Kurzweil Scorecard: The Law of Accelerating Returns, Twenty Years On

In 2005, Ray Kurzweil bet the whole house on a single curve: the exponential improvement in price-performance of computing. Twelve of the predictions in The Singularity Is Near are really variations on that one bet, staked across different horizons and different substrates — molecular logic, nanotube circuits, self-configuring hardware, $1,000-equals-one-human-brain milestones.

Two decades later, the curve is still bending in roughly the right direction. But almost none of the specific mechanisms Kurzweil nominated to keep it bending have arrived. The bet paid off; the rationale for why it would pay off looks wrong in detail and right in spirit.

The big trend line

Kurzweil’s load-bearing claim was that “the pace of human-created technological change is accelerating and the powers of technology are expanding exponentially” (ch. “The Intuitive Linear View Versus the Historical Exponential View”). In The Singularity Is Nearer (2024) he restates the same idea but with updated receipts: “one dollar buys about 11,200 times as much computing power, adjusting for inflation, as it did when The Singularity Is Near hit shelves.”

That number is, if anything, conservative for the segment that matters. The compute used to train a frontier AI model has been doubling every 5.7 months since 2010 — Kurzweil’s own 2024 accounting — a pace roughly four times faster than Moore’s Law at its best. Against his 2005 forecast that “the rate of paradigm shift or technical innovation is currently doubling every decade” (ch. “The Singularity Is Near”), the AI training curve has been moving perhaps three decades of “Kurzweil time” per decade of wall-clock time.

The companion claim — “by 2030, one thousand dollars of computing will equal roughly one thousand human brains in processing power” (ch. “Human Memory Capacity”) — depends entirely on what you call a human brain. Kurzweil’s own 2005 estimate was 10^14 operations per second. A single Nvidia H100 today delivers close to 10^15 FP16 tensor operations per second, and sells in the low tens of thousands of dollars; consumer-grade GPUs in the $1,500–$2,000 range push past 100 teraflops at reduced precision. By Kurzweil’s own brain-ops figure, $1,000 already buys something in the neighborhood of a single human brain’s raw throughput five years ahead of the 2030 milestone. The thousand-brain target by 2030 is plausible but not guaranteed; the 2050 target of one thousand dollars exceeds the processing power of all human brains on Earth (roughly 10^24 ops/sec for 8 billion people) is still a nine-orders-of-magnitude jump from here.

The mechanisms that didn’t arrive

This is where the scorecard gets interesting. Kurzweil was specific about how the curve would keep bending after silicon hit its wall. The candidates he nominated — molecular rod logic, self-configuring nanocircuits, carbon-nanotube memory replacing DRAM, Moore’s Law reaching the end of its S-curve before 2020 — have mostly not panned out. The curve kept bending anyway, but through different mechanisms.

Drexler’s mechanical molecular computer is the cleanest miss. The Singularity Is Near described “a mechanical molecular computer using locks instead of transistor gates, with each lock occupying 16 cubic nanometers and switching 10 billion times per second” (ch. “Nanotechnology”). In The Singularity Is Nearer, Kurzweil candidly notes of these designs that “no one has actually built them yet,” and concedes that “it will be at least a decade before the field starts maturing.” Two decades in, molecular rod logic remains a paper architecture. In our patent corpus, filings explicitly mentioning molecular logic gates run at roughly one patent per year — effectively dormant.

Self-configuring nanocircuits — Kurzweil’s prediction that “future nanocircuits will continuously monitor their own performance and route information around unreliable sections, analogous to Internet routing around failed nodes” (ch. “Self-Assembly”) — did arrive, but not at the nanoscale. US 12,399,781 (granted August 2025, “Methods and apparatus to increase resiliency in self-healing mechanisms”) describes systems that partition a host into a primary and shadow partition, apply a fix to the primary, and fail over to the shadow if the primary loses communication. US 12,007,832 (“Restoring a system by load switching to an alternative cloud instance and self healing”) does much the same at the VM layer. The self-healing story is happening — but at the datacenter and hypervisor layer, not inside a nanocircuit. Wrong mechanism.

