Building Tomorrow's Digital Ruins: The Economics of AI and Data Centers

We are spending more money building data centers than at any point in human history. Most of them will be obsolete in four years. What does that mean for the global economy and for us?

$650 Billion — Global data center spend in 2026, 3–4 years - Functional lifespan of AI chips (vs. 5–6yr accounting life), $3–5 Trillion - Projected cumulative investment by 2030

There is a pattern to how civilisations build themselves. They pour money into infrastructure railways, roads, electrical grids, fiber optic cables, and in doing so, they bankrupt themselves briefly before the benefits of that infrastructure unlock decades of productivity growth. History is consistent on this point.

In a recent episode of The Diary of a CEO, entrepreneur Daniel Priestley articulated the danger in our current moment with unusual precision. He noted that over the past 180 years, whenever a society spent more than 3% of GDP on a single infrastructure buildout, it produced a temporary economic crisis lasting roughly a decade. Railways, electrification, highways, each time repeating the same arc. But when we built train tracks, they lasted 100 years. Roads, 50 plus. Telecommunications and fiber optics, 30. All multi-decade investments that societies could leverage for generations.

There is something fundamentally different happening now. For the first time in the history of infrastructure buildouts, the thing we are building has a useful life not of decades, but of years. The AI data center boom is the largest concentrated capital expenditure in modern history and it may also be the most rapidly self-obsoleting.

The Scale Is Unprecedented

The numbers are difficult to fully absorb. The global data center sector is projected to expand at a 14% compound annual growth rate through 2030, with nearly 100 gigawatts of new capacity being added between now and then, effectively doubling global capacity. The sector is experiencing an infrastructure investment supercycle requiring up to $3 trillion by 2030.

In 2024, Amazon, Microsoft, Google, and Meta collectively spent over $200 billion on capital expenditures, representing a 62% year-over-year increase from 2023. Amazon's CapEx was $85.8 billion — up 78% year-over-year. In 2025, hyperscaler spending neared $400 billion annually. The race has no obvious ceiling. Google co-founder Larry Page was reportedly quoted saying he would "go bankrupt rather than lose this race."

Harvard economist Jason Furman noted that investment in information processing equipment and software represented just 4% of US GDP in the first half of 2025, yet accounted for a staggering 92% of GDP growth during that period. Without the AI data center boom, annualized US GDP growth would have been approximately 0.1%, essentially flat. One commentator quipped that the US economy had become "three AI data centers in a trench coat."

The Obsolescence Problem Nobody Is Talking About

Here is the central paradox that should concern every investor, policymaker, and citizen: the infrastructure being built at this extraordinary scale has a useful functional life of roughly three to four years.

The chips at the heart of this buildout have a real useful lifespan of one to three years due to rapid technological obsolescence and physical wear, yet companies are depreciating them over five to six years, what The Economist has called "the $4trn accounting puzzle at the heart of the AI cloud." Microsoft CEO Satya Nadella essentially acknowledged this gap when he said: "I didn't want to go get stuck with four or five years of depreciation on one generation."

Barclays has already begun cutting earnings forecasts for AI firms by up to 10% to account for more realistic depreciation assumptions. The potential impact could represent a $176 billion earnings adjustment through 2028, fundamentally altering the P/E valuations of the largest companies driving the current market.

Industry experts warn of premature obsolescence. Even facilities completed just a few years ago fall short for today's compute-heavy workloads. As power usage spikes, entire campuses require expensive retrofits or complete replacement. Even "state-of-the-art" builds risk obsolescence by completion, given lead times and fast-moving technical requirements.

The Historical Pattern and Why This Time Is Worse

Infrastructure capital expenditures on AI data centers account for 1.2% of US GDP in 2025 exceeding the telecom buildout of the early 2000s (1.0%) and trailing only the railroad boom of the 1880s (6.0%). When infrastructure booms are externally financed, they historically have tended to overshoot inviting both capital misallocation and speculative excess.

But there is a critical difference. Every previous infrastructure investment that caused temporary economic pain ultimately delivered multi-decade returns. AI data centers are being financed on those same long time horizons, the hyperscaler typically signs a lease with a developer's SPV, which borrows the debt for typically three to five years to fund construction. JPMorgan projects about $300 billion of AI and data-center-related deals every year for the next five years. The debt structures presume longevity that the technology does not deliver.

