Rails and Racks: Two Booms, Opposite Shortages

Key Takeaways

• The railway boom ran short of capital. The AI boom is short of labour and metal.
• Hard money, and later deliberate monetary contraction, made the railway busts deflationary and long.
• Plentiful money with physical bottlenecks points the other way, towards inflation.
• Railway track lasted a century. AI chips are depreciated over five to six years.
• Owners of scarce physical capacity have historically fared better than boom investors.

Two booms, opposite shortages

There are constant comparisons right now about the similarities between the rail boom between 1845 till around 1893 (there were several booms and busts over the 50 year period). But these two booms are polar opposites in what inputs created the shortages: railroads needed scarce capital, held down first by the gold standard and later by deliberate monetary contraction, vs labour and copper being the scarce resource with unlimited capital, suggests very different outcomes.

In the AI era, capital is plentiful. Central banks can create it, and tech giants are spending hundreds of billions a year. The wall is made of electricians, linesmen, transformers, copper and steel. That could change what a bust looks like, and who pays for it.

Railroads: scarce capital, plentiful labour

The railroad mania started in Europe around 1845, railways were built as independent privately owned assets. Shareholders of these railways were asked to invest, with peak investment soaking up around 7% of GDP (Odlyzko, University of Minnesota). Deposits (of gold) drained from banks raising interest rates and then causing a run on banks, causing a cascade of bank failures during the week of terror in October 1847 (New York Fed Crisis Chronicles).

Much like current runs on banks, it was the intervention of the government allowing the Bank of England to act as the lender of last resort, thereby increasing liquidity, that restored confidence. The Bank Charter Act 1844 was suspended on 25 October 1847 and the panic ended almost immediately (Panic of 1847).

America repeated the pattern for half a century, though the money was not always metal. America's biggest railroad bust, the 1873 panic that took down Jay Cooke & Company along with 89 railroads and some 18,000 businesses, happened in the greenback era, when the dollar was inconvertible paper from 1862 until 1879 (Congressional Research Service). Capital stayed scarce anyway because policy kept it scarce: Grant was contracting the money supply and the Coinage Act 1873 demonetised silver.

Railroads were often the single biggest destination for investment, and much of the money came from British and European bondholders because American savings could not keep up: railways made up 61.6% of all British investment in the US (University of Pennsylvania, Harris). Labour was the easy part. Britain's railways were dug by navvies, around 200,000 at the peak, many from rural Ireland and Scotland. America's transcontinental line was built largely by Irish immigrants and Civil War veterans in the east, and thousands of Chinese workers in the west, more than 13,000 of them on the Central Pacific alone (Gilder Lehrman Institute). Workers were cheap and could be found. Money could not.

AI: plentiful capital, scarce labour

Unlike the railroads, AI currently has no capital shortage. In 2025 the five largest hyperscalers, Meta and Microsoft among them, spent about US$443 billion in capital expenditure (Introl, January 2026), and four of them are guiding to roughly US$700 billion in 2026 (Motley Fool, 27 February 2026). The labs are building directly too: Anthropic committed US$50 billion to its own US data centres in November 2025 (TechCrunch, 12 November 2025). Much of it comes from their own cash flows, and increasingly from bonds and private credit. Behind them sit central banks that have shown since 2008 that they will create money rather than let markets collapse.

The bottleneck is people with specific skills. Turner & Townsend's 2026 survey of 112 construction markets found over 70% reporting data centre contractor capacity as tightening or overstretched, with labour shortages the leading source of cost escalation in 71% of markets, and Australia and New Zealand recording the largest shortages of any region (Turner & Townsend, 8 July 2026). One industry estimate puts the US data centre workforce needed for 2026 growth at around 650,000 (DCGeeks).

These are not workers you can find on a dock. Electricians, high-voltage linesmen, commissioning engineers and specialist fitters take years to train. Data centre projects are pulling them away from homes, factories and hospitals. Outside data centres the rest of the building market is shrinking: US private nonresidential construction spending excluding data centres was about 6% lower in August 2026 than a year earlier (US Census Bureau, August 2026), with housing weakening as rates rise. In Australia the cash rate sits at 4.60% after the RBA's 29 September 2026 increase, and home values have fallen in most capital cities (RBA cash rate, effective 30 September 2026).

