Kurzweil Scorecard: The Substrates That Didn’t Win

In 2005, Ray Kurzweil built his case for accelerating returns on a parade of
exotic computing prototypes — a DNA supercomputer in Rehovot, a liquid-crystal
molecule in Oklahoma storing a thousand bits of spin, an optical chip from an
Israeli startup doing eight trillion calculations a second with 256 lasers,
an atomic memory drive projected at 250 terabits per square inch. Twenty-one
years later, almost none of those substrates are inside the chips running
this sentence. Silicon won. It won by getting weirder — gate-all-around
transistors, backside power delivery, chiplet stacks, EUV lithography — but
the substrate Kurzweil was betting would be replaced is the one that quietly
absorbed every trick of its would-be replacements.

The interesting part is that Kurzweil knew this might happen. The one
forward-looking prediction in this batch — that quantum computers, even with
hundreds of entangled qubits, would remain special-purpose rather than become
general-purpose — has aged the best of anything he said about computing.

The predictions

This batch is a snapshot of computing circa 2005, drawn from five chapters of
The Singularity Is Near. It mixes three kinds of statement: historical
observations, contemporary lab demos treated as signal, and one explicit
forward bet. Kurzweil’s argument bundled all three together: the historical
trend plus the lab demos proved, in his framing, that Moore’s Law was just
one instance of a more general force that would ride new substrates once
silicon faltered. In The Singularity Is Nearer (2024), he reaffirms the
framing: “After integrated circuits have reached their limits, new
paradigms using nanomaterials or three-dimensional computing will take
over” (ch. “Deep Learning”). The substrates in the 2024 update are
different from the ones in 2005 — memristor networks, atomically precise
silicon lithography, single-atom qubits rather than Lenslet or the Weizmann
DNA rig. The list got swapped without comment. That swap is the story.

Where we actually are

Transistors and Moore’s Law. Kurzweil wrote that “the number of
transistors in Intel processors had doubled every two years” and
“transistors had become faster by about a factor of one thousand over
the previous thirty years” (ch. “Moore’s Law and Beyond”). The historical
claim is true. The forward momentum is more complicated. TSMC’s 2 nm (N2)
node hit mass production in the second half of 2025 at roughly 313 million
transistors per square millimeter for high-density logic; Intel’s 18A,
also in production in 2025, reaches about 238 MTr/mm² with gate-all-around
“RibbonFET” transistors and backside power delivery via PowerVia. Density
is still climbing, but the doubling time has stretched well past two years,
and the gains now come from architectural innovation layered on top of
feature-size shrinks — US 12,598,788 (silicon-germanium nanosheet
structures, April 2026) reads like normal incremental silicon engineering,
not substrate revolution.

MIPS per thousand dollars. “Computing had advanced from taking ninety
years to achieve the first MIPS per thousand dollars to adding one MIPS per
thousand dollars every five hours” (ch. “The Fifth Paradigm”). Kurzweil
updates this in 2024 with the IBM 7094 he used at MIT in 1965 — $30 million
in today’s dollars for 0.25 MIPS — against the iPhone 14 Pro’s 17 trillion
operations per second at roughly $1,000, a “two-trillion-fold improvement.”
The 2024 update concedes something the 2005 chart did not: the post-2010
acceleration in AI training compute — doubling every 5.7 months — is
mostly not from transistor scaling. It’s from parallelism, algorithmic
efficiency, and capital. The curve kept bending; the engine changed.

Wireless doubling. “Wireless communication power was doubling every
ten to eleven months” — Edholm’s law, from a 2004 Nortel presentation.
Two decades on, the underlying trend has held at roughly 18-month doubling.
4G peaks of 10–20 Mbps in the mid-2010s have given way to 5G mmWave peaks
of 20 Gbps; 6G targets published in 2025 aim for 1 Tbps peak rates by
2030 and a fivefold spectral-efficiency improvement over 5G. The wrinkle
is that wireless bandwidth has begun to commoditize in the opposite
direction from what Kurzweil was pointing at — instead of feeding a merger
of human and machine cognition, it mostly feeds video streaming.

Atomic-scale memory. “Scientists at the University of Wisconsin and
University of Basel created an atomic memory drive in 2002 with a projected
density of about 250 terabits per square inch” (ch. “Computing with
Molecules”). The projection was hit in the lab — a 2016 demonstration
stored 8,000 bits at roughly 502 Tb/in², twice Kurzweil’s number. But it
required cryogenic cooling to 4K and individually positioned atoms, and
has never produced a commercial drive. Hard-disk areal density in 2025
targets about 2 Tb/in² using heat-assisted magnetic recording. Atomic
memory is two orders of magnitude ahead of HDDs in the lab and zero
orders of magnitude ahead in any data center.

DNA computing. “In 2003, Ehud Shapiro’s team at the Weizmann Institute
demonstrated a DNA-ATP liquid computing system… performing 660 trillion
calculations per second using only fifty millionths of a watt” (ch.
“Computing with DNA”). The 2003 demo was real; its successor products were
not. DNA as a computing substrate has essentially been replaced by DNA
as a storage substrate. Twist Bioscience spun its storage division out
as Atlas Data Storage in May 2025, targeting 13 terabytes in a droplet of
water; the Atlas Eon 100, unveiled in late 2025, claims 60 petabytes in
60 cubic inches. The key missing piece — random access — shows up in
patents like US 12,441,996 (October 2025), which describes DNA origami
nanostructures used to package indexed oligonucleotides, enabling
“selective physical data access and retrieval from a molecular pool.”
The substrate is real; the application pivoted from computing to cold
storage.

