Kurzweil Scorecard: Stacked Memory, Mesh Dreams, and the Sixth Paradigm’s Actual Shape

In 2005, Ray Kurzweil bet a chapter of The Singularity Is Near on a handful of startups whose logos have since faded from investor decks — Nantero, Matrix Semiconductor, Masuoka’s cylindrical-memory venture — and on a model of distributed “mesh” computing that imagined your laptop renting cycles from the laptops around it. Twenty years on, one of those bets is buried inside almost every SSD shipped on Earth. Another is a cautionary tale of production slides deferred year after year. And the mesh never really came: the cloud ate it.

That mixed record turns out to be the most honest way to read this batch of twelve predictions. Kurzweil’s directional claim — that silicon would learn to grow upward once it could no longer shrink sideways — is among the most vindicated ideas in twenty-first-century hardware. The specific corporate names, the specific mechanisms, and especially the specific slopes of his exponentials are another story.

The predictions

The batch covers the engine room of Kurzweil’s argument: Moore’s Law and its successors, 3D memory, mesh computing, supercomputer brain-equivalence, and molecular CAD for the nanotech era he believed would open by 2025. These are the claims that had to hold for the later predictions about AGI, brain uploading, and nanomedicine to have any physical basis.

Kurzweil wrote that “information technologies, including price-performance, speed, capacity, and bandwidth, are now doubling about every year” (ch. “The Singularity Is Near”). He re-stated the underlying claim last year, more conservatively: “one dollar buys about 11,200 times as much computing power, adjusting for inflation, as it did when The Singularity Is Near hit shelves” (The Singularity Is Nearer). Eleven thousand over nineteen years is a doubling every 18 months — Moore’s original cadence, not the annual one. The sleight is important. Much of the rest of this scorecard is the gap between those two numbers.

Where we actually are

The 3D memory bet was a triumph — just not for the company he named. Matrix Semiconductor was founded in 1998 by Thomas Lee, Mark Johnson, and Mike Farmwald, and sold to SanDisk for $238 million in October 2005, months after Singularity Is Near hit shelves. Matrix’s specific trick — antifuse-based, stackable, write-once cells adapted from LCD manufacturing — did not survive. The idea that you could grow a memory chip upward rather than outward did. In our patent corpus, filings on three-dimensional or vertically stacked NAND grew from 18 in 2005 to 424 in 2024, with Micron, Sandisk, Samsung, Yangtze Memory, and SK hynix as the top assignees since 2020.

The layer counts now read like moonshot telemetry. SK hynix is shipping 321-layer NAND; Samsung is targeting 430 layers in its V10 generation for late 2025; Yangtze Memory’s 294-layer parts are in volume production. US 12,363,898, granted to Yangtze Memory in July 2025, describes a “3D NAND memory device with non-uniform channel structure” whose core claim is a second die bonded face-to-face with the memory die, carrying peripheral circuitry beneath the memory cells themselves. US 12,310,014, granted to Applied Materials two months earlier, claims new selection-gate isolation methods specifically for the 3D case. The vocabulary Kurzweil reached for in 2005 — “vertically stacked planes of transistors” — is now the house style of a multi-hundred-billion-dollar industry.

Nantero is the mirror case. Kurzweil wrote that “Nantero’s nanotube memory design, combining random access and nonvolatility, could potentially replace RAM, flash memory, and disk storage” (ch. “Nanotubes Are Still the Best Bet”). Twenty years later, Nantero is still filing. US 10,854,243, granted to Nantero in December 2020, describes a full nonvolatile memory array built from “two-terminal nonvolatile nanotube block switches” arranged in a bit-line/word-line grid. US 10,885,978, granted in January 2021, describes a single switch cell with a carbon-nanotube fabric sandwiched between top and bottom conductive terminals. The engineering is real. The market is not. A Fujitsu licensing deal announced in 2016 targeted 2018 mass production, then 2019, then 2020; there is still no consumer or hyperscaler product line built on NRAM. The bet on nanotube memory replacing RAM, flash, and disk was directionally plausible and mechanically wrong — a different vertical won.

Fujio Masuoka’s cylindrical memory was the same kind of miss. Masuoka, the Toshiba engineer who invented NAND flash, went on to pitch a cylindrical transistor geometry he claimed was ten times denser than planar chips. That specific architecture didn’t win either — but his original invention, flash, got the vertical treatment his cylinder was trying to achieve, via entirely different process chemistry. Score it as spiritually vindicated and literally beaten.

