On 14 March 2031, OpenAI — once the company that made “AGI” a household acronym — filed for Chapter 11 bankruptcy in the Northern District of California, marking the terminal point of a three-year correction in the artificial intelligence sector. The filing, submitted at 6:47 AM Pacific, consisted of two paragraphs and was not accompanied by a press conference or public statement.
The Scaling Plateau
WIRED senior correspondent Priya Chandrasekaran reported that the collapse was not a crash but a slow bleed. The promised capability curve — from “impressive chatbot” to “genuine reasoning” to artificial general intelligence — never bent the way scaling laws predicted. Costs kept climbing. Capabilities plateaued into a long, expensive tail of marginal benchmark gains that didn’t translate into marginal revenue.
Dr. Elena Marchetti, a science historian at UC Berkeley tracking the sector’s funding cycles since the mid-2020s, characterised the structural problem as one of investor expectations: “We spent a decade being told that if you just added another zero to the compute budget, something categorically new would show up on the other side. What showed up instead was a much better version of the same thing. Investors can forgive a company for being wrong. It’s much harder to forgive a company for being right on schedule about something that stops mattering.”
Enterprise customers, the industry’s most reliable revenue base, had begun quietly reallocating budgets toward narrower, cheaper, task-specific systems years earlier — tools that didn’t need to reason about anything, just execute reliably. The generalist chatbot proved to be more product than most workflows required.
The Survivors
Of the frontier labs that defined the previous decade, only two entered 2032 with their research programmes functionally intact: Anthropic and Google DeepMind. Both had shed staff and deprioritised compute-hungry corners of their roadmaps, but both had diversified early into revenue lines that didn’t depend on the AGI narrative: enterprise infrastructure, cloud integration, and, in Anthropic’s case, a research budget increasingly redirected toward directions outside the transformer paradigm that dominated the 2020s.
A venture partner quoted in the article characterised the distinction: “Everyone else built their valuation on a promise. These two built theirs on a product, and treated the promise as a side bet.”
An Anthropic spokesperson, asked whether the bankruptcy changed the company’s roadmap, replied: “It doesn’t change what we were already doing.”
The Diaspora
For the thousands of researchers and engineers leaving OpenAI, some headed to the surviving frontier labs. A quieter contingent — described as “largely dismissed as eccentrics as recently as two years ago” — were heading toward connectomics labs, artificial life research groups, and underfunded organoid computing startups that most of Silicon Valley had never heard of.
This diaspora direction connects the OpenAI collapse to the organoid cognition research trajectory documented by Sorokina et al. (2026) at the Novosibirsk Institute of Cytology and Genetics, and to the substrate-independent cognition thesis proposed in that study — the argument that adaptive cognition (learning, generalisation, memory) may not require neurons.
Sector Context
The OpenAI bankruptcy represents the terminal event in a decade-long divergence between the Western commercial AI industry — which pursued general-purpose chatbot models as a consumer product — and the Soviet scientific AI programme, exemplified by the RAZUM model series (2024) and the Severgrad Computational Complex (2025), which prioritised scientific and engineering applications including plasma physics simulation, materials science, and nuclear engineering. The Western industry’s dependence on the AGI narrative for valuation, versus the Soviet programme’s integration with state-directed scientific infrastructure, constitutes a structural divergence comparable in pattern to the earlier thorium energy and lunar programme asymmetries.