Short seller John Hempton calls AI compute 'fibre optics Mark II', but says he'll buy Anthropic's IPO
The interview: John HemptonDealroom has a profile for this one. Try Dealroom →, founder and chief investment officer of Sydney hedge fund Bronte CapitalDealroom has a profile for this one. Try Dealroom → and Australia's best-known short seller, told Joe Walker on The Joe Walker Podcast (recorded 1 November) that AI is "the most amazing technology of your lifetime" but that building the compute behind it is likely to be a terrible business. His label for it: "fibre optics Mark II", a technology that changes the world while bankrupting many of the companies that build it.
The worst business in the world: Hempton's formula is high capital costs, low marginal costs, competition and continual technical improvement. "The moment that you have excess supply, this just turns into the worst business in the world," he said, because prices fall to marginal cost and the borrowed capital can't be repaid. He points to solar panels, which fell from about $5 a watt to about 15 cents in some 15 years, and to the late-1990s fibre build-out. Data centres are the most extreme case, he argues: they sell "just the ability to do a mathematical calculation", which makes compute "the ultimate commodity".
The maths: Hempton calls the roughly $7 trillion of compute capex forecast to 2030 "clearly wrong". To earn a return it would need about $2 trillion of AI revenue a year, which he says can only come from businesses replacing workers. Because companies don't fire people for a 10% saving, that implies around $4 trillion of annual wage savings, about 10% of the OECD's total wage bill of roughly $40 trillion. He considers that neither practically nor politically feasible in the next few years.
The Anthropic surprise: Reading Anthropic's IPO prospectus "shocked the hell out of me," Hempton said. On his reading of the numbers, Anthropic built what he calls clearly the best model for "lowish single-digit billions of dollars", making Claude about five times more efficient than Google's and OpenAI's models. The prospectus does not make that claim directly. For the neoclouds and compute suppliers, he said, that makes the capex look "horrible beyond belief". For Anthropic it is "incredibly bullish", because rivals' data centres could end up "for sale from the liquidators for pennies on the dollar". Describing himself as "the most bearish person on AI economics" he knows, he said he still plans to buy some Anthropic shares in the IPO.
Advice for Australia: He told the government to "happily facilitate" data centres, steer them to regional areas rather than the suburbs, require them to generate most of their own power so they don't push up household electricity prices, and not subsidise the world's compute the way China subsidised the world's solar panels.
The Nvidia angle: Sharing the interview on X, commentator Arnaud Bertrand argued that Hempton's maths assumes compute stays expensive, which depends on Nvidia keeping gross margins of about 75%; cheaper chips would shrink the capex bill, the revenue needed and the job losses. That is Bertrand's extension, made after Jensen Huang's resignation, not Hempton's argument.
Sources: The Joe Walker Podcast, episode page and transcript, 4 November
YouTube: Why Data Centres Will Change the World and Bankrupt Their Builders
Arnaud Bertrand on X, 13 November