Molecular spin storage. “Scientists at the University of Oklahoma
demonstrated storing 1,024 bits of information in a single liquid-crystal
molecule containing nineteen hydrogen atoms” (ch. “Computing with Spin”).
This one-off lab result never became a technology. There is no
spin-molecule memory in any shipping device in 2026.

Optical SIMD. “Lenslet had developed an optical SIMD system using
256 lasers capable of eight trillion calculations per second”
(ch. “Computing with Light”). Lenslet shut down a few years after
Kurzweil wrote that paragraph. The idea came back twenty years later
in a different package. Lightmatter’s Passage M1000, unveiled in March
2025, uses a 3D-stacked photonic superchip to deliver 114 terabits per
second of optical interconnect bandwidth for AI data centers, with a
co-packaged-optics module named L200 following in 2026. Optical-computing
patents crawled along at 1–4 per year through the mid-2010s and jumped
to 8 in 2025 alone, led by inventions like US 12,387,094 (Photonic tensor
accelerators for artificial neural networks, August 2025), whose abstract
describes “photonic units for vector-vector multiplication, matrix-vector
multiplication, matrix-matrix multiplication… through coherent mixing
and square-law detection.” This is Lenslet’s vision, arrived twenty
years late on the back of silicon photonics and specifically shaped for
transformer inference. Photonic-neural-network papers grew from 60 in
2015 to more than 1,000 in 2025.

Moore’s 1965 forecast. Kurzweil quotes Gordon Moore’s original 1965
Electronics article predicting “65,000 components by 1975”
(ch. “The Sixth Paradigm”). The forecast was optimistic by about four
years — Intel’s 1974 8080 had 4,500 transistors, and 65,000-transistor
single-die designs arrived around 1979 — but right on direction.

Quantum stays special-purpose. This is the one forward-looking bet in
the batch: “Even if quantum computers with hundreds of entangled qubits
become feasible, they will remain special-purpose devices rather than
general-purpose replacements” (ch. “Quantum Computing”). In 2025,
quantum computing passed exactly that threshold and no further. Microsoft
and Atom Computing encoded 28 logical qubits onto 112 physical neutral
atoms and entangled 24 of them, with a 50-logical-qubit machine named
Magne slated for early 2027. Quantinuum demonstrated logical qubits
outperforming their physical counterparts. Google’s Willow chip, via the
Quantum Echoes algorithm, achieved a verifiable speedup of roughly
13,000× over the best classical approach on a narrow out-of-time-order-
correlator problem. IonQ and Ansys reported a medical-device simulation
on a 36-qubit machine beating classical HPC by 12%. Every one of those
wins is on a bespoke problem class. None threatens a general-purpose CPU
or GPU. Patent filings since 2022 are dominated by IBM, Google, Microsoft,
IonQ, and D-Wave, with titles that read like a catalog of error-correction
schemes — lattice surgery, magic patches, bias-error codes, multimode
grid states — rather than general-purpose architectures. Moody’s 2025
industry survey identifies the dominant trend as “specialized hardware
and software for specific problem classes rather than universal quantum
computing approaches.” Kurzweil nailed this one.

The scorecard

What Kurzweil missed (and what he nailed)

The pattern in this batch is sharp. When Kurzweil described a trend —
transistors scaling, wireless scaling, price-performance scaling — he was
almost always right in direction, if sometimes slow in timing. When he
described a specific substrate as the heir apparent — Weizmann DNA,
Lenslet optics, Oklahoma spin molecules, Wisconsin-Basel atomic memory —
he was almost always wrong in specifics, occasionally right in spirit.
DNA became a storage medium rather than a compute medium. Optical
computing came back, but not from Lenslet and not as standalone SIMD; it
came back as co-packaged interconnect on silicon-photonic wafers inside
AI data centers. Atomic memory stayed in the lab. The spin molecules
stayed where they were.

The one clean forward-looking prediction — that quantum would remain
special-purpose even at the hundred-qubit scale — has aged the best. In
2005, the default technologist fantasy was a general-purpose quantum
computer that would replace CPUs. Twenty-one years later, with verifiable
quantum advantage demonstrated and 50 logical qubits on the roadmap, the
industry consensus has rotated to exactly Kurzweil’s position: quantum is
a coprocessor for a narrow class of problems. The public caricature of
Kurzweil is the perpetual optimist. On quantum in this batch, he was the
deflationist, and the deflationist won.

The honest lesson from the rest of the batch is a forecasting asymmetry.
Betting on exponential trends in well-instrumented, economically
important capacities paid off. Betting on specific exotic substrates
mostly didn’t. The iPhone in your pocket has almost none of the substrates
Kurzweil pointed at in 2005. It is made of silicon, assembled with EUV
lithography, laid out in nanosheets with backside power, and connected by
copper and (increasingly) co-packaged photonics — a hybrid nobody in 2005
would have described as the winning architecture, and that is the winning
architecture anyway.

Method note

We pulled each prediction from a structured index of about 1,100 testable
claims extracted from The Singularity Is Near, then cross-referenced
against 9.3 million U.S. patents, 357 million OpenAlex papers, and
current industry reporting. For each prediction we surfaced recent
filings and publications on the relevant substrate, identified specific
high-signal patents by number, and grounded the narrative in verbatim
passages from The Singularity Is Nearer (2024). Verdicts reflect
evidence available through April 2026; corrections welcome.