Kurzweil Scorecard: Silicon Held. The Plumbing Didn’t.

In The Singularity Is Near, Ray Kurzweil placed two confident bets about the substrate of the digital age. The first: by the 2030s, computers would be built out of carbon nanotubes — silicon, he wrote, was “unlikely” to remain the dominant material. The second: the Internet would keep getting more resilient, because traffic naturally “routes around failed hubs or channels.”

Two decades later, both bets look upside down. Silicon didn’t just survive — it ate the alternatives. And the Internet, far from routing around failures, has spent the last eighteen months going dark in spectacular, cascading ways. The CrowdStrike update that bricked 8.5 million Windows machines in a single morning was not the system Kurzweil described in 2005.

This is a scorecard for five interconnected predictions about the engine room of the Singularity: the chips, the networks, the brain replicas, and the law that supposedly drives them all forward.

The predictions

The five claims in this batch were Kurzweil’s structural bets — the load-bearing assumptions underneath the rest of the book. If silicon scaling died early, if the Internet stayed brittle, if neural replication stalled, the whole timeline would slip.

The good news for Kurzweil is that the headline trend — exponentially cheaper computation — has held up. The bad news is that the mechanisms he named for how that trend would continue are mostly wrong. The story of computing since 2005 is a story of how the industry kept Moore-style price-performance gains alive by abandoning almost every specific path Kurzweil predicted it would take.

Where we actually are

Carbon nanotubes lost. Silicon learned new tricks.

Kurzweil wrote that “the dominant substrate for future nonbiological intelligence is unlikely to be silicon and is more likely to be something like carbon nanotubes” (ch. “Ich bin ein Singularitarian”). In the 2024 update, he concedes the field has shifted — citing “exciting new breakthroughs in graphene, carbon nanotubes, and carbon nanothreads” but framing them as long-horizon material science rather than the near-term silicon successor he originally pitched.

Patent activity tells the story bluntly. Carbon nanotube transistor grants peaked at 28 in 2013 and have collapsed to single digits — five in 2025, one so far in 2026. The most recent serious CNT logic patent in our index, US 11,121,335, granted in 2021, describes carbon nanotube field-effect transistors with end-bonded metal contacts. It is a research-grade fabrication technique, not a manufacturing roadmap.

Compare that to chiplets and advanced packaging — silicon’s actual escape hatch from the limits of monolithic scaling. Chiplet patents went from 23 grants in 2021 to 103 in 2025, a 4.5x climb in four years. Just last week, US 12,615,209 was granted for “inter-chiplet routing of transactions across multi-heterogeneous chiplets,” and US 12,616,013 for a “dielectric bondable chiplet for package architecture.” Extreme ultraviolet lithography patents — the technology that keeps silicon transistors shrinking past the 3 nm node — held steady at 176 grants in 2025.

The clearest tell that the industry has moved on from the carbon-substrate dream is the Stanford-MIT-CMU-Penn 2024 collaboration with SkyWater Technology: a monolithic 3D chip that integrates carbon nanotubes with conventional silicon CMOS and resistive RAM. CNTs ended up as a stacked layer inside a silicon package, not a replacement for it. Verdict: wrong mechanism, right destination. The substrate stayed silicon; the geometry went 3D.

The Internet stopped routing around failures.

Kurzweil’s claim that “the Internet had not been substantially disrupted by 2005, and as it continues to grow its robustness and resilience continue to strengthen because information routes around failed hubs or channels” (ch. 8, “Promise and Peril of GNR”) was correct in 2005. It is no longer correct.

On July 19, 2024, a single configuration update from CrowdStrike crashed 8.5 million Windows endpoints worldwide in what is now described as the largest IT outage in history. Airlines grounded over 5,000 flights. Hospitals reverted to paper. Direct losses to Fortune 500 companies hit $5.4 billion, with healthcare ($1.94B) and banking ($1.15B) hardest hit. Information did not route around the failure — it had nowhere to go, because the failure was inside the operating system kernel of every affected machine.

