Scaling Isn't Intelligence: 88-Hour Navier-Stokes vs. 4% of $2B

Frontier-lab roadmaps today rest on one unstated premise: more compute and data produce intelligence, and intelligence produces capability. This commentary — circulating the same week OpenAI claimed an 88-hour Navier-Stokes solve and Anthropic reported Claude writes 80%+ of merged code — attacks that premise head-on. The distinction it draws between intelligence and power matters for how ML teams set research priorities, interpret benchmark results, and pitch the scaling story to investors and regulators.

The most consequential argument in artificial intelligence right now is not whether scaling works — it is what scaling actually buys. A syndicated opinion piece published September 24, 2026 by Newsday and News India Times takes aim at the industry's most load-bearing unspoken assumption: that intelligence, once sufficient, is all you need. The essay builds its case around OpenAI's reported announcement that it had solved the Navier-Stokes Millennium Prize problem in 88 hours, roughly three years after its models were only 'good at grade-school math.' Project that curve forward, the piece notes, and you get Sam Altman's prediction that AI will cure all diseases on one side and the fear that a rogue AI will destroy humanity on the other. The commentary's claim is that this curve is real but drawn on the wrong axis.

“

Historian Alex Wellerstein calculated that Los Alamos — home to Niels Bohr, Richard Feynman and Robert Oppenheimer — accounted for just 4% of the Manhattan Project's nearly $2 billion cost.

The heart of the argument is a category confusion between intelligence and power. Santa Fe Institute scientists Melanie Mitchell and David Krakauer are cited for summarizing the field's mantra as 'Scale is all you need' — more training data, more compute, better models. Altman's exponential-growth vision and Anthropic CEO Dario Amodei's hope for 'a country of geniuses in a datacenter' both assume that sufficient reasoning capacity converts directly into the capacity to reshape the physical world. The essay's rebuttal is concrete rather than philosophical: the United States already ran the 'geniuses in a datacenter' experiment at Los Alamos, and it did not prove what the scaling optimists think it did.

Historian Alex Wellerstein calculated that Los Alamos — home to Niels Bohr, Richard Feynman and Robert Oppenheimer — accounted for just 4% of the Manhattan Project's nearly $2 billion cost. Eighty percent of the budget went to the industrial plants in Tennessee and Washington that produced uranium and plutonium. Oppenheimer's 1945 congressional testimony drives the point home: without scientists there would have been no bomb, but 'if there had been only scientists, there would have been no bomb.' Genius was necessary; the physical, industrial and organizational infrastructure was what proved sufficient to turn ideas into a world-changing artifact.

The implications for the AI industry are substantial. The scaling narrative currently underwrites hundreds of billions of dollars in compute investment, semiconductor demand, energy procurement and frontier-lab valuations, while simultaneously driving both accelerationist policy hopes and existential-safety regulation. If intelligence is not the same as capability, then the utopian and apocalyptic endpoints are both overstated. A model that solves a Millennium Prize problem has demonstrated benchmark mastery, not the ability to mine uranium, build power plants, fabricate chips, or operate the institutions that convert model output into outcomes. The real bottlenecks remain physical and social: energy supply, advanced manufacturing, rare materials, capital allocation, and human organizations able to deploy models at scale.