The $7 Trillion AI Gamble Has Big Tech Trapped

Seven trillion dollars is the kind of number that stops feeling like money. It sounds fictional, like something a villain requests before switching off the moon.

Yet that is roughly the upper-end price tag attached to the data centres now being proposed for the AI boom. Not already spent. Not already lost. Proposed. That distinction matters, because the internet has taken a complicated financial warning and compressed it into a much cleaner story: AI is failing, and Big Tech has set fire to $7 trillion.

I do not think that story is accurate. The more unsettling reality is that AI is working well enough to keep the race alive, but not yet profitably enough to make the scale of the race feel safe. Microsoft, Google, Amazon and Meta cannot comfortably stop building. They also cannot keep spending at this pace forever without proving that the machines will earn much more than they cost.

That is the trap.

First, what does “$7 trillion” actually mean?

The figure is not a confirmed Big Tech bill. A Reuters Breakingviews analysis counted roughly 110 gigawatts of announced data-centre projects. Using estimates of $36 billion to $60 billion per gigawatt, the theoretical construction cost lands between about $4 trillion and $6.6 trillion.

So the viral “$7 trillion gamble” is a scenario built from the full pipeline of planned capacity. Some projects will be delayed, resized or cancelled. Others may never leave a slide deck. Saying Big Tech has already gambled and lost $7 trillion is therefore wrong.

But dismissing the number completely would also be foolish. Even if only a portion of those plans becomes concrete, this is still one of the largest private infrastructure races in history. AI is no longer merely a software story. It is a construction, energy, financing and depreciation story wearing a chatbot as its friendly little mask.

Big Tech is trapped by a game it cannot pause

Imagine four people bidding for the only bridge into the future. Each suspects the bridge may be overpriced. None can walk away, because the moment one stops bidding, the others gain control of the road.

That is the AI arms race. If Microsoft slows down while Google keeps building, Azure risks losing customers who need more computing capacity. If Google becomes cautious while Microsoft and Amazon continue, Gemini and Google Cloud risk falling behind. Meta may not sell cloud capacity in the same way, but it believes better AI can protect its advertising machine and define the next computing platform.

The individually rational decision is to keep spending. The collective result may be too much capacity, squeezed margins and an industry forced to lower prices simply to keep expensive chips busy. This is my inference from the incentives, not a disclosed corporate plan. It is also why “trapped” is more accurate than “finished.”

The companies themselves are telling us how intense the race has become. Alphabet spent $44.9 billion on capital expenditure in the second quarter of 2026, mostly on AI-related technical infrastructure, and raised its full-year guidance to $195–205 billion. That investment helped push quarterly free cash flow to negative $5.9 billion. Yet the same Alphabet earnings call reported Google Cloud revenue growth of 82% and a backlog of $514 billion.

That single result contains the entire contradiction: the spending pressure is real, and so is the demand.

The returns exist, but they are arriving unevenly

The bearish case becomes weaker when it pretends nobody is making money. Microsoft reported that Azure and other cloud services revenue grew 43% in its latest fiscal quarter, while Microsoft Cloud revenue reached $59.3 billion. Its full-year revenue and operating income both grew by double digits. That is not what a dead technology looks like.

Meta is also seeing strong top-line growth. In the second quarter of 2026, revenue rose 28% year over year. But its quarterly capital expenditure reached $31.08 billion, while free cash flow was only $784 million and operating income fell 8%. Legal charges and severance contributed to that pressure, so it would be dishonest to blame the whole squeeze on AI. Still, the figures show how little room enormous infrastructure spending can leave, even inside one of the world’s most profitable businesses.

The weaker point is further down the chain. According to McKinsey’s 2025 global AI survey, 88% of respondents said their organisations regularly used AI in at least one function, but only 39% reported an enterprise-level impact on earnings before interest and taxes. Roughly 6% qualified as high performers generating significant value.

In plain English: companies are using AI, but most have not redesigned their work deeply enough to turn usage into major financial returns. A chatbot subscription here and an auto-generated meeting summary there will not repay a continent of data centres.

The real bottleneck is not intelligence. It is economics.

Models can improve quickly. Power stations, transmission lines, cooling systems and data centres do not materialise at software speed. The International Energy Agency projects that global data-centre electricity consumption will more than double to around 945 terawatt-hours by 2030—slightly more than Japan consumes today—with AI as the biggest driver of the increase.

This is where the conversation looks different from Ghana. From here, a trillion-dollar compute race is not only an American stock-market story. It is a reminder that the AI economy will be shaped by who has reliable electricity, cheap capital, chips, land, cooling water and political influence. Countries that mainly import AI services may pay for intelligence built on infrastructure they do not own.

That does not mean Ghana should attempt to copy a hyperscale race it cannot finance. It means our strategy should be brutally specific: use smaller models where they work, build local datasets, create products around local problems and invest in the people who can turn existing models into useful systems. Owning the most GPUs is not the only way to own value.

I have seen the smaller version of this while building with AI. It is easy to confuse a dramatic demo with a durable product. The UCC SRC app taught me that AI can help one person move frighteningly fast, but it does not remove the need for verification, maintenance and judgement. I wrote more about that gap in my account of vibe-coding the UCC SRC app.

So, is the gamble failing?

Not yet. “Failing” is too final for evidence this mixed.

AI has real users, real revenue and real productivity gains. The infrastructure providers are reporting demand that still exceeds supply in important areas. But the industry is also front-loading gigantic costs while enterprise-wide returns remain concentrated among a small minority of adopters.

The most plausible danger is not that AI suddenly becomes useless. It is that the technology becomes common, capable and cheap before today’s infrastructure owners recover what they paid to build it. Society could win while some investors lose. The internet survived the dot-com crash; many dot-com companies did not.

Big Tech is trapped because retreat looks like surrender and acceleration looks increasingly expensive. Eventually, somebody will discover whether these data centres are gold mines or the most sophisticated empty rooms ever built.

The machines are getting smarter. The bet is whether their economics can catch up.