Why an AI bust could look different

When the AI boom breaks, the first hit could look familiar. Overleveraged developers and neoclouds could default, lenders pull back, share prices fall, and investors sell whatever they can for cash. UNCTAD warns that an AI valuation correction could trigger a broader sell-off and margin calls, with bankruptcies among leveraged AI-linked firms seizing up credit markets (UNCTAD, Trade and Development Report 2026). The immediate effect would historically be deflation, just like 2008, and with current property markets weakening this may have already begun.

Central banks would be under pressure to respond by lowering interest rates to create liquidity in a failing system. Governments who also usually respond with increased spending may be slower in this cycle, with years of enormous deficits and voter resistance to more spending holding them back. That would mean monetary policy has to do the heavy lifting this time.

The scarcities change what happens next. Skilled trades would likely still be short, because the grid, housing and energy transition all need the same electricians. Copper could still be short, because mines cannot be conjured up: S&P Global projects supply falling 10 million tonnes short of demand by 2040 (S&P Global, 8 January 2026). So a rescue pours money into an economy whose physical bottlenecks have not gone away. That could be a recipe for inflation, not the long deflation of the 1870s.

The other side of the argument is that there is no bust to rescue. JP Morgan estimates 2026 data centre funding needs of roughly US$700 billion could be met entirely from hyperscaler cash flow and high grade bond markets, and puts data centre vacancy at a record low 1.6%, which argues the market is underbuilt rather than overbuilt (JP Morgan, Outlook 2026). On that reading the capital never breaks.

The labour market may split either way. White-collar workers face falling demand as AI does more of their work. Tradespeople and anyone who controls scarce physical capacity keep their pricing power. The railroad era made workers cheap and money dear. The AI era makes money cheap, office work cheaper, and skilled hands and metal dear.

A century of track, a few years of chips at best

One other comparison that needs to be highlighted is that railways lasted, in fact they are still in use. Roadbeds, cuttings, bridges and tunnels dug in the 1800s still carry trains today. When a railway company went broke, the line was sold to new owners and kept working. In 1893, companies controlling a quarter to a third of all US mileage went into receivership and the track kept carrying freight throughout (Panic of 1893). The investors lost, but the country kept the asset.

AI capital depreciates and quickly. Tech companies depreciate their servers over five to six years, and many analysts argue the frontier life of top AI chips is shorter still, with Michael Burry putting it at two to three years (Level Headed Investing). AI data centres are also built so specifically, for power density and cooling, that they cannot easily be turned to other uses.

Bankrupt AI infrastructure leaves behind chips that are steadily losing value and specialised buildings with no obvious next use. Much of the misallocated capital would simply depreciate away, causing again deflation on the capital input.

This is where private credit turns a boom into a dangerous bubble. Some AI cloud companies have borrowed using GPUs as collateral, and large data centre projects are increasingly funded through private credit and off-balance-sheet vehicles. CoreWeave secured an US$8.5 billion facility in March 2026 largely against a single customer contract (Data Center Dynamics, 31 March 2026). Fast-depreciating collateral, opaque lending and a compressed timeline: the echo comparison is 2008 when mortgage securities were the problem, but that land and asset was never going to fully depreciate. AI could, with obsolete chips and infrastructure.

What it means

The railroad era teaches that investors in the boom often lose, while the owners of whatever the boom depends on win. That held for landowners near stations and for the steelmakers. It is likely to hold again for the owners of energy, well-located land, copper and skilled labour.

The difference is the money. Capital that stayed scarce, whether by gold or by policy, made the railroad busts deflationary. Plentiful capital and scarce labour and metal make the AI bust more likely to end in government-led inflation set against supercharged deflation.

In that world, the assets that have historically held value are the scarce ones: urban land, resources and energy, and metals that cannot be created on a balance sheet. Physical gold and silver through Ainslie Bullion sit in that last category.

_______________________________________________

This article is general information only and does not constitute financial advice. Past performance is not indicative of future results. Always conduct your own research or consult a licensed financial adviser before making investment decisions.