Kurzweil Scorecard: Computing Hardware and the Sixth Paradigm

US patent 12,433,031, granted September 30, 2025, describes a three-dimensional semiconductor device where two complete layers of transistors — fabricated independently, then oxide-bonded face-to-face — communicate through vertical pillars threaded between embedded memory arrays. The claims read like an architectural diagram for the future Kurzweil predicted in 2005: computing that stacks vertically, that merges logic and memory, that escapes the flatland of Moore’s Law by building up instead of shrinking down. Except Kurzweil called it “molecular computing.” What actually arrived is copper-bonded silicon, engineered not at the molecular scale but at the scale of a few hundred atoms — close enough to smell it, far enough to matter.

What Kurzweil predicted

This batch covers Kurzweil’s hardware arc: the idea that computing power would keep growing exponentially by jumping from one physical paradigm to the next. He named six paradigms — electromechanical, relay, vacuum tube, transistor, integrated circuit — and predicted a sixth: “three-dimensional molecular computing will become the sixth paradigm and continue exponential growth after Moore’s Law” (ch. “The Fifth Paradigm”). He expected it to arrive in the 2020s.

Around that central claim orbit a constellation of specific forecasts: “by the end of the 2000s decade, computers will disappear as distinct physical objects, with displays in eyeglasses and electronics woven into clothing” (ch. “Deflation … a Bad Thing?”); “about 10^13 bits of memory, enough for functional human memory, will be purchasable for one thousand dollars by around 2018” (ch. “Human Memory Capacity”); and on the far horizon, reversible computing delivering a “factor of a billion energy reduction” over nonreversible computing (ch. “Powering the Singularity”).

Where we actually are

3D chips: the sixth paradigm arrived sideways

Kurzweil’s “sixth paradigm” was molecular computing — logic gates built from individual molecules, self-assembled nanotube circuits, DNA computing. None of those became the production technology. What happened instead was 3D stacking: physically bonding conventional silicon layers on top of each other.

TSMC holds 60% of the global 2.5D and 3D stacking market through its CoWoS and SoIC technologies, with plans to more than double output between 2024 and 2025. Their System-on-Integrated-Chips (SoIC) technology uses bumpless hybrid bonding — direct copper-to-copper connections between stacked dies — to build devices that would read as science fiction to a chip designer from 2005. Samsung has committed over $1.5 billion toward advancing 3D IC packaging with hybrid bonding and through-silicon vias. Intel’s Foveros technology enables active die-on-die stacking.

In our patent data, SanDisk Technologies leads 3D chip grants since 2020 with 317 patents, followed by Yangtze Memory Technologies (412 combining variants), Samsung (259), Monolithic 3D Inc. (195), TSMC (105), Micron (86), and Intel (70). The sheer volume — over 1,006 US grants in 2024 alone, up from 231 in 2000 — reflects an industry that has bet its future on vertical integration.

Patent 12,432,917, also granted September 30, 2025, shows just how sophisticated this has become: a three-dimensional memory device where alternating stacks of insulating and conductive layers are drilled through with “memory openings” to create vertical channels, with sacrificial spacer layers replaced by dielectric material to form cavities between the active layers. This is the kind of atomically precise layer-by-layer engineering Kurzweil would recognize as a descendant of his vision — even though the base material is still silicon, not carbon nanotubes.

The verdict: the sixth paradigm arrived, but as an engineering tour de force in silicon rather than a phase transition to molecular substrates.

Memory for $1,000: nailed it, then the floor fell out

Kurzweil predicted that 10^13 bits (about 1.25 terabytes) of memory would cost $1,000 by 2018. He was roughly right: by 2018, a 1TB SSD cost around $150-250, and 1TB of DDR4 RAM cost about $3,000-5,000. Depending on whether you count storage or volatile memory, he was either ahead of schedule or close to target.

Then something Kurzweil didn’t predict happened: memory prices stopped falling. In late 2025, DRAM prices surged roughly 90% year-over-year, driven by insatiable AI demand for high-bandwidth memory. A 32GB DDR5 kit that cost $100-200 in October 2025 now starts at $350. SK Hynix’s new HBM4 chips exceed 2 TB/s of bandwidth per chip — 60% faster than HBM3E — but the price-per-gigabyte curve has temporarily inverted. Memory shortages are expected to last through at least Q4 2027.

The long-term exponential holds. But the short-term disruption — AI gobbling the world’s memory supply — is a dynamic Kurzweil’s smooth curves didn’t capture.

Computers disappearing into clothing: fifteen years late, but arriving

Kurzweil wrote that by the end of the 2000s, “computers will disappear as distinct physical objects, with displays in eyeglasses and electronics woven into clothing, enabling full-immersion visual virtual reality” (ch. “Deflation … a Bad Thing?”). In 2009, the iPhone 3GS was the state of the art. Computers had very much not disappeared.

But in 2026, they’re starting to. Meta has sold over 2 million Ray-Ban Smart Glasses. Apple is reportedly speeding up development of AR glasses — pausing work on a second Vision Pro to get glasses-form-factor computing to market, with an announcement expected in late 2026 and mass production potentially starting in early 2027. Analyst Ming-Chi Kuo projects Apple will ship 3-5 million smart glasses units in 2027. AI smart glasses sales overall are projected to surge from 1.5 million pairs in 2024 to 90 million by 2030.

