The Late-2026 Quantum Audit: Is It a Computer Yet, and What Does a Megaquop Cost?
Year: 2026
Type: article
Abstract: This paper is the six-month follow-up to *The 2026 Quantum Audit* (April 2026), applying the Quantum Operations Model (QOM) cost framework to what changed in quantum computing between April and September 2026. The two events April named as the next audit's triggers — IBM's Kookaburra processor and the Magne neutral-atom machine — have not fired. Five other things did. Quantinuum ran a heterogeneous-code architecture (Helix) on hardware: a memory code, a compute code, and an interface between them on one machine, decoded offline. The industry adopted reliable logical operations as its unit — proposed as rQOPS and QLOPS in 2023–2025, named by Preskill's megaquop, specified by the US Department of Energy's "hard operations," and, in Sandia's QUOPS benchmark, measured for the first time on logical hardware. NVIDIA shipped a compiler that emits fault-tolerant resource estimates. DOE published a procurement specification for a quantum computer — Genesis Q, with up to $215 million in appropriation-dependent milestone pools for machines of at least 100 logical qubits performing an illustrative 10⁵ "hard" operations, evaluated in September 2028 — that is the same width and, over one run, the same total operation volume as the QOM reference computation, at a tighter error target. And Microsoft's topological program produced a second contested preprint. Two new instruments are introduced. A Von Neumann scorecard asks, for each of seven components a computer must have — memory, logic, control unit, I/O, bus, compiler, clock — whether any platform has demonstrated it below break-even, with dates. Every component has at least one public demonstration somewhere; on this report's criteria five of the seven rows have a Pass on some platform and two — the control unit, and sustained operation at the logical level — have none; no platform has more than three Passes. And a price per megaquop, from the identity C_megaquop = 10⁶ · w_sys / (U · Ω_L), where w_sys is the machine's rate in dollars per second, Ω_L its throughput in reliable logical operations per second, and U the fraction of machine time that produces the buyer's operations. On the only logical rate measured on commercially accessible hardware that we identified (Helios-1, 4.9 logical QUOPS per second, at a width of four and a completed circuit size of 40), a megaquop is a derived 57 hours — a lower bound, since the benchmark's accuracy threshold sits far above megaquop-grade reliability — and, against separately assembled cross-platform system-rate proxies, an estimated $141,000–408,000 at list prices for dedicated machine time or $61,000–184,000 at an operator's cost, with a second route from per-gate list prices giving $0.6–1.5 million; the modeled fault-tolerant machines of April's tables price it at three cents to about $550 and seconds to minutes. The distance between the two is two and a half to five orders of magnitude in price and three to four in time; the time gap is entirely the throughput term, and the system-rate term narrows the price gap rather than widening it. The April thesis holds: the QEC overhead is the dominant modeled cost, this audit identifies no independently replicated advantage over the best classical method on a workload with a classical solution in production use, and quantum computing's near-term value lies in capability rather than cost reduction. The verdict in one line: the parts exist, no platform yet satisfies the eight-row criterion for a computer, and the invoice can now be estimated per operation from a measured rate. Three errors in the April paper are corrected.
Author: Branham, Kase
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7214761876