Kurzweil Scorecard: Brain-Scale Compute Arrived Early and Nobody Noticed

In a Sandia National Laboratories server room in Albuquerque, a microwave-sized chassis named Hala Point is running 1.15 billion artificial neurons on 1,152 Loihi 2 chips. It draws 2,600 watts — less than a hair dryer. Intel and Sandia say its neuron count is roughly that of an owl’s brain. Meanwhile, on the other side of the country, the El Capitan supercomputer at Lawrence Livermore crossed 1.742 exaflops in November 2024, displacing Frontier. The machine Ray Kurzweil waited twenty years for is already here. It just looks nothing like what he described.

This batch of predictions from The Singularity Is Near (2005) is where Kurzweil staked out his most aggressive compute-hardware bets: analog native-mode neural chips, self-rewiring silicon, functional human-brain simulation by the mid-2020s, and uploading-grade resources — 10^19 operations per second for $1,000 — by the early 2030s. The scoreboard two decades in is strange. The quantity of computation arrived ahead of schedule. The shape of it did not.

The predictions

Kurzweil made twelve claims in this cluster: eight factual assertions about 2005-era compute, one 2010s prediction about analog neural chips, two 2020s predictions about emulation-grade compute and self-organizing hardware, and a capstone 2030s uploading prediction. They interlock — the emulation claim depends on the doubling claim; the uploading claim depends on both.

Kurzweil wrote that “the computational capacity needed to emulate human intelligence will be available in less than two decades from 2005” (ch. “The Software of the Brain”). In The Singularity Is Nearer (2024), he revised his own arithmetic downward: “My 2005 calculations in The Singularity Is Near noted 10^16 operations per second as an upper bound on the brain’s processing speed… A range of further research over the past two decades has shown that neurons fire orders of magnitude more slowly — not two hundred times a second, which is around their theoretical maximum, but closer to once a second.” He now pegs the brain at 10^14 operations per second. He notes that “as of 2023, about $1,000 worth of hardware can already achieve this.”

That admission reframes the whole batch. If the target is 10^14, the $1,000 brain has been sitting on a shelf since before he finished the sequel.

Where we actually are

Compute arrived early. Frontier’s HPL score is 1.35 exaflops, El Capitan’s is 1.742 exaflops. Both are 10,000× to 17,000× the revised brain estimate. On the consumer end, a single Nvidia H100 delivers 989 TF/s at FP16 for roughly $25,000–$30,000 in 2025. Divide by price and you’re already a decimal point past Kurzweil’s 2030s milestone for the top end of that range. Price-performance has not doubled yearly, as he claimed in 2005; CPU benchmarks slowed to around 3.5% a year by 2018. But training compute for frontier AI models has doubled every 5.7 months since 2010 — Kurzweil cites this number himself in 2024. The old yearly doubling broke. A faster, narrower doubling took its place, on chips designed for matrix multiplication rather than general-purpose instruction execution.

Analog native mode came back — sideways. Kurzweil wrote that “using transistors in their native analog mode to simulate neural regions can improve capacity by three or four orders of magnitude, as demonstrated by Carver Mead” (ch. “Is the Human Brain Different from a Computer?”). For fifteen years after 2005 this looked like a dead end. Then memristor crossbars and compute-in-memory chips brought it back. US 12,063,052, granted August 2024 to Hewlett Packard Enterprise, describes a crossbar array of memristors programmed with matrix values that performs matrix-vector multiplication in the physical analog domain, with a second crossbar encoding error-correction parity so single-cycle errors can be detected in the analog result itself. US 11,748,609, granted September 2023 to the University of Dayton, covers an analog neuromorphic circuit where resistive memories are trained on-chip by backpropagating error signals as voltages through the same physical synapses used for inference. These are exactly the devices Mead sketched in Analog VLSI and Neural Systems in 1989. They’re being built, at last, because deep learning made matrix multiplication the dominant workload.

