The Bubble We Built
AI — The Morning After, Part 1 of 4
Three years ago, I wrote a short series of articles asking a simple question: AI is here — now what? The articles were exploratory, honest about the uncertainty, and tried to cut through the early noise to offer something practical for business leaders. I did not have all the answers. Nobody did.
A lot has happened since.
I want to revisit the subject — not because the earlier thinking was wrong, but because we now have something we did not have in 2023: evidence. Real data. Three years of enterprise adoption, investment decisions, energy bills, and — critically — outcomes. Or the lack of them.
This is the first of four articles taking a clear-eyed look at where we actually stand with AI. Not where the vendors say we stand. Not where the breathless headlines suggest. Where we actually stand.
Let us start with the money.
The numbers that should make you pause
The scale of investment in AI infrastructure over the past two years is genuinely staggering.
The four largest technology companies — Amazon, Google, Meta, and Microsoft — spent over $330 billion on AI infrastructure in 2025 alone. OpenAI has committed to spending $1.4 trillion over eight years building new data centres, partnering with Nvidia to deliver ten gigawatts of compute capacity. That commitment is backed by $13 billion in annual revenue, funded largely by debt, and the company has acknowledged it expects to report annual losses through 2028 — reaching an estimated $74 billion in operating losses that year alone.
Morgan Stanley estimates total global spending on data centres between 2025 and 2028 at $3 trillion, half of which is covered by private credit.
These are not small bets hedged across a diversified portfolio. These are existential commitments. And when you look at the financial structures being used to fund them, some of the language starts to feel uncomfortably familiar.
Take one widely reported arrangement: a $27 billion loan for a data centre, backed by a major technology company’s lease payments, structured so the debt never appears on the company’s balance sheet. The mechanism is called a special purpose vehicle. Twenty-five years ago, that phrase entered public consciousness through a company called Enron.
I am not predicting an Enron-scale collapse. I am saying the structural resemblance is worth acknowledging, rather than ignoring.
What are we actually getting for the investment?
Here is where the evidence gets uncomfortable for anyone who has spent the past two years enthusiastically presenting AI business cases to their board.
A report published in August 2025 by MIT Media Lab assessed $30–40 billion in enterprise investment in generative AI and found that 95% of organisations reported zero measurable return. A National Bureau of Economic Research study from February 2026 found that despite 90% of firms reporting no impact of AI on workplace productivity, executives were projecting AI would increase productivity by 1.4%.
The gap between what leaders believe and what the data shows is not a rounding error. It is a chasm.
The San Francisco Federal Reserve, reviewing the macro evidence, put it plainly: most studies find limited evidence of a significant AI effect, and even firms that say the technology is useful find little evidence of transformative gains.
Now, one reasonable response is: it is simply too early. Technologies take time to permeate. The electricity grid did not immediately transform manufacturing productivity either — it took decades of organisational and process redesign before the full benefits materialised. That parallel is real and worth taking seriously.
But there is a difference between “too early to see the full benefit” and “spending at a scale that cannot be sustained by current revenues.”
The dot-com analogy is useful here, not because history repeats itself exactly, but because the pattern of debt-financed infrastructure build-out for a future that has not yet arrived is a recognisable one.
Vast amounts of fibre-optic cable were laid in the late 1990s. The infrastructure was eventually used — just not by the companies that paid for it.
The circular economy of AI investment
One of the more striking features of the current AI landscape is how circular the investment flows have become.
Nvidia sells chips to OpenAI, then invests $100 billion back into OpenAI to fund data centres, which OpenAI fills with Nvidia’s chips. CoreWeave, once a cryptocurrency mining startup, pivoted to building data centres and entered deals with OpenAI worth tens of billions, with Nvidia guaranteeing to absorb any unused capacity through 2032. OpenAI pays for CoreWeave capacity partly in OpenAI stock; CoreWeave uses that stock to offset what OpenAI owes.
One MIT economist described arrangements like this as “a house of cards.”
That may be too dramatic — but the circularity of these deals does mean that genuine end-user demand is harder to identify beneath the layers of cross-investment. When Nvidia is essentially subsidising one of its largest customers, the demand signal it receives is not a clean one.
Google’s CEO, asked how his company would fare if the bubble burst, said he thought no company would be immune, including Google. That is an unusual thing for a CEO to say publicly. It suggests the people closest to these dynamics are not as confident as the market valuations might imply.
The question nobody is asking loudly enough
In late 2025, 30% of the S&P 500 and 20% of the MSCI World index was held up by just five companies — the highest concentration in half a century, with valuations at their most stretched since the dot-com era.
I want to be clear about what I am and am not arguing here. I am not predicting a crash. I am not suggesting AI is a mirage. I wrote in 2023 that AI is a genuine shift, and I still believe that. The technology is real, the capabilities are expanding, and there will be organisations that get significant and lasting competitive advantage from it.
What I am arguing is that the gap between the scale of infrastructure investment and the demonstrable return from enterprise adoption is large enough to warrant serious scrutiny — particularly from leaders who are being asked to approve AI budgets without clear evidence of what those budgets are actually delivering.
Final Thought
Writing about speculative episodes in 1955, indeed, John Kenneth Galbraith observed that a bubble is not so much a situation in which people are fooled, as one in which they insist on fooling themselves — and that any suggestion values are unreal is fiercely resisted1. Seven decades on, that dynamic feels entirely recognisable. The next article in this series looks at the resource reality that most AI conversations quietly skip: the energy cost of running all this infrastructure, and what it means for how organisations — and societies — should be thinking about AI as a strategic choice rather than a given.
But before we get there, a wicked question worth sitting with:
If your organisation stopped all AI investment tomorrow, what would you actually lose?
If the honest answer is “a few licence fees and some pilots that haven’t delivered much yet,” that is important information. It does not mean you should stop — but it does mean the business case for continuing deserves considerably more rigour than it may currently be getting.
This is Part 1 of ‘AI — The Morning After’, a four-part series revisiting AI adoption and what it means for leaders today. The original 2023 series can be found here.
About the author
I’m Frank Smits, a change and transformation consultant with about 30 years of international experience helping organisations navigate complex business and IT-driven change. I have particular expertise in setting up and managing global HR programmes, including the implementation of HRIS solutions such as SAP SuccessFactors.
I’ve worked with global teams across industries and cultures to deliver major transformations—balancing strategy, execution, and the human side of change. Based in Europe, I work in multiple languages and thrive on making change practical, collaborative, and real.
What I can offer:
I help you shape and manage the engagements to achieve your business outcomes. By bringing in my specialist programme management, change and transformation expertise. From initiation through to implementation. Or any part thereof. This includes leading large-scale business change initiatives, from digital transformation to complex HR programmes.
As executives, you may need a discrete partner to test your ideas. Or get fresh, new ones. I will act as your sparring partner. To having the right conversations. Helping you succeed. This applies to all areas of business leadership—including how to navigate and lead major HR change or transformation initiatives.
Expertise in designing and facilitating innovative and engaging interventions. And, if needed, I can convene a variety of experts from my extensive network (academics, peers, consultants, active retirees) to open up new thinking. In HR programme leadership, this means bringing together key stakeholders—HR, IT, business leaders, and external partners—to drive alignment and engagement.
Sometimes you would like your leaders to get dedicated, customised, specialist education. To enhance your capabilities. To possibly change the conversation. HR leaders facing major change or transformation also need the right tools and perspectives to guide their teams. I can help build that capability.
Find me on LinkedIn: https://www.linkedin.com/in/franksmits
Source: John Kenneth Galbraith, testimony to the US Senate Stock Market Study, March 1955.