Kurzweil Scorecard: The Infrastructure Arrived, the Mechanisms Didn’t

Ray Kurzweil’s 2005 blueprint for mid-century computing rested on four load-bearing claims: transistors would keep halving in price every 1.6 years, DNA would do computation, IBM’s autonomic systems would run themselves, and battlefields would run on self-organizing networks with no central hub. In 2026, every one of those pieces of infrastructure is real — and almost none of the mechanisms Kurzweil named survived. The map was right. The roads are not the ones he drew.

The predictions

This batch pulls ten 2005 claims from across The Singularity Is Near: chapter “Moore’s Law and Beyond” on transistor economics; “Computing with DNA” and “Emulating Biology” on Hao Yan’s self-assembling DNA tiles; “The Criticism from Software” on IBM’s autonomic computing program; “On Warfare” on the Army’s Future Combat Systems network; and “The Criticism from Malthus” and “Going Back in Time” on Seth Lloyd’s cosmic compute ceilings and Todd Brun’s closed-timelike-curve computers. Taken together, they are Kurzweil’s bet on how the substrate of the 2020s would be built.

Where we actually are

The transistor halving that wasn’t. Kurzweil wrote that “average transistor price was halving about every 1.6 years, with one dollar buying about ten million transistors in 2002” (ch. “Moore’s Law and Beyond”). Extrapolated to 2025, that trajectory puts a dollar at roughly 220 billion transistors. Reality: the per-transistor cost curve flattened at the 28-nanometer node around 2015 and has not meaningfully declined since. TSMC announced 5–10 percent price hikes for nodes below 5 nm starting in 2026, with 2 nm wafers pushing past $30,000 apiece. For the first major node transition in the industry’s history, cost per transistor will rise.

But read Kurzweil’s own 2024 restatement in The Singularity Is Nearer: “This rate of improvement continues unabated. It is not fundamentally dependent on Moore’s law, the famous exponential reduction of feature sizes on microchips.” He walks away from the transistor metric entirely and pivots to compute-per-dollar: “some Pentium 4s were at 12 million [operations per second per dollar]… Google Cloud TPU v5e chips are likely around 130 billion operations per second per dollar.” That’s a 10,000× improvement in nineteen years, achieved without another halving in transistor price. The mechanism died; the macro curve didn’t.

DNA tiles became a drug delivery vehicle, not a computer. Kurzweil cited “Duke researchers created self-assembling DNA ‘tiles’ that formed nanogrids… coated with silver to create nanowires” (ch. “Emulating Biology”), and separately argued DNA computing was “limited to single-instruction multiple-data architectures” (ch. “Computing with DNA”). The SIMD ceiling turned out to be real: nobody is shipping a DNA coprocessor. But Hao Yan — the original Duke collaborator — moved to Arizona State and, together with Paul Rothemund’s 2006 origami paper, kicked off a research explosion. Our literature scan finds DNA-origami publication counts rising from four in 2005 to 391 in 2025, with DNA-data-storage papers hitting 112 last year. Eighteen US-granted patents since 2018 cover DNA origami explicitly; Arizona State and Harvard each hold four. US 12,441,996, granted October 2025, describes a “library of origami folded DNA data storage files… the data oligonucleotides are attached to the top surface at a density of less than 100, 80, 50, 40, 20 or 10 data oligonucleotides per 100 nm² of the DNA scaffold.” That is Hao Yan’s nanogrid, repurposed as a write-once molecular filesystem. US 11,708,601 uses the same tiled structures as a nanoscale force sensor based on FRET pairs. In The Singularity Is Nearer, Kurzweil now lists “DNA origami” and “DNA nanorobotics” among the pathways to molecular manufacturing — a quiet correction of his earlier computing framing.

Autonomic computing got rebranded Kubernetes. Kurzweil quoted IBM’s 2001 manifesto about “self-configuring, self-healing, self-optimizing, and self-protecting” systems (ch. “The Criticism from Software”) and predicted the supporting software would run to “tens of millions of lines of code.” IBM’s formally branded autonomic effort faded. What arrived in its place was container orchestration. Kubernetes ships with automatic pod restart, liveness and readiness probes, declarative rollbacks, horizontal autoscaling, and network policy enforcement as default behaviors — the original IBM checklist, implemented at hyperscaler scale. Literature output tagged “autonomic computing self-managing” actually jumped to 328 papers in 2025 after more than a decade of decline, driven by the intersection with AI-operated clusters. On the patent side, US 11,916,721 (VMware, February 2024) claims a method for “network function virtualization self-healing” that ingests “key performance indicator information relating to a virtual network function,” maps it to physical fault notifications, and “automatically perform[s] a remedial action” against a dynamic threshold. The 2001 IBM paper described that system in prose; a 2024 VMware grant describes it in claims.

