly. Everyone else is stuck with expensive infrastructure and unclear returns.

The data center squeeze and the coming reset

All of this is putting enormous pressure on the companies building and financing AI infrastructure.

Take CoreWeave, a fast-growing AI infrastructure provider. It went from $16 million in revenue in 2022 to $1.9 billion in 2024, largely by operating GPU clusters for AI workloads. But it also carries $24.5 billion in total debt, with $7.5 billion in interest payments due by the end of 2026. About 62% of its revenue comes from a single customer: Microsoft—the same Microsoft that’s already stretching its balance sheet to fund AI capex.

Financial historians see echoes of the early 2000s telecom crash: massive infrastructure buildouts, optimistic demand forecasts, and then a painful realization that monetization isn’t keeping up.

Investor Michael Burry has argued that hyperscalers are depreciating Nvidia’s chips over five to six years, even though their true economic life may be closer to two or three. He estimates that depreciation is understated by about $176 billion across the industry through 2028.

At the same time, Chinese AI labs are closing the performance gap with U.S. frontier models in weeks, at a fraction of the cost. New open-weight models can rival GPT-5 on engineering benchmarks at roughly one-sixth the price per token. Intelligence is getting cheaper faster than the data centers built to sell it can depreciate.

Roughly half of U.S. data centers planned for 2026 are already facing delays or cancellations. Nvidia has even released a $4,700 desktop machine, the DGX Spark, that can run models which needed entire server rooms just two years ago. If that level of compute can sit on a desk, the justification for endless billion-dollar data centers starts to crack.

The demand for AI won’t vanish—but it doesn’t need to collapse to cause problems. It just has to grow slower than investors were promised.

So why are CEOs cancelling AI plans?

When you put it all together, the quiet cancellation of many AI initiatives starts to make sense.

Public-facing AI agents and chatbots expose companies to legal risk they can’t insure away. The underlying models hallucinate at rates no human employee could get away with. Insurers are excluding AI from coverage. Courts are holding companies fully responsible for what their AI says and does.

At the same time, the economics are underwhelming. Infrastructure spending is measured in hundreds of billions; revenue is measured in tens of billions. Many enterprise projects don’t make it to production, and most that do fail to move the bottom line in a measurable way.

Meanwhile, GPU prices are falling, competitors are building strong models on cheaper hardware, and some of the most aggressive infrastructure bets may never pay off as originally pitched.

So CEOs are doing what they always do when the story changes: they’re narrowing scope, cutting the riskiest projects, and focusing on small, defensible use cases with clear ROI. The AI revolution hasn’t been cancelled—but the dream that it would instantly replace entire departments and mint effortless profits has.

What’s left is something more modest and more realistic: AI as a powerful tool for specific workflows, tightly governed, heavily monitored, and increasingly overseen by people whose job is to say “no” as often as they say “yes.”