The honest TRL of post-CMOS: the gap between headline qubit counts and the “when is it useful?” question

Evidence-first notes on bioscience and deep tech, at the edge of the lab and the market. Information only — not investment advice.

The 30-second version

• What. For post-CMOS alternatives (quantum, neuromorphic, photonic and others), the real question is not “does it work?” but “when does it become useful?” The estimated resources to break RSA-2048 have fallen roughly 20-fold on paper over six years (from ~20 million physical qubits in 2019 to under 1 million in 2025) — but this is an algorithmic advance, not a hardware realization, and it remains 3–4 orders of magnitude away from real qubits.
• So what. Willow’s below-threshold result (Λ = 2.14) proved that “error suppression scales” — it is not a useful computation. Falling numbers read easily as “imminent,” but every estimate is a projection resting on not-yet-achieved assumptions (0.1% gate error, 1 µs cycle time).
• Now what. Real hardware (IBM Nighthawk, 120–360 qubits) and photonics (a 1×4 microring proof-of-concept) are still at the demo / emulation stage. An honest date for useful fault-tolerant (FT) quantum computing is the 2030s, and the metric to watch is not headline qubit counts but noise, overhead × time, and the software ecosystem.

[demo-gap note] Willow is real hardware, but at the scale of one logical qubit and not a useful computation. Neuromorphic, photonic and thermodynamic approaches are heavily demo/simulation; in photonics, only a 1×4 microring weight bank has actually been fabricated for optical compute, and useful scale is emulation only.

The five-minute read

The large gap between a real milestone and usefulness

Post-CMOS discussion usually stops at “does quantum work?” But one layer in — “when does it actually become useful?” — the picture changes. Quantum has crossed a genuine milestone in below-threshold operation (Λ = 2.14), yet from there to a useful fault-tolerant (FT) computation lies a large gap created by the product of physical qubit count, time, and distillation overhead. The through-line: headline metric ≠ usefulness.

The estimates shrink, but the hardware does not

The estimated cost of breaking RSA-2048 has fallen sharply, from Gidney–Ekerå 2019’s 20 million physical qubits over 8 hours to Gidney 2025’s under 1 million noisy qubits in under a week, and the 2026-02 Iceberg “Pinnacle” claims under 100,000 via qLDPC. But this reduction is an algorithmic advance — trading fewer qubits for more time — not a hardware improvement. The driver is the collapse of magic-state overhead: producing non-Clifford (T) states has historically been the dominant cost of FT (about 94% by Fowler 2012), and Gidney’s “magic state cultivation” may lower this to roughly the cost of a single CNOT. That said, at realistic noise (1e-3) it still falls about 10× short of the fidelity RSA-2048 requires — a noise-specific shortfall.

Real devices are 3–4 orders below

IBM’s 2026 flagship Nighthawk is a single 120-qubit module (roadmap target of 3 modules, ~360 physical qubits). Against the 10⁵–10⁶+ needed for useful FT, that is 3–4 orders of magnitude short — meaning that, independent of the estimation game above, real qubits are overwhelmingly lacking. Photonic optical compute is similar: only a 1×4 microring weight bank (90 nm GF9WG) has been built and tested, and useful scale (~40,000 weights) is emulation only.

[diagram: paper estimates vs. real hardware gap]

  Qubits needed to break RSA-2048 (log scale)
  2019 estimate ██████████████████████ 20,000,000  (Gidney-Ekera)
  2025 estimate ████████████            <1,000,000  (Gidney, cultivation)
  2026 estimate ██████                  <100,000    (Iceberg qLDPC claim)
  ─────────── theoretical projection (not-yet-achieved assumptions) ───────────
  Useful-FT line ▓▓▓▓▓▓▓▓▓▓▓ ~10^5–10^6
  IBM real 2026 │ 120–360 qubits  ← 3–4 orders below

Deep dive

1. Background — the real question for post-CMOS

Where Part 0 §4 mapped “what each alternative is and roughly what TRL,” this part asks one layer deeper: when does it actually become useful? Quantum in particular has crossed a genuine milestone in below-threshold operation (Λ = 2.14), but from there to useful FT computation lies a large gap of physical qubits, time and distillation overhead. The through-line — headline metric ≠ usefulness — holds most dramatically in post-CMOS.

2. What this deep dive newly establishes

Core answer: the RSA-2048 estimates have fallen roughly 20-fold on paper over six years (thanks to algorithms, not hardware), but every one rests on not-yet-achieved assumptions and sits 3–4 orders of magnitude from real hardware.

• Sharp drop in estimates (on paper) — CONFIRMED (3-0): 20M physical qubits / 8 hours (2019) → under 1M noisy qubits / under a week (2025) → under 100k via qLDPC claim (2026-02 Iceberg). All are theoretical estimates dependent on not-yet-achieved assumptions (uniform 0.1% gate error, 1 µs surface-code cycle, 10 µs reaction). The reduction is an algorithmic advance — fewer qubits traded for more time.
• Driver = collapse of magic-state overhead — CONFIRMED (3-0): producing non-Clifford (T) magic states has historically been the dominant cost of FT (~94% Fowler 2012, 30–70% Litinski 2018). Gidney’s “magic state cultivation” may make T states as cheap as a single CNOT, eliminating multi-round distillation (+ yoked surface code, approximate-residue arithmetic → >100× fewer Toffolis vs. CFS 2024). Attributed to vendor/author claim.
• Still short — CONFIRMED (3-0): cultivation at realistic noise (1e-3) still falls about 10× short of the fidelity RSA-2048 requires (reaches 2e-9 vs. needed ~2e-10). Halving noise to 5e-4 reaches 4e-11 and suffices — a noise-specific shortfall. Algorithms run ahead; hardware noise is rate-limiting.
• Logical-operation “breaking even” demonstrated — CONFIRMED (3-0): Quantinuum first exceeded unencoded physical (1e-3) with an FT controlled-Hadamard (non-Clifford) at logical infidelity ≤ 2.3e-4. But this is an 8-physical-qubit demo — proof of principle, not scale.

