Those who worry that AI may create an all-knowing, conscious (or at least conscious-seeming) digital entity which (who?) becomes either vengeful, or perhaps simply finds no practical reason to allow humans to continue taking up valuable compute resources are labeled “AI Doomers.”
Doomer debates typically range from a flashy, singular cataclysmic moment where the machines finally take over and wipe out everything, to the more sinister and subtle long-form sabotage where the AI stays secretive for a long time, slowly suffocating our civilization through one of several possible means.
The possibilities make for excellent fodder for science fiction and anxiety. But what about a form of existential threat that is less The Matrix and more Margin Call?
Interestingly, it is entirely plausible that a transformer-based AI is deployed as a research accelerant that helps discover the breakthrough in computing or cognitive architecture that renders itself obsolete. What could a post-transformer AI paradigm look like? Neuromorphic computing, optical processors, quantum-classical hybrids, biological computing, or something not yet named. The specific successor is unknowable — that's the point. But what seems consistent is the pattern: every architecture that has ever seemed permanent was temporary.
If a new paradigm emerges that achieves superior cognitive capabilities with fundamentally different hardware requirements, the result is not a gentle transition. It is potentially trillions of dollars in stranded assets — data centers, GPU inventories, power infrastructure, cooling systems — purpose-built for a paradigm that no longer leads.
AI’s amazingly fast proliferation across the globe is seen as a triumph of the Internet-connected world. And it is. At no time in history would a new technology have been able to reach billions of people so quickly. And also, perhaps at no other time would civilization-scale capital investment be made as quickly either. Perhaps that’s another triumph, perhaps not.
The AI race today has many of the largest, most valuable companies in the world spending unprecedented amounts of money on data centers. JPMorgan estimates that the data-center build out for AI could total $5 to $7 trillion over the next few years, more than even Microsoft, Google, Amazon, Facebook, and Oracle collectively are able to fund without taking on debt.
That debt is being underwritten with long time horizons and an underlying thesis that assumes that the transformer architecture and its descendants represent the durable path to machine intelligence from which the profits are just about inevitable. The transformer concept itself is less than a decade old.
A leapfrog at this scale is not simply an investor problem. The capital is being deployed by pension funds, sovereign wealth funds, and major banks whose losses socialize to the broader economy. Occupy Wall Street was the social response to financiers destroying economic value through reckless speculation on mortgage-backed securities. The AI version could be worse: at least housing serves a genuine need. Stranded data centers full of obsolete chips don’t. They are monuments to a paradigm that moved on. Of course, the buildings and power connections could be repurposed, but they represent roughly a quarter of the investment. The other three-quarters is silicon and hardware purpose-built for transformer-scale parallelism. The political backlash from "Big Tech spent trillions on AI and left us with empty warehouses" could make the post-2008 techlash look quaint.
This is not the "existential risk" that AI Doomers warn about. It is something stranger: AI's greatest success (a genuine scientific breakthrough) becomes the AI industry's greatest threat (paradigm obsolescence). The better AI works at what it's best at, the faster it may undermine the infrastructure built to support it.
Postscript for the conspiracy-minded
Assume there is a eureka moment — a fundamental paradigm shift is discovered. Conventional wisdom in modern capital markets says that new ideas are quickly rushed to market. Breakthroughs get patented, funded, and scaled as fast as the discoverer can find a term sheet.
But given what could be at stake in this case, consider two scenarios in which the rational move is not to announce.
Internal discovery. Suppose the breakthrough happens inside a research lab funded by, or adjacent to, the very companies that have sunk trillions into the current paradigm. The researchers understand what they've found. So do the executives. But disclosure would vaporize the value of infrastructure that their investors, their lenders, and their governments have committed to at historic scale. The incentive is not to suppress the discovery forever — that's impossible — but to delay it long enough to reposition. Sell assets quietly. Restructure debt. Let the next earnings cycle pass. The history of industries sitting on disruptive discoveries to protect sunk costs is long and well-documented.
Adversarial discovery. Now suppose the breakthrough happens on the other side of a geopolitical divide (very plausible given China has roughly three times as many active AI researchers as the United States). If one of the West's adversaries achieved such a breakthrough, you might assume they'd rush to deploy it and declare victory. But perhaps the shrewder play is the opposite: stay quiet and let the token-factory buildout continue. Every additional trillion the West commits to a soon-to-be-obsolete paradigm is a trillion that cannot be redirected. The longer the silence, the larger the stranded asset base, and the more devastating the eventual revelation.
Neither scenario requires a formal conspiracy. Both require only that rational actors, faced with asymmetric information and enormous stakes, do what rational actors have always done: act in their own interest. Thus, it’s feasible that the discovery has already happened.