Every Quantum Vendor Measures Progress Differently. That’s the Point.

Introduction

Go to IBM’s hardware page today and you’ll find Nighthawk r2 listed at 120 qubits. Open the April 2025 Technology Atlas and the same processor family is scaling to 360 qubits by 2026 and 1,080 by 2027. Now read the July 2026 metrics blog and those plans become 120 “programmable qubits” in 2026, 500+ in 2027, and 1,000+ in 2028. Three IBM pages, three sets of numbers, the same processor family. Which one is right?

All of them. That’s the problem.

The hardware page describes the current chip. The April 2025 roadmap described multi-module scaling using a total qubit count that included every element on every connected chip. The July 2026 blog introduced a new framework that distinguishes “programmable qubits” (the ones users actually control) from supporting hardware like tunable couplers and reset gadgets. IBM’s November 2025 announcement already described Nighthawk as “120 qubits linked by 218 tunable couplers,” so the distinction existed before the label did. The “programmable qubits” framework is a clearer way to describe what was already being counted.

Separately, the scale-up timeline changed. The April 2025 roadmap had 1,080 qubits across nine modules arriving in 2027. By mid-2025, IBM had already revised that target, pushing 1,000+ connected qubits to 2028. The July 2026 blog reflects the updated schedule, not the original.

Two things happened in parallel: IBM clarified its counting terminology and revised its roadmap downward. Each is defensible on its own. But a reader who checks three different IBM pages on the same day and gets three different numbers for the same hardware is not experiencing a clarity upgrade. They’re experiencing a mess. And IBM, to its credit, is at least trying to clean it up by defining its terms. Most of the industry has not even done that.

Major quantum hardware vendors increasingly report different and non-equivalent measures of progress. Some are proprietary composite metrics. Others are shared benchmarks, standard engineering quantities, or experimental milestones. The strategic effect is the same: each company foregrounds the dimension on which its architecture appears strongest. The result is an industry where the people who most need to compare platforms (enterprise buyers, investors, defense planners modeling CRQC timelines) lack the normalized data required to do so.

Six Vendors, Six Scoreboards

I’ve written a comprehensive technical reference on quantum computing benchmarks covering RCS, Quantum Volume, Algorithmic Qubits, CLOPS, rQOPS, and other methodologies. This article is not a repeat of that material. What I want to examine here is the competitive logic behind each vendor’s metric choices, because the pattern is more revealing than any individual number.

IBM now emphasizes programmable qubits, QuOps (two-qubit operations as a quality indicator), and maximum circuits per second (throughput). This three-axis framework rewards superconducting systems’ speed advantage. IBM expects Nighthawk r2 to deliver roughly 25x the circuit throughput of its predecessor Heron fleet, with a target of 100,000 circuits per second. The framework also lets IBM report logical qubits and T-gate counts alongside programmable qubits. The result is a multi-tier picture that looks more sophisticated than a single number.

Google reports 105 physical qubits on its Willow processor, but the metric Google emphasizes is below-threshold operation: demonstrating that adding more physical qubits to a logical qubit reduces rather than increases the error rate. Google’s Quantum Echoes experiment measured verifiable out-of-time-order correlator (OTOC) expectation values on Willow. Google estimated that the hardest instances, which took approximately two hours on Willow, would require roughly 13,000 times longer on a classical supercomputer after extensive classical red-teaming. Google’s preferred framing rewards a research-lab posture where a 105-qubit chip with below-threshold error correction and a verifiable quantum advantage result sounds more advanced than a 1,000-qubit system that hasn’t demonstrated error suppression. For Google’s purposes, it is.

Quantinuum reports the highest published Quantum Volume in the industry: $$2^{25}$$ (approximately 33.5 million), achieved on its 56-qubit System Model H2 in September 2025. Quantinuum says this is over 16,000 times higher than the nearest competitor’s reported QV. On its newer 98-qubit Helios system, Quantinuum demonstrated 48 error-corrected logical qubits from 98 physical qubits using two-level concatenated Iceberg codes, achieving a physical-to-logical encoding ratio of approximately 2:1. (Those experiments used postselection and included both fault-tolerant and partially fault-tolerant benchmarks, so the result is not directly comparable with a roadmap claim for continuously operated universal logical qubits.) Quantinuum reports among the highest commercial-system gate fidelities: 99.9975% single-qubit and 99.921% two-qubit on Helios. This metric selection is logical: trapped-ion systems have high native fidelities, and QV is the metric that most directly rewards fidelity over raw count. Quantinuum also published a detailed technical critique of IonQ’s competing metric, calling it “a poor substitute” for QV. That tells you as much about the competitive dynamics as the technical arguments do.