The software-about-software prediction landed. Kurzweil’s corollary — that the supervising codebase would reach tens of millions of lines — is verified. The cloud-native landscape the CNCF tracks spans more than 150 graduated, incubating, and sandbox projects whose combined codebases easily exceed that threshold. The prediction was correct about magnitude and about direction.

Future Combat Systems died; its network arrived through a commercial satellite. Kurzweil predicted “self-organizing, highly distributed communications networks with no centralized hubs, routing around damage” (ch. “On Warfare”), anchored on the Army’s FCS program. DoD formally canceled FCS on 23 June 2009. The vision did not die with the program. In Ukraine in 2026, Ukrainian-built Bucha and Molniya strike drones launch in mesh-networked flights in which one aircraft trades its warhead for an extra battery to act as a signal repeater. Delta, Ukraine’s situational-awareness layer, is now cited by CSIS as a working model for the US military’s own CJADC2 concept. The institutional program that Kurzweil expected to build this network spent roughly $20 billion and delivered almost none of it; a civilian satellite constellation plus off-the-shelf drone autopilots delivered the architecture anyway.

Closed-timelike-curve computers: still thought experiments. Kurzweil attributed to Todd Brun (ch. “Going Back in Time”) the claim that “closed-timelike-curve computation would greatly expand the potential of local computation.” Brun’s 2003 paper still stands, and Scott Aaronson and John Watrous sharpened the picture in 2008: a computer with polynomial-size CTCs would have the power of PSPACE — enough to factor integers trivially and solve quantified SAT in polynomial time. No hardware exists. No experimental pathway has been proposed. The bound is rigorously established; the physics remains speculative.

Lloyd’s cosmic ceilings are still ceilings. Kurzweil cited Seth Lloyd’s estimates of “about 10^90 calculations per second” for the universe and “at least 10^70 calculations per second” for the solar system (ch. “The Criticism from Malthus”). The Frontier supercomputer at Oak Ridge currently benchmarks around 10^18 AI-relevant operations per second — roughly 52 orders of magnitude short of the solar-system ceiling. These numbers were never meant to be milestones; they are the walls of the room. They remain the walls.

The scorecard

What Kurzweil nailed and what he missed

A pattern shows up across every prediction in this batch: Kurzweil correctly identified the function that the 2020s would need, and mis-specified the mechanism that would deliver it. Self-managing infrastructure: yes, but as declarative container orchestration rather than IBM’s autonomic middleware. Molecular nanostructures: yes, but as drug delivery and storage rather than SIMD computing. Distributed battlefield networks: yes, but through low-earth-orbit satellites and commercial drone autopilots rather than a $20 billion Army integrator. Compute-per-dollar: yes, but through GPU and TPU architecture rather than another two decades of transistor halving. The meta-lesson, visible in The Singularity Is Nearer itself, is that Kurzweil’s macro curves held up better than his object-level roadmaps — and he seems to know it, which is why the 2024 book quietly swaps transistors for operations-per-second per dollar and picks up DNA origami in place of DNA computing.

The cases that aren’t yet scorable — CTC computers and the Lloyd ceilings — share a feature: they were never predictions about the 2020s at all. They were framing devices for how far compute could, in principle, still go. Twenty years on, they remain exactly that.

Method note

Claims about publication and patent volumes come from the repo’s own index of US utility patents from 1976 onward (9.3 million documents) and OpenAlex’s literature corpus (357 million works), both full-text-searched for the relevant terminology. The specific patents named above were pulled and their claims read in full. The Singularity Is Nearer passages were located in the extracted Kindle text and quoted verbatim. Military, semiconductor, and self-healing infrastructure context came from TSMC and TSMC-adjacent reporting, CSIS analysis of Ukraine’s Delta system, and the Aaronson-Watrous CTC complexity paper.