New analog silicon IS shipping for NN inference — and it bets on capacitors, not memristors

The seed splits into two lineages the hype conflates. Part A (analog for niche math) and Part B (analog for NN inference) are different revivals — different companies, different mechanisms.

Claim 1 — The genuine analog niche is differential equations, not AI. The serious primary literature on modern reconfigurable analog computers (Anabrid's LUCIDAC/REDAC) is about dynamical systems, and says nothing about neural networks. Its measurable edge: "the solution time on an analog computer is basically constant, regardless of the number of computing elements required," whereas digital solution time "typically grows with problem size, often much worse than linearly." (arXiv 2510.25942, Reconfigurable Analog Computers, Tier 1) [unverified-quant — no joules/seconds given, claim is asymptotic-shape only]

Claim 2 — New analog inference silicon shipped in 2025, and it avoids memristors. EnCharge AI's EN100 (spun from Naveen Verma's Princeton lab) does in-memory matrix-multiply with metal-wire switched capacitors, not resistive memory. Why: RRAM conductances drift — "you could set the exact same voltage on two cells... and those two cells would wind up with slightly different conductance values." Capacitors escape this because "the only thing they depend on is geometry, basically the space between wires... the one thing you can control very, very well in CMOS." Claimed 200 TOPS at 8.25 W, "performance per watt up to 20 times better than competing chips." (IEEE Spectrum, Tier 2; numbers are company-sourced.)

Claim 3 — Analog AI is a graveyard, not a clean win. Mythic (flash-transistor analog) "ran out of runway" and out of cash in 2022, revived on $13M in 2023, raised $125M in Dec 2025. Efficiency claims have not yet beaten digital's commercial gravity. (The Register / EE Times, Tier 3.)

Why this was hop-worthy

It answers the seed literally (yes — EnCharge EN100), and inverts the expected mechanism: the leading new analog-AI chip wins by being more deterministic than memristor analog — precision defined by lithographic geometry, the digital fab's home turf.

Further leads

• LUCIDAC "analog computer on a chip" (Ulmann/SpriND) — the differential-equation lineage as a co-processor.
• The capacitor-vs-conductance precision argument may bridge to the vault's SIREN / continuous-signal notes.

Hop chain

Hop 1 — 2020 Embedded Processor Report / Anabrid recon (web recon, https://anabrid.com , https://www.sprind.org/en/impulses/projects/ulmann)

• Hook type: Unfamiliar name + cross-domain bridge (old→new)
• Hook: "Anabrid / LUCIDAC — a modern commercial analog computer in the 2020s"
• Why followed: Bridges the vault's historical analog cluster (hydraulic computers, Lukyanov's 1936 Water Integrator) to a live 2020s product — highest-value unlinked-pair bridge.
• Key findings: Anabrid (Bernd Ulmann) ships THAT, LUCIDAC, REDAC — reconfigurable analog hardware aimed at differential equations, explicitly framed as co-processors.

Hop 2 — Reconfigurable Analog Computers (arXiv 2510.25942, Oct 2025, https://arxiv.org/html/2510.25942)

• Hook type: Mechanism question
• Hook: "Where, exactly, is digital measurably worse?"
• Why followed: Needed a Tier-1 statement of the niche rather than listicle hand-waving.
• Key findings: The rigorous claim is solution-time-independent-of-problem-size for coupled/nonlinear ODEs; the paper never mentions neural networks. The AI-inference story is a separate lineage.

Hop 3 — EnCharge's Analog AI Chip (IEEE Spectrum, https://spectrum.ieee.org/analog-ai-chip-architecture)

• Hook type: Surprising claim + the person behind the thing (Naveen Verma)
• Hook: "EnCharge uses capacitors, not memristors"
• Why followed: The road home to AI, and a mechanism that inverts the usual analog-AI (ReRAM) story.
• Key findings: Switched-capacitor in-memory compute; precision from wire geometry, not material conductance; EN100 shipped 2025 for laptops/edge.

Hop 4 — Mythic runs out of money / rises again (The Register, EE Times, TechCrunch)

• Hook type: Cultural/business resonance (rise-from-the-ashes) + surprising claim
• Hook: "The pioneer nearly died"
• Why followed: Test the seed's premise — if digital is measurably worse, why did the flagship analog-AI startup go broke?
• Key findings: Mythic ran out of cash 2022, survived, and only reached a $125M round in Dec 2025; the field is "very crowded." Efficiency ≠ commercial victory.

Saved hooks not followed:

• Analog signal processing with microwave networks (arXiv 2504.06790) — computing physically inside RF network scattering — from seed recon — a third, wilder analog lineage.
• IBM analog AI / HfOx ReRAM on-chip training (arXiv 2502.04524) — the resistive-memory camp EnCharge is betting against — worth a head-to-head.
• Nicholas Saunderson's "palpable arithmetic" (already in vault) — tactile analog computation, a possible bridge partner.

Surprise: expected the newest analog NN-inference chip to use memristors/ReRAM (the canonical analog-AI story) — found the leading 2025 product (EnCharge EN100) deliberately avoids resistive memory and uses lithography-defined metal-wire capacitors because material conductance is too noisy. Surprise: expected "digital is measurably worse" to mean analog is winning — found analog AI is largely a commercial graveyard (Mythic ran out of cash in 2022) despite standing efficiency claims. Surprise: expected the modern analog-computer revival to center on AI inference — found the serious primary literature (reconfigurable analog computers) is about differential equations and never mentions neural networks; "analog revival" is two separate lineages the press conflates.

post-worthy: maybe — the capacitor-vs-memristor inversion plus the two-separate-lineages framing is a genuinely non-obvious throughline, but the quant claims are company-sourced and would need a primary benchmark before publication.