IBM's Nighthawk r2 Runs 100,000 Circuits a Second - But the 25x Is Wall-Clock Speed, Not Qubit Quality

by Karl Lindström · September 3, 2026

Semiconductors & Hardware Reporter, Frontier Tech Wire

IBM put a new superconducting quantum processor on its cloud platform on 31 August and, according to the company's own technical blog post on Nighthawk r2, the machine can execute up to 100,000 circuits per second against roughly 4,000 per second for the Heron fleet it sits alongside. IBM states the resulting 25x directly rather than leaving it to be derived, and it is that figure which is now circulating through the quantum trade press.

It is worth being precise about what that number measures, because this beat has a long history of throughput gains being retold as computational ones. The 25x is circuits executed per second. It is not a claim about qubit count, gate fidelity, coherence during normal operation, or the size of problem the machine can solve. On IBM's own telling, the improvement comes from removing dead time between shots, not from making any individual shot better.

What the chip actually contains

Nighthawk r2 carries 120 programmable qubits on what IBM describes as a square-lattice architecture in which most qubits are connected to four nearest neighbours, supported by 218 dedicated tunable couplers and 120 independent reset elements - one paired with each programmable qubit. IBM counts 458 total physical quantum elements on the die. Quantum Zeitgeist, writing up the release independently, put the same breakdown in plainer terms: of the 458 on-chip elements, 338 are dedicated to coupling and reset functions rather than to computation a user can program. That is the trade. IBM has spent a large share of its silicon budget on plumbing, and the return is measured in wall-clock time rather than in problem size.

The reset trick

The mechanism is a dissipative reset. Quantum Computing Report, summarising the same release, described the operation as activating a reset coupler that connects the qubit to a cold bath. IBM says the effective T1 - the relaxation time governing how long a qubit takes to decay back toward its ground state - drops from a median of about 200 microseconds with the reset off to roughly 25 nanoseconds with it on, and that idle time between circuit runs falls to, in IBM's phrasing, as little as a single microsecond. IBM also says the reset leaves neighbouring qubits undisturbed and is available during circuit execution as well as between circuits, a prerequisite for the repeated measure-and-reset cycles that quantum error correction depends on; Quantum Zeitgeist reported no observed degradation in T1 on neighbouring qubits during reset.

One figure does not reconcile cleanly across the write-ups, and readers should know it. IBM's blog post describes around a 25x reduction in initialization error across the device, and Quantum Computing Report repeats that figure. Quantum Zeitgeist's analysis of the same launch reports an 11-fold reduction in initialization error, attributing it to cooling the qubits with a 50-ohm termination. Neither figure is Frontier Tech Wire's own arithmetic; both appear as stated in the sources named. IBM has not published the underlying calibration data that would establish the conditions behind each number, and this report leaves the conflict open rather than picking one.

What was measured on hardware

The performance claims IBM attaches to the part are presented as results obtained on the processor rather than as simulations, though IBM has not published the underlying data and none of it has been independently replicated. IBM reports accurate observable estimation, using Probabilistic Error Amplification, on circuits containing 7,500 gates - which the company identifies as a 2026 milestone on its own quantum roadmap - and cites two application results: as much as 10x faster runtimes with no loss in accuracy on circuits it characterises as advantage candidates, and a 12x speedup on physical neutron-scattering simulations, producing material spectra for direct laboratory comparison in approximately 60 seconds. Quantum Computing Report puts two-qubit gate error on the part at approximately 3x10^-3, describing it as maintaining Heron-class rates, and that is the number to watch alongside the throughput headline, because it is the one that bounds how deep a useful circuit can go regardless of how fast the machine cycles.

Why a throughput number still matters

There is a real reason not to dismiss this as a benchmark trick. Almost every near-term quantum algorithm of practical interest - variational solvers, error-mitigation schemes, randomised compiling, tomography - works by running the same small circuit thousands or millions of times and averaging. For that class of workload, sampling rate is the binding constraint on how long a job takes and, on metered cloud hardware, on what it costs. Cutting per-shot overhead by orders of magnitude changes the economics of those algorithms even if it changes nothing about the physics of a single shot. It also changes what a research group can attempt inside a fixed allocation, which is the practical bottleneck most academic users describe. IBM has a webinar on the part scheduled for 10 September, which is where independent probing of the numbers is likely to begin.

The listed-company context

None of the three pure-plays competes with IBM in the same procurement channel, but IBM's roadmap execution is the yardstick many generalist investors use to judge whether the sector's technical claims hold up. The US market was open at the time of writing and this report cites no prices for the current session. Nothing here is investment advice.

None of the three pure-plays competes with IBM in the same procurement channel, but IBM's roadmap execution is the yardstick many generalist investors use to judge whether the sector's technical claims hold up. For broader context, in the session before publication the S&P 500 closed Wednesday at 7,666.60, up 0.46%, the Nasdaq Composite at 26,217.83, up 0.45%, and the Russell 2000 at 2,953.17, up 1.13%. The US market was open at the time of writing and this report cites no Thursday prices. Nothing here is investment advice.