Boom, Bust, Build:
Where AI Fits in the Great Technology Cycle
Every few generations, a new technology arrives that promises to reshape the economy. Capital floods in, valuations soar, and a wave of new companies races to claim the future. Then, almost without exception, the wave breaks. Yet what’s left behind on the shore – the tracks, the roads, the cables – often turns out to be the real prize.
Artificial intelligence is the latest technology to ride this wave. To understand where it might be heading, it helps to look at where the great booms of the past began, peaked, crashed and ultimately paid off.
The shape of every technology boom
Most major technologies move through a recognisable lifecycle. Economists describe it in five broad phases:
Irruption → Frenzy → Crash → Golden Age → Maturity
Irruption.
A breakthrough technology appears and early adopters begin experimenting.
Frenzy.
Investment runs hot. Companies spend aggressively, often out of fear of being left behind rather than clear demand. Valuations detach from profits.
The Crash.
Reality catches up. Over-leveraged players fail, and a great deal of paper wealth disappears.
Golden Age.
The infrastructure built during the frenzy survives the crash – and becomes cheap. Disciplined survivors use it to build genuinely productive businesses.
Maturity.
The technology becomes ordinary, ubiquitous business infrastructure, and attention moves to the next wave.
The crash, in other words, isn’t the end of the story. It’s often the moment the technology finally becomes useful to everyone.
History: the three booms that
built the modern world
The railway boom (1840s–1890s)
The technology: high-pressure steam locomotives, standardised steel rails, telegraph lines and air brakes.
The infrastructure: a vast network of graded track beds, tunnels, bridges and urban rail terminals.
Through the mid-1800s, private companies raced to lay competing lines, often running track straight alongside a rival’s purely to grab market share. The over-building eventually triggered the financial panics of 1873 and 1893, which bankrupted well over a hundred rail companies.
But the tracks didn’t go anywhere. Ownership consolidated into a handful of efficient operators, and the standardised network went on to power American industrial logistics for the better part of a century. The steel rails laid during the frenzy had a useful life of more than 100 years.
The automobile boom (1910s–1970s)
The technology: the internal combustion engine, the moving assembly line, hydraulic brakes and cheap oil.
The infrastructure: sealed roads, the interstate highway system, service stations, refineries and entire suburbs.
Henry Ford’s assembly line made cars affordable, and through the 1920s hundreds of manufacturers sprang up, fuelled by a boom in consumer credit. The Great Depression of 1929 then wiped out most of them, leaving only the “Big Three” – GM, Ford and Chrysler.
After the Second World War, government-funded highways turned the car into the defining engine of the American economy. The roads and bridges built in that era still carry traffic today – infrastructure with a 50-year-plus lifespan, and a suburban geography that proved permanent.
The dot-com boom (1990s–present)
The technology: microprocessors, web browsers, the HTML and HTTP protocols, and early databases.
The infrastructure: hundreds of thousands of kilometres of underground fibre-optic cable, server farms and network routers.
In the late 1990s, almost any company with “.com” in its name could attract a multi-billion-dollar valuation regardless of profit. Telecoms laid enormous quantities of “dark fibre” – cable installed in anticipation of demand that hadn’t arrived yet. When the market crashed in 2000–2002, trillions in value evaporated and major hardware providers collapsed.
Here’s the pattern again: roughly 85% of that fibre sat unused in 2001 – but the glass stayed in the ground, and leasing it became incredibly cheap. That cheap, abundant infrastructure is precisely what allowed Amazon, Google, Netflix and others to build the modern digital economy a decade later. The cloud and mobile boom that followed simply extended this maturity into the ubiquitous infrastructure we take for granted today.
The current wave: AI in the frenzy phase
By this framework, the US economy is now in the frenzy phase of a new cycle driven by generative AI and automation. Capital is pouring into data centres and chipmakers, echoing the early infrastructure build-outs of the railway and dot-com eras. Three signals stand out.
1. An extreme infrastructure build-out
In a healthy cycle, infrastructure spending tracks demand. In a frenzy, it tracks fear of missing out.
The four biggest “hyperscalers” – Microsoft, Alphabet, Amazon and Meta – are projected to spend a combined US$650 billion in a single year on capital expenditure, largely on data centres and AI chips. For scale, that one-year figure is estimated to exceed the inflation-adjusted cost of building the entire US Interstate Highway System.