Then came the cloud cascade. On October 20, 2025, AWS’s us-east-1 region went down for hours due to a DNS-DynamoDB failure. Nine days later, Microsoft Azure went critical across every region on Earth. On November 18, 2025, an internal database permissions change at Cloudflare ballooned a bot-management config file past an internal limit and crashed the proxy infrastructure that fronts an estimated 20–25% of the public web. Cloudflare had a second outage on December 5.

These are not the failures Kurzweil’s 2005 model predicts. They are the failures of a centralized network — one where four companies (Cloudflare, AWS, Akamai, Fastly) are the chokepoints for most traffic, and a bad config at any one of them cascades globally. The Internet got faster, fatter, and more useful. It did not get more resilient. Verdict: wrong direction.

Computation kept getting cheaper. Just not the way he said.

Kurzweil writes in The Singularity Is Nearer that “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 independent Epoch AI dataset broadly agrees: GPU performance per dollar has been improving roughly 30% per year, doubling every 2.5 years. Computational price-performance jumped 2x from the A100 to the H100 in just 16 months. The headline curve bends the right way.

But “exponential growth in price-performance, capacity, and rate of adoption” was Kurzweil’s general-purpose bet, and the way it has held up is via specialization, not the universal substrate gains he described. The doublings live in tensor cores, HBM stacks, and rack-scale interconnect — not in a successor material to silicon. Verdict for information_technology_continues_exponential_growth: on track on outcomes, wrong on mechanism.

The companion claim that “the overall pace of technology trends has continued despite world wars, the Cold War, and major economic, cultural, and social upheavals” (ch. “Ich bin ein Singularitarian”) looks vindicated. COVID, the chip-war export controls, the Russia-Ukraine semiconductor supply shock — none of them dented the FLOPS-per-dollar curve. Verdict: verified.

Neural replicas are real, and they are scaling.

Kurzweil predicted that “the resolution and accuracy of attempts to create neural and body replicas will improve at the same exponential pace as other information-based technologies” (ch. “Who Am I? What Am I?”). The evidence here is the strongest in the batch.

Connectomics literature in our index has grown from a single paper in 2005 to 600 in 2025 — three orders of magnitude. The capstone is the FlyWire consortium’s October 2024 Nature paper, “Neuronal wiring diagram of an adult brain” (DOI 10.1038/s41586-024-07558-y), which mapped all 139,255 neurons and over 50 million synapses in an adult Drosophila brain. The same issue identified 8,453 distinct cell types, 4,581 of them new. The paper has been cited 447 times in less than two years.

This is not a human brain. Drosophila has roughly 0.0014% of our neurons. But the resolution and accuracy curve Kurzweil predicted is real, and steeper than most observers expected. Verdict: on track.

The scorecard

What Kurzweil missed (and what he nailed)

The pattern across this batch is a forecaster who got the trajectory right and the mechanism wrong. Compute kept getting cheaper, but through chiplets and EUV rather than carbon. Neural replication scaled, but flies came before humans. Where he was clearly wrong was where the prediction depended on a self-organizing property of distributed systems — the Internet was supposed to keep getting more resilient because failures would route around themselves. Instead, the network consolidated into a handful of providers whose outages now define what “the Internet is down” means.

The honest read is that 2005-era Kurzweil underweighted centralization as an attractor. A network where any node can be replaced survives the loss of any node. A network where 25% of HTTP traffic depends on a single company’s bot-management config file is not. The Singularity may still arrive on something like the schedule he predicted — the price-performance curve is doing its job. But it is arriving on top of a substrate (silicon plus chiplets) and a network (cloud-centralized) that look almost nothing like the 2005 vision.

Method note

We pulled patent grant counts and full text from a 9.3-million-document U.S. patent index, joined to assignee and CPC tables. Literature trends came from a 357-million-paper academic corpus indexed for full-text search. Outage figures came from CISA, IBM, Cybersecurity Dive, and post-mortems published by Cloudflare, AWS, and Microsoft. The price-performance numbers were cross-checked against Epoch AI’s GPU dataset. Quotes from The Singularity Is Nearer are from the 2024 print edition.