What is NISQ, and are we leaving that era?

What did NISQ mean when Preskill coined it?

A diagnosis, not a product tier. In 2018, John Preskill named the era the field was entering: Noisy Intermediate-Scale Quantum (Quantum 2, 79). "Intermediate scale" meant machines of 50 to a few hundred physical qubits — large enough that brute-force classical simulation of the device stops being feasible. "Noisy" was the operative word: these machines run without quantum error correction, so every gate leaks error into the state, and usable circuit depth is capped at roughly the inverse of the physical error rate. With two-qubit gates failing about once per thousand operations, circuits more than a few hundred gates deep return noise, not answers.

The nuance the marketing usage buried: NISQ was coined as something to exit. Preskill's 2018 paper framed the era as scientifically interesting and commercially unproven, and his 2025 retrospective restates the diagnosis without cosmetics: "these quantum machines are not error-corrected, and noise severely limits their computational power" (arXiv:2502.17368). Treating "NISQ device" as a stable product category inverts the intent of the word.

What would leaving NISQ actually require?

That error correction — not error mitigation — is what makes circuit outputs reliable. The distinction decides the era. Mitigation runs the noisy circuit many times and post-processes the results classically; its cost grows exponentially with circuit size, so it stretches NISQ rather than exiting it (mechanism in our error-mitigation explainer). Correction encodes one logical qubit across many physical qubits, and — if the hardware sits below the code's error threshold — the logical error rate falls exponentially as the code grows.

That is why the exit is measured in logical qubits and achievable depth, not physical-qubit counts (what is a logical qubit). A 1,000-physical-qubit machine with no correction is still a NISQ machine. A 98-physical-qubit machine running 48 logical qubits is, for those 48 qubits, something else.

What has actually been measured toward the exit?

Three dated measurements carry most of the weight; everything else is roadmap.

The table carries its own two flags. Peer-reviewed rows and vendor-reported rows are not the same grade of evidence, and the last row is not evidence at all — it is a plan with a date. Rosetta Q lists roadmaps in gold, next to measurements, never mixed with them.

Where is the exit bar?

Preskill drew it himself in February 2025: the megaquop machine — a computer that can execute about one million reliable quantum operations, which he sizes as circuits of order 100 logical qubits at depth of order 10,000, at roughly 10⁻⁶ error per logical operation (arXiv:2502.17368). His resource guess: "tens of thousands of high-quality physical qubits could suffice." His timeline: "may be realized soon" — with no date attached.

Against that bar, the largest measured logical-qubit count on any machine today is 48, and the strongest peer-reviewed results are below-threshold demonstrations and architecture papers — not million-operation runs. The gap is no longer rhetorical; it is two numbers: ≈100 logical qubits and depth ≈10,000, neither reached in public yet.

Is early fault-tolerant a real term or a marketing one?

Both — it depends on who is using it. The technical term exists and has a published definition: Katabarwa, Gratsea, Caesura and Johnson (PRX Quantum 5, 020101 (2024)) define early fault-tolerant quantum computing as machines with tens of thousands to millions of physical qubits running fault-tolerant protocols close to the threshold — a regime with its own algorithms, and one that has not arrived either. In vendor materials, the same phrase increasingly attaches to roadmaps, or to machines whose spec sheet is still written in physical qubits.

That is not an accusation — terms migrate from papers to press releases in every field, and none of this implies bad faith by any vendor. It is a reading rule. When "early fault-tolerant" appears, ask for three numbers: logical qubits, achievable depth, and logical error rate, each with a measurement date. If the answer comes back in physical qubits, you are reading a NISQ spec sheet with a newer adjective.

So is the NISQ era over, as of September 2026?

No — but for the first time the exit is measurable instead of rhetorical. The most accurate description of the 2026 frontier is NISQ hosts with fault-tolerant islands: machines still noisy at the physical layer, on which small blocks of logical qubits have begun to operate below threshold. Whether the era ends in 2027 or 2031 depends on physical:logical ratios and error rates crossing thresholds — the numbers Rosetta Q tracks vendor by vendor in the 2026 map — not on qubit-count announcements.

One separation is worth keeping even then. Leaving NISQ is necessary for most of the algorithms with proven speedups; it is not sufficient for advantage. A megaquop machine still has to beat a strong classical baseline on a real instance, end to end — and that scoreboard, as Rosetta Q measures it, stands at 0 today for every vendor and every architecture.

What do we know, and what do we not know?

What we know. NISQ has a precise origin and meaning (Preskill, 2018). Error correction below threshold is demonstrated on real hardware (Nature 638, Dec 2024). A commercial machine advertises 48 logical qubits (Nov 2025). And the post-NISQ bar has been quantified by the person who named the era: ≈100 logical qubits, depth ≈10,000, ≈10⁻⁶ per logical operation (Feb 2025).

What we don't know. Whether vendor-reported logical-qubit counts would hold under independent, uniform benchmarks — no neutral ledger of logical-qubit performance exists (that gap is why Rosetta Q exists). Whether 48 logical qubits in one error-correcting code are comparable to 48 in another — codes differ in distance, gate sets and suppression factors, and there is no conversion standard. When the megaquop bar will be met — Preskill gives no date, and the 200-logical-qubit entries on vendor roadmaps are targets, not measurements. And whether the first post-NISQ machines will produce any end-to-end advantage on a commercial problem — that gets measured, not assumed, and today the measured count is zero.

Sources

• Preskill, "Quantum Computing in the NISQ era and beyond", Quantum 2, 79 (2018)
• Preskill, "Beyond NISQ: The Megaquop Machine", arXiv:2502.17368 (Feb 2025)
• Google Quantum AI, "Quantum error correction below the surface code threshold", Nature 638, 920 (Dec 2024)
• Quantinuum, Helios commercial launch press release (Nov 5, 2025)
• Harvard/QuEra et al., "A fault-tolerant neutral-atom architecture for universal quantum computation", Nature (Nov 2025)
• Katabarwa, Gratsea, Caesura & Johnson, "Early Fault-Tolerant Quantum Computing", PRX Quantum 5, 020101 (2024)
• IBM Quantum roadmap (Starling, 2029)
• Rosetta Q, "What is a logical qubit and why does the physical-to-logical ratio decide the timing?"
• Rosetta Q, "What is error mitigation and why does it make runs more expensive?"
• Rosetta Q, "Who is ahead in quantum computing in 2026?"

Rosetta Q publishes verdicts with reproducible raw data. This is educational content, not a product claim. The "48 / ≈100" framing above juxtaposes a vendor-reported logical-qubit count with Preskill's published megaquop sizing; it is our own comparison of two published numbers, not a measurement of progress fraction.