The patent record shows this acceleration: US grants for wearable computing, smart clothing, and AR glasses leapt from 20 per year in 2000 to 775 in 2024. Patent 12,429,701, granted September 2025, describes augmented reality glasses built on a curved waveguide — the key optical breakthrough that makes AR glasses slim enough to wear as normal eyewear rather than as a headset. The claims detail an input optical compensator that pre-distorts the projected image so that when it bounces through the curved waveguide, it arrives at the eye undistorted. This is the engineering that separates a visor from a pair of glasses.

Kurzweil was fifteen years early. But the direction is exactly what he described.

Protein folding: the prediction nobody noticed

Buried in this batch is a 2005 claim that “supercomputers scheduled to come online around 2005 are expected to have enough computational capacity to simulate protein folding and interactions between three-dimensional proteins” (ch. “Life’s Computer”). This prediction was almost absurdly conservative. Not only did the hardware arrive — it was surpassed so completely that by 2021, DeepMind’s AlphaFold 2 solved protein structure prediction for most known proteins, earning a Nobel Prize equivalent in biology. The key paper, cited over 42,000 times in our literature database, is the most-cited computational biology paper of the century.

AlphaFold 3, announced May 2024, extended the model to predict structures of protein-DNA, protein-RNA, and protein-ligand complexes — the interactions that drive drug design. It showed at least a 50% accuracy improvement over existing physics-based methods on the PoseBusters benchmark, and over 3 million researchers in 190 countries have used the freely available AlphaFold database.

Kurzweil predicted the hardware would be ready by 2005. He didn’t predict that the software would take another 16 years — or that when it arrived, it would render decades of physics-based simulation approaches largely obsolete overnight.

Reversible computing: behind schedule, but suddenly relevant

Kurzweil predicted that “over the next several decades computing will transition to reversible computing, reducing energy needs by as much as a factor of a billion compared with nonreversible computing” (ch. “Powering the Singularity”). For twenty years this prediction looked dead. Reversible computing was a curiosity confined to theory papers.

Then AI ate the power grid. Frontier model training runs now consume megawatts. Data centers are negotiating directly with nuclear plants for power. Suddenly, a 1,000x energy reduction isn’t academic — it’s existential.

Vaire Computing, a startup, demonstrated its “Ice River” prototype in early 2025: an arithmetic circuit that recovers 40-70% of computational energy using adiabatic resonators. Their roadmap targets a 4,000x energy reduction within ten to fifteen years. Michael Frank at Sandia National Laboratories projects that reversible computing could ultimately deliver the kind of efficiency gains Kurzweil predicted, with IEEE’s semiconductor roadmap acknowledging that conventional digital logic efficiency will plateau later this decade.

Patent 12,431,188, granted September 2025, describes an efficient Muller C-Element — a fundamental building block of asynchronous, potentially reversible logic — optimized for high bit-width applications. The claims detail a circuit that uses active resistors and parallel OR configurations to minimize switching energy. It’s not molecular computing. But it’s the kind of building block that could eventually compose into the energy-efficient architectures Kurzweil envisioned.

The scorecard

What Kurzweil missed (and what he nailed)

The hardware batch reveals a forecaster who understood economics better than materials science. Kurzweil’s fundamental claim — that computing power per dollar would keep growing exponentially by jumping to new paradigms — has held up remarkably well. What he got wrong, consistently, was which paradigm would do the jumping. He bet on molecular computing, DNA logic, nanotubes. What won was relentless refinement of silicon: extreme ultraviolet lithography pushing to 2nm, 3D stacking bonding complete layers of transistors face-to-face, and specialized accelerators (GPUs, TPUs) that do one thing at blinding speed.

The protein folding prediction is the sleeper hit of this batch. Kurzweil framed it as a hardware story — enough FLOPS to simulate folding. But the actual breakthrough was algorithmic: AlphaFold used machine learning to bypass the simulation entirely, predicting structures from amino acid sequences alone. Hardware was necessary but not sufficient. The algorithm was the revolution.

And the reversible computing prediction, laughable for two decades, has suddenly acquired urgency. AI’s power appetite has made energy efficiency an industry-wide crisis. Whether the solution is Kurzweil’s reversible logic, or something he didn’t foresee — like running inference on smartphones instead of data centers — the problem he identified is now the hottest constraint in the field.

Method note

This scorecard draws on 9.3 million US patent grants from USPTO bulk XML, with full-text claims and descriptions read for key patents cited in the text. Patent numbers and grant dates are verified from the database. The 357-million-paper OpenAlex literature corpus provided citation counts and publication trends. AlphaFold citation data (42,000+) is from OpenAlex as of April 2026. 3D chip market share, HBM pricing, and memory shortage data are from Counterpoint Research, Tom’s Hardware, and industry reporting accessed via web search. Apple AR glasses timeline is from Bloomberg reporting and Ming-Chi Kuo’s analyst notes. Vaire Computing’s reversible computing prototype details are from IEEE Spectrum (January 2025). All patent claims quoted are from the USPTO grant text as stored in our database.