The cost-per-transistor cliff is where Kurzweil was genuinely caught out. He asserted in 2005 that “the cost per transistor cycle was halving every 1.1 years” (ch. “Moore’s Law and Beyond”). That stopped being true at the 28-nanometer node in 2011. Today, a TSMC 28-nanometer wafer costs roughly $3,000; a 3-nanometer wafer is around $20,000 and rising, with 2-nanometer wafers expected to clear $30,000. TSMC announced 5-10 percent price hikes for sub-5-nanometer nodes starting in 2026. The density keeps climbing — NVIDIA’s GB202 graphics processor packs 92.2 billion transistors — but the cost per transistor has flattened and in many cases risen. Intel’s Paolo Gargini predicted in 2005 that Moore’s Law “could continue for at least the next 15 to 20 years” — 2020 to 2025, now. In density terms, he was right; in cost-per-transistor terms, he was wrong by at least a decade.

Supercomputers and the human brain finally shook hands — eleven years late. Kurzweil predicted that “supercomputers will have the requisite hardware to emulate human intelligence by the end of the 2000s decade” (ch. “The Singularity Is Near”), citing a brain-scale benchmark of roughly 10^16 operations per second. IBM Roadrunner cleared 10^15 in 2008; 10^16-class machines didn’t appear until Summit in 2018 and Fugaku in 2020. The first true exascale machine — Oak Ridge’s Frontier at 1.102 exaFLOPS — arrived in 2022. Lawrence Livermore’s El Capitan, verified at 1.742 exaFLOPS in November 2024, is now the fastest computer on Earth. In The Singularity Is Nearer, Kurzweil regroups: “Oak Ridge National Laboratory’s Frontier… can perform on the order of 10^18 operations per second. This is already on the order of 10,000 times as much as the brain’s likely maximum computation speed.” The claim is right now. It just wasn’t right by 2010.

Mesh computing is the cleanest “wrong mechanism” in the batch. Kurzweil imagined every networked device contributing idle cycles to its neighbors, adding “another factor of 100 to 1,000 in effective price-performance” by the 2020s. The closest thing we have to that today are BOINC and Folding@home. BOINC runs about 20 petaFLOPS across 18,000 participants; Folding@home, which briefly hit 2.4 exaFLOPS during the COVID surge, is back to roughly 13 native petaFLOPS as of October 2025. Meanwhile, hyperscaler cloud capacity is measured in hundreds of exaFLOPS, and virtually every training run that matters happens there. The economics of coordination, data gravity, and memory bandwidth all favored centralization. The mesh isn’t dead — it’s just two rounding errors below the cloud.

Molecular CAD exists, but not for the reason Kurzweil expected. His 2005 prediction imagined CAD systems that could accept a 3D scan of an existing product and spit out molecular assembly instructions. What actually emerged is a rich ecosystem of DNA-origami design tools — caDNAno, originally from the Wyss Institute and now maintained by UCSF’s Douglas Lab, plus ATHENA, MagicDNA 2.0, CanDo, and Parabon’s commercial inSēquio. These tools let biologists design specific nanostructures strand-by-strand and folded structure by folded structure, not replicate arbitrary macro products. It’s genuine molecular engineering, but it points at biology and pharmaceuticals, not at the universal assembler Drexler and Kurzweil described.

The scorecard

What Kurzweil missed (and what he nailed)

The pattern across this batch is consistent: Kurzweil was right about direction and wrong about rate and mechanism. When he said memory would go three-dimensional, he was right enough that the entire flash industry has been rebuilt around the idea. When he said it would go three-dimensional via carbon nanotubes from a Woburn, Massachusetts startup named Nantero, the market picked another mechanism. The structural prediction survives. The operational one — the bet you could actually have placed in 2005 — often does not.

A second pattern: Kurzweil’s slopes are aggressive by about a factor of two. Doubling every year became doubling every 16–18 months. Brain-scale supercomputers by 2010 became brain-scale supercomputers by 2020. Nantero production in 2018 became production in 2020 became production in an undetermined year. The curves bend the way he said they would — they just move slower. Being off by a factor of two on timing while right on shape is closer to a hit than a miss. That the mesh never came — and the cloud took its place — is the single largest architectural surprise in this batch, and it rewrote the next twenty years of the industry.

Method note

We read the twelve predictions out of The Singularity Is Near, then checked each against three evidence streams: a corpus of 9.3 million US patents queried for vertically stacked NAND and carbon-nanotube memory filings by year and assignee, with specific claims pulled from the top grants; public supercomputer rankings (Frontier, El Capitan), semiconductor analyst 3D NAND layer counts, and TSMC wafer-cost data; and recent review coverage of the DNA-origami CAD ecosystem. Kurzweil’s own updated figures in The Singularity Is Nearer (2024) are his admission of the revised slope, and we used them as the scoring benchmark wherever available.