The Scale, The Plan, and The People — No One's Happy

My previous post documented what happens to people and teams when output is decoupled from understanding. That was the micro picture: what the pattern looks like from inside the room where the work is supposed to happen. The response was larger than I expected. A considerable number of readers wrote in to corroborate the pattern from inside their own organizations. Because of this, I am drafting guidelines on recommended organizational AI use, downstream of those conversations, which will appear separately and soon. If you’d like to contribute please email.

This article takes up the macro question. If the pattern is as visible as the feedback suggests, if every consulting firm keeps utilizing agentic tools and workflows that produce subpar output, if every enterprise keeps buying tools that deliver little to no measurable productivity gain, the question worth asking is structural. Why does a system that produces so many failures of this shape keep being built, sold, and scaled?

It requires considering various facets of the current landscape: what the people closest to the architecture have begun to say about what these systems can and cannot do; the scale of the infrastructure bet being made on their behalf; the financial structure that locks each organization into continuing the spend; why if opinions are becoming more modest, spend is not following.

The age of scaling, in retrospect

The people who built the current generation of AI systems have begun saying, in public, that the approach that got them here will not get them further. Ilya Sutskever, the OpenAI co-founder and former chief scientist whose work in the 2010s established the scaling hypothesis the entire industry runs on, told Dwarkesh Patel in late 2025 that the age of scaling — which he dated with characteristic precision as 2020 through 2025 — is over. “Scaling the pre-training is essentially tapped out,” he said. “If you just multiply the scale by 100x now, you won’t get a qualitative change in capability.” [1] The recipe, on his account, is finished; what comes next requires research. Demis Hassabis, who runs Google DeepMind, put the window to AGI at five to ten years in October 2025, and in the same month publicly corrected an OpenAI researcher who had overclaimed results on a math benchmark [2] — a small, in-room professional discipline that does not happen unless he has stopped believing the field can afford the inflation.

These are not fringe critics. Yann LeCun left Meta in November 2025 to serve as Executive Chairman of Advanced Machine Intelligence Labs, a company built on the premise that the dominant architecture is a dead end. He raised $1.03 billion against that claim in March 2026, and has said publicly that nobody in their right mind will be using large language models of the current type within three to five years. [3] Andrej Karpathy, who helped build OpenAI, said in October 2025 that current agents “just don’t work,” that there is “some over-prediction going on in the industry,” and that anything resembling general intelligence is roughly a decade away. [4]

The frontier labs have begun adjusting their timelines. Dario Amodei at Anthropic predicted powerful AI “as early as 2026” in Machines of Loving Grace, published in October 2024. [5] By Davos in January 2026 he was calling for AGI-level capability “in two years” — meaning 2028. By his Dwarkesh interview the following month he was saying “I don’t believe we’re basically at AGI” and openly acknowledging that if his revenue forecast was off by a year, “there’s no force on earth, there’s no hedge on earth that could stop me from going bankrupt.” [6] That shift may well represent genuine calibration rather than retreat — an honest revision as results come in. But the capital expenditure commitments that now underpin the industry were made against the 2024 timeline, not the 2028 one. The estimates have moved. The money has increased.

The numbers

The San Francisco Federal Reserve’s Economic Letter in February 2026 reviewed the macro literature and rendered the consensus in central-bank register: “While GenAI and related applications are useful, they are not the innovation that spurs broad-based reorganization of the economy.” [7]

Goldman Sachs’ chief economist Jan Hatzius, surveying the same terrain, concluded that AI had contributed “basically zero” to U.S. economic growth in 2025 and observed that “FOMO, not ROI, is driving hyperscaler capex.” [8] This shows simply that the technology is real, its uses are real, and at the scale of the spend, the productivity it returns is not what the spend requires.

The mismatch becomes specific in the revenue math. J.P. Morgan, modeling what the buildout would need to earn to clear a ten percent return on current capex, arrived at roughly six hundred fifty billion dollars per year in AI-sector revenue. The current run-rate, by the most permissive count, is about twenty-five...