In November 2025, El Capitan at Lawrence Livermore clocked 1.809 exaflops on the Top500 — about 5,000 times faster than the Blue Gene/L that Ray Kurzweil cited in 2005 as the machine approaching brain scale. Twenty years, three and a half orders of magnitude. That is the kind of number that flatters a forecaster.

But walk into an electronics store. The fastest consumer graphics card you can buy, the RTX 5090 at $1,999, reaches roughly 105 teraflops in 32-bit math — about one percent of what Kurzweil predicted the personal computer of 2025 would deliver. The exponential is intact. The shelf label is not.

This batch is Kurzweil’s hardware roadmap: eight predictions, all in the “computing” category, covering everything from transistor shrink cadence to ambient devices to the thermodynamics of price. The pattern is sharp and uncomfortable: he got the direction right almost every time and the calendar wrong just as often.

The predictions

All eight come from chapters in The Singularity Is Near (2005) where Kurzweil lays out the physical substrate for everything else in the book — without the hardware, there is no strong AI, no uploaded mind, no merger. He was betting on silicon to keep delivering.

He wrote that “Personal computers will achieve about 10^16 calculations per second by 2025 based on exponential computing trends” (ch. “Accelerating the Availability of Human-Level Personal Computing”). He claimed feature sizes were “shrinking by half every 5.4 years in each dimension, doubling elements per square millimeter every 2.7 years” (ch. “Moore’s Law and Beyond”). He projected that IBM’s Blue Gene/L would deliver “360 trillion calculations per second and about 10^15 bits of main storage” (ch. “The Computational Capacity of the Human Brain”). And on ambient computing: “By the second decade of the twenty-first century, computing will be highly distributed throughout walls, furniture, clothing, bodies, and brains” (ch. “Setting a Date for the Singularity”).

In The Singularity Is Nearer (2024), Kurzweil updated his accounting: “A roughly $900 (2023 inflation-adjusted) computer chip in 1999 could perform more than 800,000 computations per second per dollar. By early 2023 a $900 chip could do nearly 58 billion computations per second per dollar” — a 72,500-fold improvement in 24 years. That doubling cadence is the engine powering every other prediction here.

Where we actually are

Blue Gene/L and the supercomputer ceiling. Kurzweil’s 2005 number was correct — IBM’s Blue Gene/L did hit 360 teraflops. What he did not predict was how far the ceiling would move. El Capitan now sits at 1.809 exaflops on AMD Instinct MI300A nodes; Frontier is at 1.35 exaflops. The second-place machine today is 3,750 times faster than Blue Gene/L was. This is the prediction where the substrate not only delivered but outran its own schedule. Ahead of schedule.

Personal computers at 10 petaflops. This is the bruising one. An RTX 5090 hits 419 teraflops at FP8 dense, and 1.676 petaflops at FP4 with sparsity. NVIDIA’s DGX Spark, marketed as an AI desktop, hits one petaflop at FP4 and costs around $3,999. None of these numbers reach 10^16 “calculations per second” in any precision strict enough for the comparison Kurzweil intended. In FP32, the gap is two orders of magnitude. In FP4 AI precision, we are roughly six times short. The direction is right and the curve has not stalled — it has just not yet crossed the line he drew for 2025. Behind schedule, closing the gap at AI precision.

Feature-size shrink. Kurzweil’s 2005 cadence was half the linear dimension every 5.4 years. TSMC began volume production of its 2nm N2 node in Q4 2025, using gate-all-around nanosheet transistors for the first time — a 15 percent performance gain at the same power versus N3E, and roughly 15 percent denser. From the 90nm node in 2005 to N2 in 2025 is about 45x on the label, which would match Kurzweil’s rate. The problem is the label. Foundries long ago stopped naming nodes by actual feature size; real contacted gate pitch has shrunk far less than the number on the wafer box implies. And the cost per transistor, which used to fall with each node, has gone the other way: by one accounting, per-transistor cost at 3nm is $2.16, a level last seen around 2005. The transistor is still arriving; the dollar stopped following it to the fab. Verified on the marketing number, behind on the physical geometry, wrong on the economics. Call it verified in letter, slowed in spirit.

