Very few AI companies have produced repeatable, audited ROI (return on investment) metrics. Meanwhile, capital expenditures (CapEx) and OpEx keep climbing.
The chain runs in one direction. Enterprises buy AI without baselines. Vendors book the revenue. Mega-scale cloud operators read that revenue as demand and borrow against it, and the resulting AI data centre debt is increasingly parked in special purpose vehicles that never touch a balance sheet. Bondholders buy the paper, and the derivative market sells insurance on the paper. Every link in that chain rests on the first one holding, and the first one is the weakest documented claim in the entire economy.
MIT‘s NANDA initiative reported last year that 95% of enterprise pilots showed no measurable P&L impact. Morgan Stanley found only 21% of S&P 500 companies mentioned an AI benefit at all in 2025, and by the second quarter of 2026 about 25% reported a measurable one.
Of course, there are the outliers that have succeeded in adopting AI into existing workflows. But there’s a dearth of granular information that gives a sense of what AI is doing and what it’s not doing.
While enterprise AI is taken on by more and more companies, ROI is often an afterthought. Meanwhile, employees utilize unauthorized AI to speed workflow, a practice known as shadow AI, which most often isn’t included in internal metrics.
Part of the problem seems to be an accounting issue. Financial disclosures pertaining to AI are rather scant. A few years ago, reporting how many employees were using enterprise AI and how many hours they logged seemed to satisfy investors.
The Numbers Nobody Kept
In the mad dash to inject AI into corporate workflows, profitability in companies like Meta and Microsoft took a back seat to AI adoption. Tokenmaxxing, the practice of turning AI agents loose across a company with no controls on what they consume, emerged as a trend. One executive told Fortune their spend on Anthropic went from $20,000 in December to nearly $1 million by July, a 50x increase in seven months, with no matching gain in output.
Since then, after soaring enterprise costs, the practice has been discouraged. Now, as Big Tech is taking on billions of debt to finance infrastructure buildouts, boards and investors want to see the numbers. The problem is that they often don’t exist.
While the economy is increasingly being propped up by a spending spree on AI infrastructure, a loss of demand could have severe consequences. For now, companies are continuing to spend on enterprise AI, although the metrics aren’t always there to support the cost. What happens when boards and investors demand a proper audit?
Where Marketing Meets Accounting
Much of what is marketed as artificial intelligence isn’t. Far from the frontier models being put out by Anthropic and the like, companies are marketing basic services that utilize standard statistics, data processing, and automation as AI. The reason is simple: to boost product sales and investor hype.
Beyond AI washing, in which companies rebrand as tech-focused, major companies are en masse employing AI. But, given the lack of information, it’s difficult to decipher what is driven by hype and what is driven by sound business practices.
The CEO of IBM, Arvind Krishna, questioned ROI on data center spending in December of 2025:
“There’s no way you’re going to get a return on that in my view because eight trillion of capex means you need roughly 800 billion of profit just to pay for the interest.”
How Did Credit Default Swaps End Up on AI Data Centre Debt?
Krishna also spoke of chip depreciation cycles, a reality Big Tech has largely brushed aside. Instead, major players have moved beyond self-funding and issued billions of debt via special purpose vehicles, debt which lives outside corporate balance sheets.
As insurance for these corporate bonds issued by Big Tech, the derivative market is bubbling with credit default swaps (CDS). If anyone can remember the 2008 financial crisis, credit default swaps didn’t cause it, but they made it a whole lot worse when issuers didn’t have the money to pay bondholders.
Given AI companies’ lack of focus on ROI, the emergence of credit default swaps attached to data center debt is a bit disconcerting.
Half the Companies, None of the Proof
Plug and Play Tech Center, a venture capital firm based in California, recently released a report on enterprise AI. The report shows that 74% of large enterprises run at least one AI solution in production. However, the report, which surveyed leaders of Fortune 500 companies, also showed that half of the respondents were either too early to assess ROI or were not measuring it consistently.
Really, that shouldn’t be surprising. The dearth of metrics for AI spend has publicly shown itself already. Uber infamously used up its entire 2026 AI budget in 4 months earlier this year.
Another 2026 survey initiated by Writer, an AI company based in San Francisco, found that 75% of executives admit that AI strategy is “more for show than actual guidance.” Writer, for their part, surveyed 2,400 executives and employees.
Lanai, a company that focuses on enterprise AI accountability, put out its own report that focused on AI labor and poor accounting. The report, which surveyed 200 executives at companies with 1,000 or more employees, found that 88% of businesses have no formal methodology for attributing outcomes to AI. Moreover, 87% of organizations credit AI output entirely to the human employee, sometimes or always.
In the report, Lanai makes recommendations to board members on what issues to focus on in order to gain actionable insight on ROI. Among them, Lanai cites the absence of baseline metrics from which to extrapolate gains made by AI. It also suggests that AI spend should be treated as a labor cost, not as an IT expense.
Does the AI Data Centre Debt Math Survive an Open-Weight Pivot?
While Big Tech has taken on debt to fund massive AI infrastructure projects, the idea is that the demand will be there to pay for it all. There’s very little available evidence to suggest this is the case.
Meanwhile, major companies entrenched in AI don’t have the information to know exactly what their AI spend is accomplishing. They do know, however, that open-weight models are increasingly a better cost-effective option.
A large pivot away from closed models to open-source AI could affect the viability, if not the solvency, of AI infrastructure being built. This could mean massive impairment cycles in the future, visible in mid-cap software before the hyperscalers.
While the surveys and research reports referenced above present a range of stats, they collectively show AI adoption moving faster than methods for its accountability. The lack of information means that Big Tech is building out the future based on a first wave of enterprise AI adoption, which could dry up or change directions.
CFOs, board members and investors are growing wary of AI platitudes and hype marketing. They want the nuts-and-bolts metrics that show ROI. What they find could have real ramifications for the tech sector and the larger economy.
When the Auditors Show Up
Eight trillion in capex needs roughly 800 billion in profit to service the interest. Nobody at Meta, Microsoft, Amazon or Google has produced a public figure that gets close, because the companies buying the output cannot produce one either.
Lanai found 88% of organizations with no methodology for attributing outcomes to AI. Plug and Play found half of production deployments unmeasured. Writer found three quarters of executives calling their own AI strategy a performance.
None of that is fatal on its own. Enterprises have overbought technology before and absorbed the loss quietly. What makes this different is the leverage sitting on top of it, and the fact that CDS volume tied to the AI sector rose almost 600% in a year while the underlying evidence stayed exactly where it was.
The audit is coming, from CFOs first and bondholders second. Big Tech is building the future on a first wave of enterprise adoption that nobody bothered to measure. When someone finally does the measuring, the number will either justify the debt or it won’t, and everyone from Santa Clara to the corporate bond desk will find out at the same time.
Author: Tim Tolka, Senior Reporter
The editorial team at #MRKT3.0 has taken all precautions to ensure that no persons or organizations have been adversely affected or offered any sort of financial advice in this article.
See Also:
OpenAI Needs to Grow 20x in 5 Years. Europe Might Not Play Along.
Why AI Super PACs Are Avoiding the Word “AI” at All Costs
AI Billionaires Are Donating Billions to Study the Danger of AI Billionaires