IBM’s self-diagnosing microprocessors were claimed circa 2005 as “microprocessor designs that automatically diagnose problems and reconfigure chip resources accordingly.” This one was effectively already true when Kurzweil wrote it — IBM’s 2001 Autonomic Computing initiative was already commercial — and has matured in step with server-class silicon since. Modern Power and Xeon chips do core sparing, thermal throttling, and isolation of faulty cache banks as a matter of course. Call it on track.

Moore’s Law reaches the end of its S-curve before 2020 is the most consequential miss in the batch. It didn’t. TSMC started 5 nm mass production in 2020 at 173 million transistors per square millimeter; the N3 node arrived in 2022 at 290 M/mm²; N2, with gate-all-around nanosheet transistors, entered mass production at the end of 2025 at roughly 490 M/mm². Each node stretch has taken longer than the last — the N3-to-N2 transition took three years, not eighteen months — but the paradigm change Kurzweil predicted for “before 2020” is in fact happening in the 2025–2030 window, and the replacement mechanism isn’t molecular rod logic or carbon nanotubes. It’s 3D stacking, chiplet disaggregation, and nanosheet FETs.

Chiplets, notably, are the quiet story. Filings in our patent index that explicitly describe chiplet architectures went from 4 in 2016 to 77 in 2024 and 76 through 2025 — a 19× increase in less than a decade. US 12,423,250 (“Integrated chiplet-based central processing units with accelerators,” granted September 2025) describes a system-on-chip with a CPU coupled to an accelerator via a die-to-die interconnect and uniform memory accessed through a second die-to-die link. The claims read like the architectural skeleton of every frontier AI accelerator currently shipping. This is the paradigm that actually took over from monolithic CMOS. Kurzweil didn’t name it.

Molecular computing as the mainstream “sixth paradigm” by 2005 was simply wrong. It never went mainstream. Instead, the field forked into two unrelated successors: neuromorphic computing, whose 2025 flagship Intel Hala Point runs 1.15 billion spiking neurons across 1,152 Loihi 2 chips, and quantum computing, which is still pre-commercial. Neuromorphic patents in our index grew from 1 in 2013 to 29 in 2025. The sixth paradigm arrived — it just isn’t mechanical.

The scorecard

What Kurzweil missed, and what he nailed

The pattern across this batch is striking. Kurzweil was extraordinarily right about the curve and extraordinarily wrong about the mechanism. He named molecular rod logic, nanotube memory, and self-configuring nanocircuits as the successors to silicon. What actually took over was chiplet disaggregation, 3D stacking, gate-all-around nanosheet transistors, and neuromorphic accelerators — none of which appear in the 2005 book with anything like the prominence he gave to molecular computing.

This is a useful corrective for anyone reading today’s forecasts. The direction and steepness of a technology trend are often easier to call than the specific substrate that will carry it. Anyone who in 2005 had bet their career on molecular rod logic would be nowhere; anyone who had bet on “compute price-performance keeps bending regardless of substrate” would have been right for two decades straight. That is Kurzweil’s real legacy in this batch: not the specific roadmap, but the refusal to get distracted by the apparent end of any given S-curve.

The one forecast that looks genuinely endangered is the “$1,000 exceeds all human brains on Earth by 2050” target. It requires roughly nine more orders of magnitude of price-performance improvement in 25 years. That is within the range of the post-2010 AI compute trend, but well outside Moore’s-Law pace. Which mechanism gets us there — neuromorphic, photonic, quantum, or something not yet named — is the live question. Kurzweil’s record suggests the curve will bend. It also suggests whatever he names as the mechanism will probably not be the one that does it.

Method note

Scorecards in this series are built by interrogating a local index of 9.3 million U.S. patent filings with full-text search, plus 357 million scientific papers from the open literature, and cross-checking against live web sources for current benchmarks, product launches, and roadmap updates. Patent numbers cited above are real U.S. grants; the chiplet and self-healing patents referenced were pulled in full and read before being quoted. Historical compute-per-dollar figures and AI training compute trends are taken verbatim from Kurzweil’s own 2024 update, The Singularity Is Nearer.