MRAM enters the market. This one Kurzweil called cleanly. Everspin shipped the first commercial MRAM parts around 2006, and spin-transfer-torque MRAM is now offered as embedded memory by GlobalFoundries, Samsung, TSMC, and UMC. Samsung demonstrated an 8nm logic-compatible 128Mb eMRAM for automotive; TSMC has 22nm and 16nm eMRAM in production with 12nm and 5nm in the pipeline. The 2025 MRAM market is forecast at $912M growing to $4.8B by 2035. Our patent database shows steady flow: 53 granted US patents in 2025 alone mentioning magnetoresistive random-access memory, including US 12,588,216 (IBM, March 2026), which embeds an MRAM pillar directly into a backside power rail — a design that would have been science fiction in 2005. US 12,376,498 describes a spin-orbit-torque memory cell with an MTJ pillar and a selector on opposite sides of the SOT layer, fabricated into logic wafers. Embedded MRAM is quietly replacing NOR flash at every advanced node. Ahead of schedule.

Spintronics, cobalt-doped silicon-iron (March 2004). Historical claim. Verified at the time. The relevant question today is whether the broader bet paid off, and it has: every eMRAM mentioned above traces back to that class of research. Verified historical, extended.

NTT 10nm electron-beam 3D lithography. Another historical claim. Also verified at the time. Electron-beam lithography did not become the mainstream path to 10nm and below — EUV did. But the underlying capability NTT demonstrated — arbitrary 3D sub-10nm patterning — is now routine in research lines and has informed how features are stitched at advanced nodes. Verified historical, right direction, wrong mechanism to mass production.

Price-performance, 10M–100M factor by 2030. From Kurzweil’s own updated accounting, 1999 to 2023 delivered roughly 72,500x improvement in computations per second per dollar. Extrapolating his cited doubling of ~1.4 years forward to 2030 adds maybe another 30x. That puts the 2005-to-2030 factor in the range of 5,000x–50,000x — three or four orders of magnitude short of his 10 million to 100 million target. The trend is unambiguously real. The multiplier is not on pace. Behind schedule.

Computing everywhere, in walls and clothing and bodies, by the 2010s. Ambient computing did arrive, but not where Kurzweil looked for it. The phone, the watch, the earbud, the car — these absorbed the computing load that he expected would migrate into furniture and garments. Amazon’s Alexa+ rollout in 2025 drew early complaints of reliability and hallucinations; global smart speaker and display shipments are forecast to decline in 2025 and 2026. Computing is indeed distributed throughout daily life. The walls stayed quiet. Arrived through a different door.

The scorecard

What Kurzweil got and what he missed

The hits on this list are the ones about specific pieces of silicon: a named machine at a named throughput, a named memory technology entering the market. When he pointed at hardware, he was usually right, sometimes early.

The misses are a pattern. Every time Kurzweil projected a consumer number — price per computation, flops on a desktop, ambient presence of devices — he was optimistic by roughly one order of magnitude on the calendar. The exponential is real. The friction between the exponential and the retail shelf is larger than his 2005 model accounted for: packaging costs, EUV capital intensity, yield on leading nodes, the end of Dennard scaling, and the stubborn fact that consumer wallets did not grow at 1.4x per year.

The interesting prediction on this list is “computing everywhere.” He was right that computing would become ambient. He was wrong about where it would hide. It did not move into walls; it moved into pockets, then wrists, then ears. The substrate was ready by 2015. The form factor humans wanted turned out to be different from the one he imagined in 2005. That is a pattern worth keeping in mind when reading his predictions about the 2030s.

Method note

Eight predictions from the source text were scored against three inputs: recent product and market data gathered from current industry reporting, a sample of US patent grants from our patent corpus that concretely illustrate where hardware claims are being filed today, and Kurzweil’s own 2024 update, which restates the underlying price-performance trend in inflation-adjusted terms. Patent counts by year and individual patents were pulled from a local copy of the US patent grant record.