But the order-of-magnitude claim has been verified only in specialized workloads. IBM’s NorthPole, presented at ISSCC 2024, achieves 25× better energy efficiency than a 12 nm GPU on ResNet-50 image classification, and 72.7× better energy efficiency than the next-lowest-latency GPU on a 3-billion-parameter language model. Hala Point reports deep-network efficiency up to 15 TOPS/W. Three-to-four orders of magnitude across general computing? Not yet. One-to-two orders of magnitude for specific neural workloads? Measured.

Neuromorphic patent activity confirms the pivot. Grants mentioning neuromorphic chips climbed from single digits a year through 2012 to 62 in 2024. The assignee list reads as an industry roadmap: IBM (84 grants since 2020), Samsung (47), TDK (20), SK Hynix (19), Intel (18), HRL Laboratories (13). Samsung’s US 11,620,505, granted April 2023, describes a package of neuromorphic chips in a systolic array that sequentially passes weights between neighbors — a physical implementation of the dataflow pattern Google’s TPU made famous, now shipped with analog-synapse hardware. Seoul National’s US 12,099,919, granted September 2024, builds flash-memory synapses into a cross-bar with forward and backward neurons so the network can train without leaving the chip.

Self-organizing hardware didn’t happen. Self-organizing software did. Kurzweil predicted that “self-organization like the brain’s rewiring will be implemented in hardware” by the 2020s (ch. “Is the Human Brain Different from a Computer?”). Grants for “self-healing” systems filed in the last five years are almost entirely about IT operations — JPMorgan’s US 12,411,726 on multi-tier network resiliency, Intel’s US 12,399,781 on self-healing mechanisms in cloud infrastructure, Bank of America’s US 12,007,832 on load-switching to an alternative cloud instance. The adaptive-rewiring property moved up the stack, from transistors to Kubernetes.

Brain emulation is stuck on data, not silicon. The Human Brain Project, a €607 million EU flagship, wound up in September 2023. Its external reviewers praised the EBRAINS infrastructure, 3,000 publications, and 160 digital tools — and quietly did not claim a whole-brain simulation. The State of Brain Emulation Report 2025 puts the bottleneck plainly. C. elegans and adult Drosophila now have fully proofread connectomes; a 2024–2025 result embedded an adult fruit-fly whole-brain model (125,000 neurons) in a physics-simulated body and reproduced feeding and grooming. Mouse cortex connectomics has reached 1 cubic millimeter — 120,000 neurons, 523 million synapses — out of roughly 70 million neurons total. Fewer than 500 people globally work on brain emulation full-time. Kurzweil’s compute forecast beat the field. The rate-limiting step turned out to be experimental data from electron microscopes and calcium imaging, not floating-point operations.

The scorecard

What Kurzweil missed (and what he nailed)

The right pattern to see in this batch is not optimism versus pessimism. It is a forecast that was directionally correct and mechanistically wrong. The brain-scale computer exists. It is plural. It is cheap. And almost every specific architectural detail Kurzweil attached to the prediction — yearly doublings of general-purpose performance, analog native-mode transistors as the efficiency path, hardware that rewires itself to solve problems — either didn’t happen or got replaced with something else. Digital systolic arrays, HBM stacks, and mixed-precision matrix engines carried the water that analog neurons were supposed to.

One consolation prize: his newer revision, the one in The Singularity Is Nearer, is quietly closer to the reality he expected. When he lowers the brain to 10^14 operations per second and admits that $1,000 can already do it, he is describing a world that makes the 2030s uploading milestone look not obviously wrong on compute — only on biology, ethics, and the eleven levels of emulation fidelity that Sandberg and Bostrom catalogued in 2008. Those are the interesting constraints now. The FLOPS argument is over.

Method note

Patent landscape counts came from a full-text index over 9.3 million U.S. grants and pre-grant publications. Key patents were read in full — title, abstract, claims, description — to confirm what they actually claim. Literature signal came from a 357-million-record index of scholarly works, filtered to high-impact papers on Loihi, NorthPole, and memristor crossbars. Price-performance and supercomputer figures come from the November 2024 TOP500, Intel’s Hala Point announcement, IBM Research’s NorthPole publications, and Kurzweil’s own updated estimates in The Singularity Is Nearer. Brain-emulation status draws on the State of Brain Emulation Report 2025 and the final Human Brain Project review of November 2023.