Kurzweil Scorecard: Mead Was Right. The Market Didn’t Care.

In 1989 Caltech’s Carver Mead published Analog VLSI and Neural Systems and made the wager that defined a field he had also just christened. The brain, Mead argued, is not a digital machine. Neurons integrate continuously varying voltages, communicate with stochastic spikes, store weights in the conductances of their own membranes, and do all of this on roughly twenty watts. If you wanted to match that efficiency in silicon, the answer was obvious: build chips that work the same way. He coined a word for it — neuromorphic — and spent the next decade demonstrating single-transistor synapses, floating-gate non-volatile weights, and silicon retinas that processed visual information in analog, on the chip, without ever touching a digital bus.

In 2005, Ray Kurzweil cited Mead as proof his timelines were achievable. The bet, he argued, would pay out: “machines can use the same digital-controlled analog methods used by the human brain” (ch. “A Panoply of Criticisms”). And in the same chapter pair, he insisted on the inverse: “digital computation can simulate analog processes to any desired level of accuracy, whereas analog systems cannot necessarily simulate digital computation equivalently” (ch. “A Panoply of Criticisms”). Mead, he wrote, had already shown that “silicon circuits can implement digital-controlled analog circuits analogous to mammalian neuronal circuits” (ch. “The Criticism from Analog Processing”).

Twenty-one years later, the wager has paid out — and the payout went to the wrong table. The mechanism Mead and Kurzweil pointed at exists, ships, and is patented at an accelerating clip. It runs about half a percent of the global AI chip market.

The predictions

Three claims drive this batch, all from the criticism-and-rebuttal chapters where Kurzweil defended the Singularity Is Near compute model. The first is a technology bet: real machines will use brain-like analog methods. The second is a near-tautology that Kurzweil deployed strategically: anything analog can be approximated arbitrarily well in digital, and the converse does not hold. The third is a name-check of Mead’s lab as proof of the first.

In The Singularity Is Nearer (2024) Kurzweil did not restate the analog claim directly. He did spend a chapter celebrating the cost-performance arc of digital silicon — “a factor of about 2.8 billion” from 1969 to 2016 — and reciting the case for connectionism. The closest he came to revisiting the analog bet is a footnote pointing at three-dimensional memristor crossbar work (Lin et al., Nature Electronics 3, 2020). The book is, in effect, an updated digital manifesto from the man who said in 2005 that digital would always win on flexibility.

Where we actually are

Mead’s mechanism is real, and it is shipping under patent. Strip the word “neuromorphic” out of the corpus and what remains is a population of analog-circuit grants that read like Mead’s late-1980s lab notebook updated for 28-nanometer process. Our patent search pulls 56 neuromorphic grants in 2025 and 30 in the first third of 2026 — up from 2 grants in all of 2010. Among them:

• US 12,361,272, granted July 2025 (“Analog neuromorphic circuit implemented using resistive memories”), claims a network where “each input voltage is multiplied in parallel by the corresponding resistance of each corresponding resistive memory to generate a corresponding current” and the currents are summed by wire. That is Ohm’s law and Kirchhoff’s current law performing a matrix-vector multiplication, with the weights stored as conductances. It is the analog neural circuit Mead sketched, materialized as silicon you can buy a license to.
• US 12,393,833, granted August 2025 (“Systems and methods for optimizing energy efficiency of analog neuromorphic circuits”), specifies “an analog network of analog components including operational amplifiers and resistors” where “each operational amplifier represents an analog neuron, and each resistor represents a connection between two analog neurons.” The patent’s optimization target is energy: it benchmarks against digital implementations and claims orders-of-magnitude gains for inference.
• US 12,347,421, granted July 2025 (“Sound signal processing using a neuromorphic analog signal processor”), is a multi-core analog audio chip explicitly architected as “a plurality of analog neuromorphic cores” with a digital switch routing sound streams between them — Kurzweil’s digital-controlled-analog hybrid, claimed verbatim.
• US 12,586,636, granted March 2026, drops the marketing pretense and titles itself “Analog-digital hybrid computing method and neuromorphic system.” The phrase that made it through the examiner is Kurzweil’s, give or take a word.

The assignee roll call reads like a list of companies that did not bet on neuromorphic in 2005 and now are: IBM (84 grants since 2020), Samsung (47), Intel (18), HRL Labs (13), TDK (20), SK hynix (19), and a long tail of universities. The literature signal is steeper. Papers on memristor crossbar and in-memory analog compute went from 129 in 2015 to 436 in 2025. Total neuromorphic-tagged paper output went from 63 in 2005 to 5,658 in 2025 — roughly 90×, on a base where the global research enterprise grew about 4×.