IonQ made Algorithmic Qubits (#AQ) central to its roadmap, reaching #AQ 35 on Forte in January 2024 and announcing #AQ 64 on a Tempo development system in October 2025. #AQ scores the largest width-and-depth region in which a suite of structured algorithmic circuits (quantum Fourier transform, Grover’s search, QAOA, and others) meets a defined success threshold. It is not the logarithm of Quantum Volume, despite a rough correspondence for unmitigated systems. #AQ uses structured algorithm circuits rather than QV’s random square circuits, and it permits disclosed full-stack optimization and error mitigation. Quantinuum argues those rules make the metric susceptible to compiler and mitigation effects that inflate scores. IonQ argues that measuring the delivered system rather than bare hardware is precisely the point, because users care about the final answer.

More revealingly, IonQ introduced a 13-workload application benchmark framework in April 2026 that reports solution quality, time-to-solution, and energy-to-solution alongside #AQ. Even the company most closely associated with a single-number metric now appears to accept that one number is insufficient.

QuEra and the neutral-atom community foreground logical-qubit count, logical error rate, and reliable-operation budgets. QuEra’s current roadmap targets its fault-tolerant Libra system by 2028: 256 logical qubits from more than 10,000 physical qubits at a target logical error rate of $$10^{-6}$$, followed by more than 1,000 logical qubits from over 20,000 physical qubits in 2028/29. In January 2026, a QuEra-led team demonstrated 96 logical qubits from 448 physical atoms using high-rate [[16,6,4]] codes, the largest verified logical-qubit result to date. These claims depend on high-rate codes, real-time decoding, atom-loss detection, and active reloading. Logical count alone is no more sufficient for evaluating QuEra than physical count is for evaluating IBM.

Microsoft introduced rQOPS (reliable quantum operations per second) in 2023, defining it as the product of logical-qubit count and logical clock speed, qualified by a specified logical error rate. Microsoft proposes one million rQOPS at a logical error rate of at most $$10^{-12}$$ as the entry point for a useful quantum supercomputer. The metric is forward-looking by design: it describes fault-tolerant machines that don’t exist yet. This positions Microsoft as thinking at a systems level while competitors are still counting physical qubits. Microsoft has yet to register its own hardware on its own scale.

The Pattern

The industry is not merely counting qubits differently. It is competing through different definitions of meaningful progress.

IBM foregrounds programmable scale, circuit complexity, and throughput. Google foregrounds below-threshold error correction and verifiable beyond-classical experiments. Quantinuum emphasizes physical fidelity, Quantum Volume, and high-rate logical encoding. IonQ has moved from #AQ toward application-level solution quality, time, and energy. QuEra reports logical-qubit roadmaps alongside logical error targets and reliable-operation budgets. Microsoft describes future fault-tolerant systems through rQOPS.

Some of these measures are proprietary. Some are open benchmarks. Some are ordinary technical quantities, and some are experimental milestones. No single one is fraudulent, but each selects a different slice of system performance. No one lies. Everyone curates.

A buyer presented with “1,000 programmable qubits,” “#AQ 64,” “QV $$2^{25}$$,” “48 encoded logical qubits,” and “one million rQOPS” is not looking at five scores on the same scale. It is like comparing a car’s horsepower, fuel efficiency, cargo volume, and safety rating and trying to decide which car is “ahead.” The numbers answer different questions, under different assumptions, for machines at different stages of development.

Three Audiences Getting Hurt

This metric fragmentation does real damage to three groups that cannot afford confusion.

