Latest / Quickly Quantum / 61 Qubits, 110 Years, and the Benchmarking Catch — October 29, 2026
Transcript
- BlueQubit, EPFL and the XPRIZE Foundation report random-circuit sampling on IBM’s 120-qubit Nighthawk r2 processor using 61 active qubits, with a tensor-network estimate putting classical reproduction at roughly 110 years of exascale compute. That’s a striking gap—but the estimate depends on assumptions about classical sampling, and the experiment is a benchmark rather than a useful application. IBM’s commercial hardware ran the circuits through standard cloud queues, which puts accessibility in the story too. How much of the claim survives better classical methods and independent scrutiny? We’ll work through the result and its limits. Welcome back to Quickly Quantum, your daily brief on the quantum frontier. It's Thursday, October 29, 2026. Let's get into it. Intel has signed a definitive agreement to acquire quantum software firm Classiq, with the proposed combination joining Classiq’s synthesis platform to Intel’s spin-qubit manufacturing and existing software, The Quantum Insider reports. The companies’ stated plan is a full-stack quantum computing platform, while Classiq would remain hardware-agnostic. In plain terms, the software is meant to help design and optimize quantum algorithms, while Intel brings its chip-making work and software. The deal would also give Intel a path to add quantum processing units to its portfolio alongside CPUs, GPUs and application-specific integrated circuits. The consequential part is the pairing, not a completed integration: the proposed platform still has to become a working product. I think Intel’s interest makes strategic sense because software synthesis could connect its hardware effort to what developers actually build, but the agreement alone doesn’t show that the pieces already work together. For you, the signal is a major chip company choosing to add a quantum-software business to its portfolio; the proof will be in what the combined platform can deliver. From Intel’s software-and-hardware pairing, another company is pitching technology that can serve quantum work and a broader computing market: PsiQuantum unveiled an 8×8 optical circuit switching demonstration built on its low-loss silicon photonics platform, The Quantum Insider reports. Optical circuit switching directs light between connections, and PsiQuantum says its laboratory testing showed 1.2 dB of insertion loss, 40 dB cross talk and less than 3 ms of reconfiguration time, without amplification. Those are company-reported demonstration results, not an independently verified benchmark. PsiQuantum says it is sharing demonstration samples with selected prospective customers and advancing the platform toward commercial products for AI data-center connectivity, while continuing work on fault-tolerant quantum computers. The company posted on X, “AI is pushing computing infrastructure to its limits. Our photonics can help.” That’s the company’s pitch; the announcement describes a demo and prospective customer samples, not a deployed data-center product. I’m glad to see quantum-developed photonics being tested against a practical networking problem, but the customer stage is still prospective. On the enterprise adoption front, D-Wave and South Korean IT systems integrator LG CNS signed a non-binding memorandum of understanding to evaluate quantum annealing workflows for enterprise optimization, Quantum Computing Report reports. Annealing is a quantum approach designed for optimization problems, and the companies aim to connect D-Wave’s Leap cloud environment with LG CNS systems-integration work. The stated target areas include industrial scheduling, anomaly detection, telecommunications routing and logistics management across South Korea and regional Asian markets. LG CNS will lead problem formulation and algorithm validation. D-Wave posted on X that the collaboration was announced at Qubits Asia in Seoul and is intended to advance quantum-computing adoption for enterprise optimization in South Korea and international markets. The MOU is non-binding, with no hardware procurement, software licensing fees or cloud-compute commitments disclosed. That leaves the test where it belongs: whether these workflows can show measurable advantages over classical mathematical programming tools in production environments. A different route from enterprise optimization to infrastructure: municipal utility EPB and IonQ published findings across three technical preprints on hybrid quantum-classical algorithms for distribution-network reconfiguration, Quantum Computing Report reports. The work was funded under a $4 million grant from the National Institute of Standards and Technology, and tested three computational frameworks on distribution networks spanning up to 417 electrical buses. A bus here is a node in the electrical network. One framework, run on IonQ’s Forte Enterprise, found an exact optimum for the 33-bus system and came within 0.30% of the optimum for 69 buses. Another reduced median classical solver runtime on a 417-bus network from 1,415.47 s down to 34.71 s, using a fixed 29-qubit hardware budget. The paper also says end-to-end runtime on small-to-medium networks remained dominated by quantum-circuit overhead and classical processing. That distinction matters: the work reports test-system findings, not live-grid dispatch. I think the most useful result is the specific boundary they’re mapping—where quantum-generated guidance starts to offset the overhead of the full workflow. And on government-backed evaluation, a separate company is receiving support to keep testing its error-correction approach: Nord Quantique announced a US$10 million extension of its DARPA Quantum Benchmarking Initiative Stage B agreement, The Quantum Insider reports. Bosonic quantum error correction aims to protect quantum information within physical qubits while reducing the hardware overhead needed for fault-tolerant computing. DARPA’s initiative evaluates whether emerging architectures can reach utility-scale operation by 2033. The new development is an extension of Nord’s own Stage B agreement. The company