The largest neuromorphic systems built in 2024–2025 are credible silicon, not slideware. Intel’s Hala Point, delivered to Sandia National Laboratories in April 2024, packs 1,152 Loihi 2 processors into a 2,600-watt chassis the size of a microwave. The system supports 1.15 billion neurons and 128 billion synapses distributed across 140,544 cores, and benchmarks at up to 50× faster inference at 100× lower energy than CPU/GPU baselines for specific workloads (Intel). IBM’s NorthPole, unveiled in late 2023 by Dharmendra Modha’s team, hits 25× higher energy efficiency and 22× lower latency than a same-node 12-nanometer GPU on ResNet-50, and runs a 3-billion-parameter language model at 72.7× higher energy efficiency than its nearest digital competitor (Open Neuromorphic). On the commercial edge, BrainChip launched Akida 2 in August 2025, opened the Akida Cloud for remote evaluation, raised $25 million in December 2025, and announced Neuromorphyx (defense and industrial sensing) and EDGEAI (Korean SoC integration for smart metering) as 2026 customers (BusinessWire, SiliconAngle).

These are not toys. Loihi 2’s energy claims have been independently reproduced in robotics control, optic-flow MAV landing, and EMG gesture recognition; the highest-citation neuromorphic survey in our literature corpus — Davies and colleagues’ “Advancing Neuromorphic Computing With Loihi” (Proceedings of the IEEE, 2021, 584 citations) — catalogs dozens of workloads where event-driven analog-hybrid silicon beat conventional digital by one to three orders of magnitude on energy-per-inference.

The mechanism shipped. The market refused to follow. Outside analysts put the 2025 neuromorphic-and-brain-inspired hardware market at roughly $50 million in commercial revenue — projected to reach $185 million by 2030 — against a global AI chip market estimated between $30 and $95 billion. That is, by the high end of the analyst range, 0.05% to 0.17% of the AI hardware economy (ResearchAndMarkets via BusinessWire). Hala Point is a research instrument at a national lab. NorthPole has no announced commercial deployment. BrainChip’s customer list runs to two named names. Every other line in the 2025 datacenter capex spreadsheet is dense digital matrix multiplication on NVIDIA Blackwell, Google TPU, or AMD Instinct.

The diagnosis offered by every market analyst we found is unanimous: it is software, not silicon. There is no CUDA for spiking networks. Pytorch and JAX assume dense float operations on regular tensor shapes. Compiling a transformer onto an event-driven analog substrate requires a research project per model. Backpropagation does not natively map to analog dynamics; the field’s most-cited recent paper, Eshraghian and colleagues’ “Training Spiking Neural Networks Using Lessons From Deep Learning” (Proceedings of the IEEE, 2023, 613 citations), is the field admitting it must import the digital toolkit to train its own hardware. Mead built the substrate. The compiler stack never came.

Kurzweil’s second claim — digital can simulate analog to any accuracy — was vindicated more completely than he could have anticipated. In 2005 the claim was a defensive flourish, an argument that he didn’t need analog silicon for his timelines because digital would catch up. By 2026 it is a statement about the entire AI economy. Frontier transformers running on NVIDIA Blackwell simulate the cognitive functions Mead’s analog chips were designed to embody — pattern recognition, sequence prediction, sensory integration — at scales Mead never imagined and through a mechanism Mead would not have predicted. The simulation is so complete that the thing being simulated lost the market. Twenty trillion floating-point operations per second of dense digital matrix multiplication does the work that twenty watts of biological wetware does at home, and the architectural cost of doing it the brain’s way has not justified the energy savings except at the edge.

The scorecard

What Kurzweil missed (and what he nailed)

The Mead-Kurzweil thesis was that brain-like computation would be a category of hardware — a sibling to the digital line, eating share over time as efficiency demands grew. The post-2010 history says brain-like computation is, instead, a category of abstraction, instantiated on whatever silicon turns out to be cheapest. Transformers borrow the layer-and-weight idea, the activation-function idea, the back-propagation idea, and the local-update idea from connectionism. They do not borrow the analog mechanism, the event-driven communication, or the in-memory weight storage. They run on chips that descend from graphics pipelines and systolic arrays designed for radar.

That is the deeper pattern across these scorecards: Kurzweil was usually directionally right about the destination and almost always wrong about the mechanism. Brain-like processing won; brain-like hardware did not. The thing called intelligence is being produced by silicon that, viewed from inside, would baffle a 1989 Caltech graduate student — and would also confirm that graduate student’s professor was correct about everything except who would write the checks.

Mead’s mechanism is patient. Memristor crossbars do not need to displace GPUs to matter. They need to win the edge — the watch, the cochlear implant, the always-on industrial sensor — where 100× energy efficiency on a specific inference workload is the binding constraint and there is no datacenter to lean on. BrainChip’s Akida 2 selling into smart metering and defense sensing is the only story where the analog bet pays off on its own terms, not on the terms Kurzweil set in 2005.

Method note

We took the three Kurzweil claims, pulled the named-entity targets (Mead, analog VLSI, digital-controlled analog), and ran them against the US patent grant corpus from 2010 through April 2026, the global scientific literature corpus through early 2026, and a set of vendor and analyst sources for commercial revenue and product status. Patent abstracts and claims were read directly for the four named patents. Paper findings were taken from titles, abstracts, and citation counts in the literature corpus. Market sizing combines ResearchAndMarkets, Polaris, and a 2026 industry-commissioned capital-allocation report cited by GlobeNewswire. All numbers above were retrieved during this session.