Microsoft, QOLAB Propose Higher Bar for Scalable Logical Qubits

When two quantum companies state how many logical qubits they possess, the numbers often refer to different things. One might count qubits that suppress specific errors under controlled conditions, whereas another might count those that endure repeated correction during a running program. This discrepancy in terminology is at the heart of a proposal from Microsoft Quantum and QOLAB, published on arXiv on September 22, 2026, which advocates for a more stringent definition of “scalable” when applied to a logical qubit.

What Constitutes a Logical Qubit, and Why Do Counts Vary?

A physical qubit is a single, fragile unit of quantum information stored in hardware, susceptible to noise, imperfect control, and measurement errors. A logical qubit distributes this information across many physical qubits to detect and correct errors. The challenge in reporting arises because “logical qubit” currently encompasses various achievements.

Some demonstrations showcase error protection, where encoding across multiple physical qubits improves memory or outperforms a single physical qubit under specific conditions. Others demonstrate error correction, where the system repeatedly identifies and rectifies faults during a computation. Both types of achievements are reported under the same label, making vendor comparisons difficult.

Microsoft has published its own milestones in this area. In April 2024, Microsoft and Quantinuum reported logical qubits with an error rate 800 times better than the underlying physical qubits. Later that year, the company described what it called the best-performing logical qubits on record, linked to priority access on Azure Quantum. These are Microsoft’s figures, promoting its hardware partnerships. The Quantinuum collaboration announcement presents the same results from the partner’s perspective.

The new proposal partly responds to such results, including those from Microsoft itself. It urges the field to move beyond treating a single strong metric as definitive proof that a machine can execute a long program.

What Microsoft Quantum and QOLAB Propose

The paper was authored by Microsoft Quantum researchers Matthias Troyer and Chetan Nayak, along with QOLAB’s John Martinis. QOLAB is a US-based company, with 11 to 50 employees, that develops high-quality qubits using semiconductor manufacturing methods. Martinis previously led Google’s quantum hardware efforts.

Their proposed term is “scalable logical qubit.” By their definition, such a qubit must endure during computation through repeated quantum error correction, not merely one-time detection or retrospective cleanup. It must support a universal set of fault-tolerant operations, including those that allow a program to react to measurement outcomes mid-execution. It must belong to a code and hardware architecture where adding physical qubits predictably reduces the logical error rate. Furthermore, it must have a credible path to replication at the scale demanded by applications.

The authors categorize these requirements into reliability, scale, capability, and performance. Reliability asks whether logical error rates surpass physical ones and continuously improve with increasing resources, with the paper targeting one failure in a trillion to one in a quadrillion operations for utility-scale tasks. Scale questions whether hundreds or thousands of logical qubits can operate simultaneously without correlated errors or crippling slowdowns. Capability spans a progression from preparing and measuring logical states to non-Clifford operations and measurement-conditioned control flow. Performance encompasses speed and cost.

The paper specifies concrete minimum numbers for usefulness: over 100 high-quality logical qubits and error rates near one in 10 billion operations. Cryptanalysis and large chemistry problems might necessitate more than 1,000 logical qubits and one failure in a quadrillion operations or better.

Why the Proposed Bar Extends Below the Qubit Layer

One of the four criteria relies on hardware that underlies the qubits. Measurement-conditioned control flow requires a classical decoder that processes error-syndrome data in real time and provides corrections quickly enough to avoid stalling the computation. This is a control-systems challenge that vendors are currently addressing.

Fujitsu’s software release illustrates the divide between older and newer assumptions within a single stack. In September 2026, the company open-sourced a quantum application package that includes noisy-era algorithms alongside fault-tolerant primitives like quantum phase estimation. A codebase containing both sets of tools indicates that the field has not yet converged on a single operating regime, which is precisely the ambiguity the Microsoft Quantum and QOLAB definition aims to resolve.

Therefore, the proposal’s demand for low-latency feedback is not abstract. The infrastructure to meet this demand is being validated concurrently with qubit development, and a definition that overlooks it would miss a significant part of the engineering difficulty.

How Reported Systems Measure Against the Proposed Bar

Testing any current claim against these four criteria quickly reveals the existing gaps. A demonstration might achieve an excellent logical error rate on a few qubits, satisfying reliability in a small system but offering no insight into scalability. Another might handle a full set of Clifford operations but lack the fast feedback loop required for capability. Post-selection improves retained results by discarding runs where errors appear, but it becomes inefficient as computations lengthen, since the probability of a clean run decreases rapidly. The paper considers post-selection a tool that can complement error correction, not replace it.

Two points maintain the integrity of the comparison. First, much of what companies advertise is measured on small hardware and extrapolated to larger systems; these should never be conflated. Second, the Microsoft Quantum and QOLAB framework is a proposal, not an adopted standard. It presents definitions, questions, and an evaluation method. It does not report a new experiment, declare a winning error-correction code, or establish a universally agreed-upon benchmark.

What Adoption Would Change for Buyers and Benchmarking

If this definition gains traction, buyers comparing systems would have a checklist rather than relying on a headline number. Does the qubit endure repeated correction during computation? Can the platform demonstrate error rates that decrease as resources increase? Is there a genuine decoder providing real-time feedback, or only offline processing? Is there a credible path from a small demonstration to hundreds or thousands of these qubits, encompassing control electronics, calibration, and interconnects?

The authors also factor in cost. They cite DARPA‘s perspective of utility-scale computing as the point where a computation’s value exceeds its cost, reframing the question as reliable computation per unit of hardware and time. Different applications set different standards, so a chemistry machine and a cryptanalysis machine would not be judged identically.

For the label to carry weight, standards bodies, funders, or a coalition of vendors would need to adopt it. However, hardware differences make a uniform standard challenging. Some platforms may exhibit fast feedback before large-scale replication; others, the reverse. This framework does not address, nor should it be interpreted as addressing, post-quantum cryptography. PQC protects classical encryption against future quantum attacks, a distinct objective from verifying qubit scalability. These two questions should remain separate.

The takeaway for anyone reading a logical qubit announcement in the coming year is to inquire which of the four criteria the number actually addresses. Until the field agrees on a shared definition, that question is more informative than the count itself.