THE QUANTUM INTELLIGENCER

At the closing keynote of IBM's 2026 Think conference in Boston, Jay Gambetta, Director of Research and IBM Fellow, declined to announce what was coming next and instead described what the company already has. The ledger, as presented, includes a simulation of a protein-ligand complex of 12,635 atoms, conducted by researchers at the Cleveland Clinic and RIKEN on IBM quantum hardware in concert with the Fugaku supercomputer, and reported as the largest quantum-centric simulation of its kind to date, with a 210-fold improvement in accuracy over previous quantum-centric methods. It includes an Oak Ridge National Laboratory calculation of the free energy of FLiBe, the molten salt proposed for fusion reactor blankets, which agreed with leading classical methods without relying on crude approximations. It includes a Q-CTRL materials simulation of up to 60 interacting electrons, finished in two minutes on the IBM Quantum Platform, where a leading classical method required one hundred hours, with accuracy within one percent. A separate benchmark study, applied to leading processors from Quantinuum, Google, and IBM, quantifies the distance that remains. Computational capability, by that benchmark, must grow by five orders of magnitude before recognized challenge problems can be executed at useful scale. The announcements describe a capability; the benchmark describes the interval to the next one. Both are true, and the interval is where the work now sits. Certain terms recur in the reports from Boston, and they repay translation. A quantum computer works on qubits, the quantum analogue of the bit. A bit is either zero or one; a qubit may hold a weighted combination of both until it is measured. The calculation proceeds in circuits, sequences of operations on those qubits. The benchmark study is called QUOPS, for quantum universal operation performance system, and it measures two things: the size of the largest computationally relevant quantum circuit a machine can execute successfully, and the speed at which it can execute them. The same paper distinguishes physical qubits from logical ones. A physical qubit is hardware. A logical qubit is an engineered construction: several physical qubits made to behave as one dependable unit, so that an error in a single member can be detected and corrected. A fault-tolerant processor is one that carries out that correction automatically. The processor tested in the study used up to eight logical qubits in a [7,1,3] encoding. Quantum-centric supercomputing, known by the initials QCSC, is the name for a division of labour between machines. A quantum processor and a classical supercomputer exchange intermediate results, each performing the step it does best; the protein simulation just described ran in this mode. Sample-based quantum diagonalization, the algorithm used in that simulation, is a hybrid procedure of the same family: the quantum machine produces a sample of probable electron configurations, and the classical machine performs the diagonalization, the operation that resolves those configurations into allowed energies. Two chemical terms carry weight. The free energy of a system is the portion of its energy available to perform work at constant temperature. FLiBe, the molten salt named for its constituents, fluorine, lithium and beryllium, is the mixture proposed for the lining of a fusion reactor wall, where it would capture neutrons and turn lithium into tritium. Interacting electrons, finally, are electrons whose behaviour must be computed together, since each responds to the others. The number of joint possibilities grows faster than a classical machine can track, and the point at which exact classical calculation stops is the ceiling the reports refer to. An analogy may fix the arrangement in the mind. Quantum-centric supercomputing is not a single machine but a division of labour, and the division is an old one. In a counting house, the one clerk who can take a difficult extraction performs that single operation, hands the figure to the ledger room, and receives back the amended columns; neither the specialist nor the room could close the month's books alone. So it is with the protein simulation announced from Boston. The quantum processor produces a sample of probable electron configurations, and the classical machine, Fugaku, takes up the surrounding arithmetic and returns a refined problem; the reports describe the work as passing results back and forth. The specialist and the ledger room are both necessary, and the gain is not that one replaces the other. The benchmark's five orders of magnitude may be mapped with equal plainness. A vessel that has crossed a harbour has not thereby crossed an ocean. The remaining distance is a hundred thousandfold, and the craft that will manage it is not the harbour boat made faster, but a vessel of another kind. The distinction is not a rebuke to the harbour boat. It is a description of the interval, and the interval is long. One concrete instance of the thing working is not in the Boston ledger. A quantum-language paper reports an experiment in which the semantic space of ordinary sentences was taken to a real quantum computer. An embedding is the list of numbers a language model assigns to a sentence so that meanings can be compared arithmetically; the standard measure of that comparison is the cosine of the angle between two embeddings. The authors mapped embeddings from the Google Sentence Transformer into quantum circuits and calculated that cosine similarity on hardware. The calculation is a demonstration, not a service. The point is that it ran. A companion paper sets out six principles for a quantum-inspired model of semantic representation, with an argument that models built on those principles could become significantly more powerful and efficient. For the record, then: the principle has left the chalkboard and executed on hardware. Two further figures belong in the ledger. The fault-tolerant processor that carried the benchmark's eight logical qubits is Quantinuum's Helios-1. The Q-CTRL simulation, reported as two minutes against one hundred hours, is elsewhere stated as a speedup of more than three thousand times over a leading classical method. The same report records that global tritium production is a few pounds per year, a constraint the fusion work must answer. —Inspector Grey Dispatch from The Prepared E0