Enterprise buyers evaluating quantum hardware for pilot programs face vendor presentations where every competitor appears to lead. IBM foregrounds throughput and scale-out plans. Quantinuum foregrounds quality and encoding efficiency. IonQ foregrounds algorithmic benchmarks. QuEra foregrounds logical-qubit count and error rates. Even technically sophisticated buyers rarely receive the normalized, independently audited, and workload-specific data required to reproduce cross-platform comparisons. The buyer who defaults to the simplest-sounding number (physical qubit count) is using the least informative metric available.

Investors and analysts encounter the same problem in quarterly earnings calls and funding rounds. When IonQ reports that each increment in #AQ doubles the computational state space, and IBM reports that Nighthawk’s throughput is 25x higher than its predecessor, and Quantinuum reports QV $$2^{25}$$, each company appears to be making extraordinary progress. They may well be. But no investor can assess relative progress when every company defines its own scoreboard. Public-market incentives compound the problem. IonQ is already publicly traded, and Quantinuum completed its public listing in June 2026 on Nasdaq under ticker QNT. Metrics reported in investor communications now carry consequences beyond technical benchmarking.

Defense planners and policymakers assessing when quantum computers become a cryptographic threat face the most consequential version of this problem. If one vendor announces “1,000 qubits by 2027” and another announces “200 logical qubits by 2029,” the numbers seem to describe comparable progress. They do not. A thousand noisy physical qubits cannot run a cryptographically relevant instance of Shor’s algorithm. But 200 logical qubits do not constitute a CRQC either. A leading 2025 surface-code estimate for factoring RSA-2048 requires roughly 1,537 logical qubits and almost 900,000 physical qubits under demanding hardware assumptions. Logical-qubit count is only one part of the threat calculation. Logical error rate, logical clock speed, decoder latency, magic-state throughput, non-Clifford gate budget, and sustained operation duration are equally decisive. Treating headline metrics as interchangeable can materially distort threat timelines.

Why Single Numbers Will Always Fail Here

The deeper issue is that quantum computers are not one-dimensional objects, and every single-number metric (QV, #AQ, qubit count, rQOPS) collapses a multi-dimensional reality into a single axis.

I built the CRQC Quantum Capability Framework to address exactly this problem. The framework tracks ten distinct capability dimensions — from quantum error correction and below-threshold operation to decoder performance, magic state production, continuous operation, and engineering scale. All ten must mature in concert before a quantum computer can break RSA-2048. Progress in one dimension (say, qubit count) means nothing if another dimension (say, decoder throughput at scale) remains unsolved. I’ve shown in detail why these constraints compound nonlinearly as you scale.

No single number captures this picture. A machine with 1,000 physical qubits and two-qubit gate fidelity of 99.99% is a categorically different instrument from a machine with 1,000 physical qubits and fidelity of 99.5%. A system that can sustain an error-corrected logical computation for 10 hours is a different instrument from one whose calibration, loss handling, or logical error budget limits it to 10 minutes, even when their nominal qubit counts and gate fidelities look similar. For a fixed code family, noise model, and target logical error rate, better physical operations can reduce the required code distance and physical-qubit overhead. Collapsing these interdependencies into a single metric destroys exactly the information that matters most.

The vendor metric fragmentation is, ironically, evidence that the industry has outgrown single-number benchmarks. Qubit count served as a rough proxy for progress when every quantum computer was too small and too noisy to do anything useful. Now that companies are actually competing on different capability axes, the pretense that a single number captures “how good” a quantum computer is has become unsustainable.

The problem is what’s replacing it. Instead of converging on a shared multi-dimensional assessment standard, the industry has produced a collection of scoreboards, each optimized for the vendor that designed it. The standards vacuum is not total. QED-C maintains an open application-oriented benchmark suite. DARPA’s Quantum Benchmarking Initiative is evaluating whether vendor architectures can credibly reach utility scale. CEN-CENELEC JTC 22 has a dedicated working group for quantum computing and simulation standards. What is missing is not activity but adoption: no independently audited, procurement-grade, cross-modal framework has become the industry’s common scoreboard.

For the community that needs to track progress toward cryptographically relevant machines, this is a step backward wrapped in the appearance of sophistication. The vendors know this. Whether their customers do is a different question.