AI Could Wipe Out 98% Of Big Tech's Income | Ed Zitron, Gary Marcus & Julien Garran

A 90-minute panel on George Noble’s Market Talk, recorded on 2 October 2026 and released three days later, also as an episode of his audio podcast. Noble, a former Fidelity fund manager, hosts with Jack Gamble of Nobody Special Finance and Alessandro of Risk Takers. The guests are Ed Zitron, Gary Marcus and Julien Garran of MacroStrategy Partnership. All six think AI is a bubble, and the episode is them agreeing from different desks.

Marcus takes safety. “Rogue AI”, he says, shifts the blame onto the product: the agents did what badly built software does, and the fix is a recall of agents until someone can build a sandbox, not a ban on a superintelligence that doesn’t exist. Language models imitate how instructions get answered without grasping what they mean, so they can’t be counted on to follow one. He thinks extinction talk uses up the attention that cybercrime and attacks on infrastructure deserve.

Garran takes the economics. Big Tech, on his numbers, has put $1.7 trillion into AI and plans about a trillion a year more, and even on a generous depreciation schedule that would eat 98% of its non-AI income by 2034. To earn it back the companies would have to rebuild their whole franchise, and he argues language models offer none of what made it: no network effect, costs that rise with scale, nothing that stops a customer switching, no brand. No moat, so no killer app and no payback.

Zitron takes the books. He trusts nothing from Anthropic or OpenAI that isn’t audited, sets Anthropic’s 2025 loss beside its run-rate claims, and puts the two labs’ compute commitments at $1.3 trillion. His newest claim is that the chips aren’t running: Microsoft, he says, has at least $80 billion of GPUs it can’t plug in, and the industry $200 to $350 billion. Unlike the dark fibre of 2000, an idle GPU still needs a building, power and cooling, so there is no cheap second life. Gamble adds the financing, from Amazon’s $8 billion chip sale-leaseback to insurers spreading the risk on chip-backed loans, and says the losses land in pensions and retirement accounts. Noble and Garran end on a credit downturn, a weaker dollar, and gold.

Garran’s 98% and his spending totals come from subscription research and could not be checked. His older “17 times the dot-com bubble” measures capital misallocated across the whole economy, housing included, not AI spending. Zitron’s chip totals rest on one unnamed source and his own arithmetic; the part others have reported is Bloomberg’s, that 2 of Microsoft’s claimed 12 gigawatts is AI capacity.

Smaller things are wrong, and wrong in the panel’s favour. COVID killed between 0.1% and 0.4% of people, not 1%. The heat channel between air-gapped machines carries 1 to 8 bits an hour, not a second. 404 Media found Meta routing Muse’s phone calls to a call centre in an internal test before launch, not in the product. The study Garran cites found no rise in how much apps were used, and did not measure revenue. There are about 47 million developers, not 5 million. On the Navier–Stokes row, Tristan Buckmaster has said he is “not claiming” OpenAI took his approach; the demand that he drop his Anthropic co-author is his allegation, which OpenAI’s Sébastien Bubeck denies; and OpenAI’s terms leave ownership of what a user types with the user.