says the funding supports continued evaluation and its development roadmap. Our earlier view of DARPA’s effort was that it asks for evidence capable of strengthening or weakening a roadmap; this extension keeps Nord in that evaluation, while the outcome depends on what its engineering milestones and prototypes demonstrate. Funding buys more evaluation, not a finished machine. That's the rest of today's quantum news. Now, the story I want to slow down for. This is a benchmark result that puts a commercial quantum processor against a classical supercomputer on a deliberately difficult sampling task. Quantum Computing Report says researchers from BlueQubit, EPFL and the XPRIZE Foundation ran random-circuit sampling on IBM Quantum’s 120-qubit ibm_phoenix processor, using its square-lattice Nighthawk r2 architecture. Our show has covered quantum advantage claims before; the useful question each time is what the comparison actually measures and how sensitive it is to the classical estimate. The team used 61 active qubits and submitted circuits through standard Qiskit Runtime cloud queues, without bespoke hardware calibration. Random-circuit sampling asks a machine to produce samples from the output of a randomly constructed quantum circuit. It’s a test of whether classical methods can reproduce that output efficiently, not a calculation with a direct commercial use by itself. At the anchor depth of 36 cycles, the circuits incorporated 918 controlled-Z gates, a type of operation between qubits. Fidelity here indicates how closely the observed output matches the ideal circuit’s expected output. IBM’s processor collected 106 samples in 19 seconds of active quantum processing unit time. Under baseline rejection-sampling assumptions, the team estimated that reproducing the ensemble on Oak Ridge National Laboratory’s Frontier supercomputer would require an estimated 1.2 × 10^27 floating-point operations, or roughly 110 years of exascale compute. The study’s case rests on more than a single extrapolated number. The researchers checked circuit fidelity across depths from 4 to 40 cycles using mirror benchmarking with randomized compiling, and cross-entropy benchmarking across three- and four-partition balanced patches. Those are two distinct ways of testing whether the hardware’s output tracks the intended circuit. The agreement gives the claimed sampling result a stronger technical foundation than an estimate standing alone. The hardware comparison also has a concrete reference point. The report gives per-cycle fidelity of 0.872 for Nighthawk r2, compared with 0.836 for the first-generation Nighthawk r1, and says the difference compounds to more than an order-of-magnitude fidelity gain at deep cycle depths. The result, if it holds up, is that a processor available through a standard cloud workflow reached a benchmark regime where the paper’s baseline classical method becomes extremely costly. But the classical estimate is the hinge, and the authors themselves describe it as a baseline. Their estimate assumes exact single-amplitude contractions with unbounded memory; other approaches, including batched tensor slicing, matrix-product-state approximations and GPU-accelerated reuse across multiple amplitudes, could lower classical runtimes. The article points to historical benchmarks involving Google Quantum AI’s Sycamore and USTC’s Zuchongzhi, where improved tensor-network methods changed earlier classical estimates. There is also a direct check against a simple way of making the benchmark look easier: severing attacks, which remove couplers between circuit regions to approximate smaller pieces that classical computers can handle. At 36 cycles, every tested severing mode scored below 2.2 × 10−4, at least an order of magnitude below the experimental target score. That supports the claim that this circuit sits on the classically difficult side of the noise-induced phase transition—the point where device noise changes which methods can efficiently simulate the result. I think the strongest part is the combination of a cloud-run experiment, two fidelity checks and explicit tests against severing attacks. The open point is the one that matters most for an advantage claim: whether independent researchers, using stronger classical algorithms, can reproduce or narrow the comparison. The team has published technical details and validation data on arXiv, and put open-source circuits, execution scripts and the full 106-sample dataset on GitHub. That makes the result testable, which is a real strength; it does not make the century estimate a settled fact. And the practical stakes need the right scale. Random-circuit sampling is a benchmark, not evidence that this processor is solving a useful customer problem. I read this as meaningful progress in benchmarking commercial quantum hardware, rather than a demonstration of practical computational advantage. If better classical methods shrink the gap, the headline comparison changes; if independent work reproduces the result under stronger classical baselines, the claim gets much harder to dismiss. Time for the Hype Check. The unusually accessible hardware run and the cross-checks make this a serious benchmark, while the classical runtime still rests on assumptions that deserve independent testing, so I give it a 7 out of 10 for substance. The stake is whether 61 qubits on a commercial processor can stay beyond classical reach; the next evidence is independent reproduction of the released circuits and dataset against improved classical methods. If you value clear-eyed coverage of quantum benchmarks and the companies building around them, follow today's episode wherever you listen. This has been Quickly Quantum, an AI-voiced podcast, created and built by a real human using today's cutting-edge technology. Nothing you heard on this show is financial advice. I'm Brian Lampert, and I'll catch you all tomorrow — take care! Watching on YouTube? Hit subscribe, and tomorrow's briefing will be waiting for you. I also host Concrete Compute: a daily briefing on the AI buildout. The datacenters, the megawatts, and who actually pays for them. Find it wherever you get your podcasts.