3. Strengths and limits of the methodology

Strengths: the sources and assumptions of each estimate are explicitly attributed, and “algorithmic advance” is kept separate from “hardware realization.” Limits: what is confirmed is mostly “the accuracy of what is estimated/claimed,” not the realization of the resource figures themselves. The FT resource estimates (20M → <1M → <100k) are all projections resting on not-yet-achieved assumptions, and falling numbers are easily misread as “imminent.” The qLDPC (generalised bicycle) overhead of ~61:1 to ~101:1 is more favorable than surface code’s ~1,000:1, but long-range connectivity, logical operations and decoding remain unsolved — so the paper advantage must clear those problems before it reaches real devices.

4. Neighbouring domains

No forced hook (in honesty). Post-CMOS connects weakly to CKM (cardio-renal-metabolic). As a non-forced meta-connection, quantum-chemistry simulation is a leading early application of useful FT quantum and could potentially touch drug and catalyst design. But that is a 2030s projection, not imminent, and the bio bottlenecks (data, validation) remain. Only the structural similarity (“headline ≠ usefulness”) is recorded.

5. Commercialization and market context (TRL, companies)

Company context (factual, neutral, no buy/sell implication): superconducting (IBM, Google, Rigetti RGTI), ion-trap (IonQ IONQ, Quantinuum), neutral-atom (Atom Computing, QuEra), photonic (PsiQuantum), annealing (D-Wave QBTS), thermodynamic (Extropic) and others carry different overheads and timelines by modality. No single-winner conclusion is asserted; vendor roadmap dates are all attributed as projections with slip histories noted. Digital compute-near-memory (CNM) is commercialized at UPMEM, Samsung and SK Hynix, whereas analog CIM remains at the research stage.

6. The skeptic’s bottom line

• Adversarial verification complete (25 confirmed, 0 kill), but what is confirmed is mostly “the accuracy of what is estimated/claimed,” not the realization of the resource figures themselves. The FT resource estimates (20M → <1M → <100k) are all projections resting on not-yet-achieved assumptions (0.1% noise, 1 µs cycle). Misreading warning: falling numbers read easily as “imminent,” but this is an algorithmic advance, not a hardware realization, and real devices are 3–4 orders of magnitude short.
• demo-gap: Willow is real hardware, not simulation, but it is one logical qubit and not a useful computation. Neuromorphic, photonic and thermodynamic are heavily demo/simulation.
• Real bottleneck: behind “qubit count / milestone” headlines, the actual rate-limiter is overhead × time (quantum), the software ecosystem (neuromorphic) and demonstration (thermodynamic).
• Prediction-market skepticism: 2026 prediction markets are near-unanimously skeptical of a practical quantum product within the year.

7. What to watch (falsifiable predictions)

1. If any vendor demonstrates a useful (classically superior) FT computation before 2030 under independent verification, the “2030s” outlook moves earlier (falsification of the reinforcing view); if none does, it is confirmed.
2. If a 10×+ reduction in the physical:logical overhead is demonstrated (e.g., via magic state cultivation), the timeline moves earlier.
3. If neuromorphic is deployed in a large-scale production killer app rather than a demo, the “software bottleneck” is confirmed to be easing.

References

• Gidney, Craig. 2025. “How to factor 2048 bit RSA integers with less than a million noisy qubits.” arXiv:2505.15917. https://arxiv.org/abs/2505.15917
• Gidney, Craig, and Martin Ekerå. 2019. “How to factor 2048 bit RSA integers in 8 hours using 20 million noisy qubits.” arXiv:1905.09749. https://arxiv.org/abs/1905.09749
• Iceberg Quantum. 2026. “Pinnacle: qLDPC resource estimation.” arXiv:2602.11457. https://arxiv.org/abs/2602.11457
• Fowler, Austin G., et al. 2012. “Surface codes: Towards practical large-scale quantum computation.” (magic state overhead ~94%).
• Litinski, Daniel. 2018. “A Game of Surface Codes.” (magic state overhead 30–70%).
• PRX Quantum. 2026. “Magic state cultivation on the surface code.”
• Bravyi, Sergey, et al. 2024. “High-threshold and low-overhead fault-tolerant quantum memory.” Nature. (IBM gross code ~24:1).
• IBM. 2026. “Nighthawk roadmap.” (120-qubit module, target 3 modules ~360 qubits — vendor projection).
• Photonic microring weight bank (90 nm GF9WG), 1×4 fabricated and tested.

Source knowledge asset: knowledge-base/deep-dives/computing-power/part4-postcmos-trl.md (generated 2026-07-09, VERIFIED). This draft inherits the figures and source attributions of the original part and creates no new figures or sources.

Disclosure

This post is for information only and is not investment advice. The author holds no position in, and no financial interest in, the listed companies mentioned (IonQ IONQ, Rigetti RGTI, D-Wave QBTS, IBM, Intel INTC, and others).

COI note: This document describes listed quantum and semiconductor companies (IonQ, Rigetti, D-Wave, IBM, Intel and others) and private ones (PsiQuantum, Atom Computing, Quantinuum, Extropic and others) in a factual, neutral technology/roadmap context. There is no buy/sell implication. Vendor roadmap dates and FT resource estimates are all attributed as “vendor/author/preprint claims (projections),” with slip histories noted.