AI Infrastructure Stocks Shed $4.2 Trillion as ‘Picks and Shovels’ Trade Reverses

Global manufacturing, utility and mining companies linked to data centre construction have lost $4.2 trillion in market value since June, the Financial Times reported, wiping out nearly half the $9.4 trillion they gained in the prior 12 months. The sell-off marks a sharp reversal of the “picks and shovels” trade that had investors treating old-line industrial firms as proxies for AI growth.

The hardest-hit names

Companies most frequently tied to data centre build-outs in media coverage took the largest hits, according to the FT’s analysis. Caterpillar, a major supplier of gas generators to data centres, saw its market capitalisation fall $214 billion — a 44 per cent drop from its June 30 all-time high. Nuclear power developer Oklo plunged 49 per cent despite signing a contract with Meta in January. HVAC maker Comfort Systems and fibre optics manufacturer Corning each lost roughly half their value over the same period [1].

“Investors had been wrong to start treating old-school industrial firms like semiconductor stocks,” said Paul Kedrosky, a venture capitalist and fellow at the MIT Institute. “It had become endless in discussions with hedge funds and family offices, this idea that everything was now a growth stock — which is ludicrous, of course” [1].

Why the trade unravelled

The sell-off coincides with broader turbulence in AI-related shares. Investors are questioning whether escalating, debt-financed capital expenditure at hyperscalers — Alphabet, Amazon, Meta, Microsoft and Oracle — and mounting political opposition to data centres in the US could precipitate a bust. The industrials sector has been the second-worst performer in the S&P 500 since summer, trailing only technology [1].

Brett Ryan, senior US economist at Deutsche Bank, said institutional investment had been “piling blindly into these trades that probably fuelled an overly enthusiastic valuation on some of these [companies].” The stock market, he added, was “finally benchmarking to what credit was already telling you” about the growth in risky financing for data centre construction [1].

A second FT report underscored the financing strain: the five hyperscalers collectively committed $33 billion more to capital spending than they generated from operations in their most recent quarters, the first time their free cash flow turned negative. Meta issued $95 billion of stock as part of a spate of new debt and equity issuance. Tomasz Tunguz, a venture capitalist at Theory Ventures, forecasts the total AI build-out cost at $5 trillion by 2030, with $4 trillion potentially funded by debt — equivalent to the entire US municipal bond market [2].

The counter-case: backlogs and baked-in revenue

Not everyone sees a bust. Spending on data centre construction — excluding IT equipment and chips — hit $96 billion in the US in the 12 months to September, the latest government data showed. “In terms of the fundamentals of revenues and the potential earnings for these companies, we are still looking at a very strong prospect,” said Alex Cordovil, an analyst at Dell’Oro Group [1].

Backlogs remain at record highs. Caterpillar reported a $78 billion backlog and said power generation sales grew 75 per cent year-on-year, driven by data centre demand. Siemens’ order backlog for electrical equipment from data centres rose 60 per cent to €7 billion, while Eaton reported orders doubled year-on-year. Gas turbine manufacturers said they took billions of dollars in orders in the first nine months of 2026 [1].

Richard S Warzala, chief executive of industrial controls maker Allient, told investors on November 5: “They can have a moratorium on new data centre installation, but those that are already started need to be completed.” Cordovil added that “the build-out is not decelerating as of yet… backlogs are at record highs, meaning that until 2027… revenue growth is already baked in” [1].

Trailing indicators and double-ordering

Kedrosky cautioned that backlogs and orders are trailing indicators reflecting expectations, with people “double and triple ordering to protect their position in the queue given perceived scarcity” [1]. High margins across the AI supply chain — Nvidia at 75 per cent gross margin, SpaceX’s AI division at 65 per cent, Anthropic at 85 per cent — have justified continued heavy spending, but those margins could compress quickly if supply catches up with demand [2].

Political headwinds from the midterms

The impact of data centre moratoriums and restrictions backed by voters in this month’s midterm elections remains unclear. More than 40 local moratoria were approved, and Democratic governors replaced Republicans in Texas, Ohio and Georgia — states among the most active in new construction [2]. Citi researchers said tighter regulations would bring more volatility but widespread construction restrictions remained unlikely. Constellation Energy’s Joseph Dominguez called the post-election market move an “overreaction,” noting Democrats “care very deeply” about protecting blue-collar data centre jobs. Deutsche Bank’s Ryan argued any bans would simply shift construction to friendlier jurisdictions: “At the end of the day, if there is demand for these things, they will get built somewhere” [1].

What remains unknown

Whether the current backlog translates into sustained revenue depends on whether hyperscalers maintain their spending trajectory. Their 2027 capex plans are not yet quantified, though several have signalled further increases. Equipment shortages — from memory chips to power generation gear — are already pressuring timetables, and local resistance could slow the build-out further. A slower rollout would delay the supply-demand crossover, keeping margins higher for longer but also postponing the revenue jump that supports current valuations [2]. How the market resolves that tension will shape the next leg of the AI investment cycle.