By mid-1999 it was already clear to veteran investors and students of economic history that the dotcom bubble had reached parabolic insanity.
The speculative momentum was still unstoppable – and would run a lot further – but grown-ups knew by then that few of the high-flying start-ups were ever going to generate a viable revenue stream. The authentic success stories would have to fight each other in a cannibalistic struggle for survival.
The physical constraint is rock hard. “Our grid in the United States hasn’t had any meaningful upgrade since the 1970s,” said Bobby Majumder from the industrial law firm FBT Gibbons.
The threat to AI stock mania is not so much lack of energy – though that is serious – but rather the global bottleneck of transformers, substations, switchgear, transmission lines and all the unsexy stuff we rarely think about, leaving aside the acute shortage of skilled workers in the US able to install and run such kit.
A single big campus in the data centre hub of Hays County, Texas – an area where I once played a lot of golf (misspent youth) and know well – can use 10 million gallons of water a day for evaporative cooling and power generation, draining the Edwards Aquifer that also supplies the Austin-San Antonio corridor.
“Nobody is talking about cooling; nobody is talking about water,” said Majumder, speaking at the recent Marshall & Stevens forum on energy infrastructure. “The farmers are not going to be happy at all about you pumping down their aquifer for cooling.”
There are other obvious catalysts that could puncture the bubble. Stubborn US inflation – input prices are rising at the fastest pace in four years – may force the Federal Reserve to stop its “stealth-QE” via bill purchases. The bond markets may hold Kevin Warsh’s feet to the fire as he takes over the institution.
Inflation may stop Scott Bessent, the poacher turned gamekeeper now running the US treasury like a hedge fund, from using the $8tn money market to help soak up massive fiscal deficits at the peak of the economic cycle.
Cheaper “commoditised” AI from the likes of DeepSeek in China may start to undercut American rivals, threatening the implicit pricing model behind today’s equity valuations. If it is true that DeepSeek v4 can achieve 80pc-90pc of the performance of Anthropic’s Claude at 10pc of the cost, you start to see the problem.
Liaquat Ahamed, author of the wonderful Lords of Finance covering the Great Depression and now releasing his new book 1873, likens the AI boom to the American railway mania after the Civil War. Routes were duplicated in the rush for dominance.
Costly lines passed through sparsely inhabited regions where there would never be enough human traffic in time to justify the scale of debt issuance. [...]
Hyperscalers can try to leapfrog the grid bottleneck by building their own power plants, but that will not solve the problem either, at least not in time to alleviate the burden of fast-mounting and opaque AI debt.
It took 17 years to plan, license and build the recent Vogtle nuclear plant in Georgia. Costs ballooned from $12bn to $30bn. Small modular reactors may be cheaper per gigawatt – don’t hold your breath – but none yet exist in the West, and there will be no serious supply chain until circa 2040.
Shale gas frackers can drill until they drop, but that makes no difference if there are no gas turbines available on the world market. The waiting list for heavy-duty models used in combined-cycle plants has stretched to seven years, although hyperscalers with the deepest pockets are jumping the queue for a fat fee with 2030 delivery dates. [...]
The AI revolution is real. The language models are fabulous. The technology will make economic life almost unrecognisable by mid-century.
But the internet revolution was also real in 1999 before the Nasdaq index dropped 77pc, flushed out the commercial nonsense and overshot in the other direction.
Don’t track Nvidia chip orders if you want to know where the AI market is heading. Track the metaphorical picks and shovels that make it all possible.
by Ambrose Evans-Pritchard, The Telegraph | Read more:
Image: Richard Newstead
[ed. See also: How bad is AI for the environment? (Yale Climate Connections).]