2. A widening “revenue gap”
A classic sign of a bubble peak is when spending races ahead of the revenue that’s meant to justify it.
By some estimates, there’s around a US$600 billion annual gap between what’s being spent on AI infrastructure and what the AI software ecosystem actually earns. Analysts have put this divergence at roughly 46% – notably wider than the telecom infrastructure bubble that preceded the dot-com crash, which is estimated to have peaked nearer 32%.
3. Enterprise disillusionment
The open-ended excitement of 2023–24 has given way to harder questions about return on investment. High computing costs mean that, for some tasks, advanced AI can cost more than the human labour it was meant to replace – without delivering the promised productivity gains. In one widely cited finding, more than 80% of surveyed businesses reported that generative AI had not yet made a material difference to their earnings, leaving many stuck in the pilot stage.
How the AI boom differs from the dot-com bust
Prominent investors – including The Big Short’s Michael Burry – have argued that today’s market resembles the final months of the dot-com bubble. There’s a strong case for that view, but there are also important structural differences that could change how any correction plays out.
| Feature | The dot-com bubble (2000) | The AI boom (2020s) |
| Primary speculators | Unprofitable startups funded largely by retail investors | Cash-rich tech giants spending from existing balance sheets |
| The bottleneck | Under-used fibre-optic capacity | Shortages in energy grids and semiconductor supply |
| Fallout mechanism | Mass bankruptcies of small internet firms | Margin compression and share-price corrections for Big Tech |
The takeaway isn’t that one is “safer” than the other – it’s that the type of pain could differ. A dot-com-style event hit thousands of fragile startups. An AI correction might instead show up as squeezed profit margins and falling share prices among a smaller number of very large, well-capitalised companies.
The twist: AI infrastructure may be built to die fast
There’s one feature that makes the AI build-out genuinely different from every boom before it – and it’s easy to miss. In past cycles, the infrastructure laid down during the frenzy lasted for decades, which is exactly why it could be repurposed cheaply during the golden age.
AI infrastructure splits into two very different lifecycles: a durable outer shell, and a highly disposable silicon core.
| Boom | Primary infrastructure asset | Useful lifespan | What happened after the crash? |
| Railways | Track beds, steel rails, tunnels, bridges | 100+ years | Companies failed; the tracks remained and powered 20th-century freight |
| Automobiles | Roads, bridges, interstate highways | 50+ years (with resurfacing) | Created permanent suburbs and lasting logistical value |
| Dot-com | Underground fibre (“dark fibre”) | 20–25 years | ~85% unused in 2001, but the glass survived to fuel streaming a decade later |
| Artificial intelligence | AI chips (GPUs), servers, custom switches | 1–3 years | High burnout and rapid model advances mean old hardware can’t easily be reused |
The Silicon Problem
The GPUs that train AI models face a double threat:
- Physical burnout. Training clusters run heavy mathematical workloads at close to full capacity around the clock. The heat and power density wear down the chips, often causing failure within one to three years.
- Rapid obsolescence. Chipmakers release dramatically faster hardware every 12 to 18 months. A two-year-old cluster simply can’t keep pace with the next generation of models.
The Shell that Survives
Not all of the spending evaporates. The buildings, advanced cooling systems and – crucially – the high-voltage grid connections have a longer life of roughly 15 to 20 years. If the frenzy cools sharply, these shells won’t run cutting-edge AI. More likely, they’ll be repurposed as cheap, commoditised capacity for cloud storage, corporate databases and lighter computing tasks.
In short: the long-lived legacy of the AI boom may be power and real estate, not the chips themselves.
What history suggests happens next
If AI follows the path of the railways or the internet, a correction isn’t a freak accident – it’s a structural feature of the cycle. When the frenzy cools and over-valued players are repriced, the physical infrastructure built during the boom becomes cheap and abundant.
That’s usually the moment the golden age begins: when practical, affordable, deeply integrated applications finally drive broad economic productivity, rather than headline valuations.
The honest answer is that no one can time these turning points, and the AI cycle has features – fast-dying hardware, energy bottlenecks, Big Tech balance sheets – that make it genuinely new. History doesn’t repeat, but it does tend to rhyme. Understanding the verse can help you read the market with a steadier hand.
What you learn here has been used in our Trade for Good software.
Click on the button to find our software education videos.
You can read more of our educational articles in the Trade for Good Learn section
Trade for Good Learn
