Picture a cutting somewhere in the north of England in the autumn of 1847. A gang of navvies, a good share of them Irish, is shifting clay by hand. In London, railway shares have been falling for months, and the men who bought them on a 10% deposit are being called for the rest of the money they don't have. Fortunes are disappearing.
The gang keeps digging. The line they're working on opens the following year.
In Issue #1, I mentioned in passing that the railway boom crashed but "the track stayed". It's a neat line, and neat lines deserve to be tested. So this issue asks the question properly: is the AI data centre boom a bubble, and if it is, what happens to the people building it?
The 1840s: the mania, the crash, the track
Railway Mania was the AI of its day: a genuinely transformative technology, and a stampede of money chasing it. In 1846 alone, Parliament passed 263 Acts for new railway companies, authorising around 9,500 miles of line. Shares could be bought for a 10% deposit, which made it easy for almost anyone to get in.
Then the Bank of England raised interest rates, money moved to safer places and, by 1847, the market collapsed. By one historian's estimate, railway investment that year was running at almost 8% of British GDP. Investors lost heavily, and around a third of the lines authorised in 1846 were never built.
But two-thirds were. Projects authorised between 1844 and 1846 produced roughly 6,220 miles of railway, including trunk routes that still run today. By most accounts, the busiest year for opening new lines was 1848, after the crash. The money stopped coming in, but the work already committed got finished. For the 200,000 or so navvies on the books at the peak, the crash meant fewer new starts. It didn't mean the work vanished overnight.
The 1990s: the fibre nobody used (yet)
The more recent echo is the telecoms boom. Between the late 1990s and 2001, telecoms companies put more than $500bn, mostly borrowed, into laying fibre-optic cable. Demand didn't arrive on schedule. By early 2002, estimates suggested that around 95% of the fibre laid in the US was "dark", carrying no traffic at all. Big names went bankrupt, and bondholders got back around 20 cents on the dollar.
The cable stayed in the ground. Over the following decade it was lit up, bought cheaply and used for broadband, video streaming and, eventually, the cloud. Some of the data centres being built today sit on the back of fibre that was once written off as waste.
The pattern is the same in both cases. The financial losses were real, and they mostly fell on investors. The physical asset lasted, and it ended up being used.
The case that this is a bubble
It's worth taking the worry seriously, because some serious people hold it.
The largest technology companies will spend more than $600bn on infrastructure this year, and some trackers now put the figure closer to $700bn. As far back as 2024, Sequoia Capital partner David Cahn called this "AI's $600B question": the annual revenue the industry would need to generate to justify the spending on chips and data centres. That gap hasn't obviously closed.
Some of the people building AI have said as much. In August 2025, OpenAI's Sam Altman told reporters: "Are we in a phase where investors as a whole are overexcited about AI? My opinion is yes." A few weeks later, Jeff Bezos called it "a kind of industrial bubble", and added that in the middle of the excitement investors "have a hard time... distinguishing between the good ideas and the bad ideas."
The construction side has its own warning signs. In May 2026, Goldman Sachs analysts estimated that only around 50–60% of the US data centre capacity scheduled for the next one to two years will come online on time. Our own market brief puts it at 30–50% of planned 2026 AI capacity slipping to 2028.
The case that it isn't (or isn't that kind)
The other side is just as serious.
Unlike the dark fibre of 2002, today's capacity isn't sitting idle waiting for customers. Microsoft has disclosed an $80bn backlog of cloud orders it can't fulfil because it can't get power. Nvidia's Jensen Huang responded to the bubble talk last November: "From our vantage point we see something very different." And the buyers this time are a handful of companies with some of the strongest balance sheets in the world, not debt-funded start-ups.
Even Bezos's "bubble" came with a twist. He said industrial bubbles are different from financial ones, and can even be good, because "when the dust settles and you see who are the winners, societies benefit". That's the railway argument, made by someone who lived through the fibre one.
The honest answer is that nobody knows yet. But for a contractor, "is it a bubble?" might be the wrong question.
What the history actually tells a builder
Cahn made a point in that 2024 piece that matters more to us than to investors. GPUs depreciate fast, because each new generation of chips makes the last one cheaper. Physical infrastructure, he noted, "does not follow any 'Moore's Law' type curve."
Put simply, the chips are the speculative part; the building is the track. A powered, cooled and connected hall with a secured grid connection is a long-life asset whoever ends up owning it. If a correction comes, it hits valuations and chip orders first, and the halls are still needed.
The market data points the same way. The market is splitting, not slowing. Tier-1 markets look undersupplied through 2027, while speculative Tier-2 and Tier-3 builds face oversupply risk. And 2026 has so far been a year of delays, not cancellations, which is exactly what 1848 looked like.
So the useful question isn't "is it a bubble?" It's "which of my projects would survive one?" There are three tests worth applying to any project:
- Is the grid connection secured? It's not enough to have applied for it; it has to be signed. Power is the scarcest thing in this industry, and a site that has it keeps its value.
- Is there an anchor tenant? A hyperscaler or major operator committed to the capacity is very different from a developer building on speculation.
- Has the financing closed? Equity and debt in place, not "in discussion". In 1847, the schemes that collapsed were the ones still calling on shareholders for money.
A project that passes all three can slip, but it is very unlikely to disappear.
What this means on site
- Plan for the programme to stretch. If a third or more of planned capacity is sliding to the right, assume your programme will too. Get extensions and the cost of keeping a team on site into the conversation now, rather than when the delay arrives.
- A slipped project is a retention problem. When a programme stretches by six months, your best supervisors and commissioning leads get phone calls from other sites. Work out now who you can't afford to lose, and what it will take to keep them through a longer build.
- Resequencing is a skill in itself. When the power arrives late or the chips come early, the order of works changes. People who can re-plan a build mid-flight, and still hand over a hall that passes integrated systems testing, are worth more in a slowdown than in a boom.
Three numbers worth knowing
6,220 miles: the railway that was actually built from the schemes authorised in 1844–46, despite the crash.
~95%: the share of US fibre estimated to be unused in early 2002. Much of it was eventually lit.
50–60%: the share of US data centre capacity due in the next one to two years that Goldman Sachs analysts expect to arrive on time.
So, is this the railways of the 1840s?
Possibly, and that may not be the warning it sounds like.
If there's a correction, investors will feel it first, speculative projects will feel it next, and well-founded projects will feel it as delay rather than disappearance. The halls with power, a tenant and money behind them will get built, and the people who know how to build, commission and run them will keep their value either way. They did in 1848 and in 2002.
One question back to you: has one of your projects slipped this year, and what did you do with the crew while you waited? Reply and tell me. I read every answer, and the best ones will shape a future issue.
Alex The DC Build Network