The Global Economic Ripple Effects

Data centres and related high-tech investment activities have recently become a key driver of US growth, with 80% of the increase in final private domestic demand in the first half of 2025 attributable to data centres and related high-tech spending. Most high-tech hardware used in US data centres is imported, with the surge beginning in late 2023, the main beneficiary economies have been Taiwan and Mexico, with Malaysia and South Korea also showing recent increases.

A March 2026 Brookings Institution report documented that electricity costs have risen 42% since 2019, significantly outpacing inflation. Utilities requested $31 billion in rate hikes during 2025 alone. "The fundamental question is whether middle-class families should subsidize the electricity needs of companies worth trillions of dollars."

The Lawrence Berkeley National Laboratory predicts that data center demand will grow from 176 terawatt hours in 2023 to between 325–580 TWh by 2028 from 4.4% to potentially 12% of total US electricity consumption.

The Productivity Question

None of this necessarily means the investment is irrational. The counterargument is that the world will ultimately be glad to have this computing infrastructure, even if markets overpay for it in the short term.

But a 2025 MIT Media Lab report found that 95% of custom enterprise AI tools and pilots fail to produce a measurable financial impact. Gartner reported that 85% of AI initiatives fail to deliver on their promised value. A University of Chicago economics paper found AI chatbots had "no significant impact on workers' earnings, recorded hours, or wages" at 7,000 Danish workplaces.

These are early-stage findings. The technology is moving fast. But they suggest the productivity payoff that would justify the investment and absorb the obsolescence cost is not yet materializing at scale.

Priestley, characteristically, points to the Jevons Paradox as a reason for measured optimism: that transformative technologies often create more economic activity than they destroy, just not in the forms we predicted. The minute an AI learns how to be a lawyer in one place, it can be a lawyer in every place. It's this instantaneous rollout because it sits on top of a network that already exists. The question is whether that wave arrives before the debt structures financing today's buildout begin to mature.

What Comes Next

2026 is the year infrastructure buildout begins to shift, data centers are transitioning from cost centers to revenue generators, and the new top-of-mind metric in industry circles is "tokens per watt per dollar." The focus is no longer on simply using less energy, but using energy as efficiently as possible.

Construction costs are already rising, averaging $11.3 million per MW in 2026, up 6% year-over-year, with AI infrastructure tech fit-out costing as much as $25 million per MW on top of that.

The companies that will fare best are those who build with modularity in mind, who phase their investments to remain agile, and who don't confuse the accounting life of their assets with their actual economic life. The companies that will struggle are those who locked in fixed-cost structures against assets whose obsolescence is accelerating faster than their depreciation schedules.

For everyone else, governments, grid operators, ratepayers, and communities being reshaped by this buildout the most important question is whether the extraordinary productivity potential of AI will materialize quickly enough to justify the costs being distributed across the broader economy. The historical pattern says this takes about a decade to resolve. And this time, the infrastructure at the center of the bet will likely need replacing twice before that decade is done.

Whilst I love the promise of AI, it has some real hurdles. Economic, environmental and other, and these hurdles are not small. A doomed business model. Increasingly high demands on power needs and the moral question "just because we can, should we?" are all at the forefront here.

I am a Star Trek fan, and thinking about AI in terms of a world fashioned after its pricniples excites me. It doesn't excite everyone though, as it removes the need for what we know as our current economy. It puts the betterment of mankind first and knowledge and exploration above all. It's idealistic. But it can be real, we just need to change the way we think about the future. And maybe thinking about it in terms of not allowing the economy to fail, by removing the need for our society to be underpinned by an economy is the first step. Then upgrading these data centers every 3-4 years doesn't mean a collapse of the economy and hard times ahead.

I would love to know your thoughts...

If you are interested in my sources, the conversation that sparked this post was on the Diary of a CEO podcast, the Daniel Priestley episode.

Here is a full list of sources: Daniel Priestley on Diary of a CEO (March 2026); JLL 2026 Global Data Center Outlook; Harvard economist Jason Furman; Princeton CITP; Belfer Center; World Economic Forum; St. Louis Fed; Brookings Institution; MIT Media Lab; Deloitte AI Infrastructure Survey 2025; JPMorgan/Bloomberg data center debt analysis.

Are Humanoid robots and AI data centers the backbone of the future?