The significance of the result lies in how it sidesteps the central limitation of conventional photonic graph-state synthesis. In linear optical systems, the standard tool for combining smaller resource states into bigger ones is the fusion operation, a divide-and-conquer strategy first formalized by Browne and Rudolph in 2005. Fusion works by interfering photons and measuring the outcomes, which links the constituent states into a larger graph. The catch is probability: with ordinary beam splitters and detectors, fusion succeeds only some of the time. Failed attempts destroy the participating photons, forcing experimenters to build elaborate multiplexing schemes, extensive feed-forward electronics, and massive hardware overheads simply to make the overall process behave deterministically. Resource-cost analyses, such as the 2015 study by Li and colleagues, have quantified how punishing these overheads become for fault-tolerant linear optical quantum computing.

The Tsinghua group, led by Hongyi Zhang and Luming Duan of the Center for Quantum Information at the Institute for Interdisciplinary Information Sciences, took a different route. Rather than relying on passive linear optics, they implemented the fusion operation on a superconducting circuit device, where microwave photons interact with engineered quantum circuits in a controllable, coherent way. Their scheme non-destructively projects pairs of flying microwave photons into Bell states, the maximally entangled two-qubit states that serve as the connective tissue of graph-state fusion. Because the projection is deterministic and the photons survive the operation, every fusion attempt contributes to the final state, eliminating the exponential losses that plague probabilistic protocols.

The photons themselves are encoded in time bins, a format in which the logical states of a qubit are represented by a photon arriving in one of two well-defined temporal modes. Time-bin encoding is attractive for propagating qubits because it is robust against many channel imperfections and fits naturally with sequential generation schemes. The team generated small time-bin-encoded cluster states on demand, meaning that each elementary building block could be produced precisely when the protocol required it. On-demand availability is crucial for a scalable architecture: if the photonic qubits must wait in flight while the apparatus catches up, decoherence erodes the fragile quantum correlations before the graph can be completed.

Managing decoherence is, in fact, a central theme of the experiment. Superconducting qubits, the stationary quantum circuits that mediate the photon-photon interactions, have finite coherence times, and a long graph-state synthesis sequence would normally exhaust them. The researchers therefore built active reset and reuse of the superconducting qubits into the protocol. After each qubit has performed its role in generating or fusing photonic qubits, it is rapidly reset to a known state and pressed back into service. This recycling circumvents the effects of decoherence that would otherwise accumulate over the many rounds of photon emission and fusion needed to build a large graph, and it echoes the complete methods sets developed in other platforms, such as the scalable ion trap techniques demonstrated by Home and colleagues in 2009.

Programmability is the second pillar of the demonstration. Because the fusion operation can be applied selectively to chosen pairs of photonic qubits, the same hardware can produce different graph topologies under software control. The team showed this by fusing two linear cluster states of three qubits each, joining qubit 2 and qubit 4 to create a new vertex and thereby producing a five-vertex caterpillar graph state, a structure whose entanglement pattern resembles a chain with additional rungs. Verifying such states is itself a formidable task: full quantum state tomography of a five-qubit density matrix would be prohibitively costly because of the sheer dimension of the Hilbert space, so the experimenters instead measured a subset of stabilizer operators, the commuting observables that uniquely characterize stabilizer states such as graph states. One stabilizer, involving Z and X operators across all five qubits, was deliberately omitted from the measurement set for exactly this reason.

To certify that the generated states were genuinely multipartite entangled, meaning that the entanglement cannot be reduced to any mixture of states over smaller partitions, the team computed fidelities with respect to the ideal target states and compared them against the standard threshold of 0.5. The measurements, performed with datasets of five times ten to the seventh power shots and analyzed with bootstrap standard errors derived by parametric resampling, confirmed genuine multipartite entanglement. Pushing further through repetitive fusion, the researchers scaled the approach to 13 photonic qubits, establishing what they describe as a scalable framework for photonic graph-state synthesis. The stabilizer formalism used for entanglement detection draws on well-established tools, including the entanglement-witness techniques of Toth and Guhne.

The microwave domain in which this demonstration takes place deserves particular attention. Circuit quantum electrodynamics has matured into a platform where itinerant microwave photons can be generated, routed, and measured with high fidelity, and where quantum non-demolition detection of individual propagating photons became possible in 2018 through the work of Kono and colleagues and Besse and colleagues. Deterministic quantum state transfer and remote entanglement using microwave photons were demonstrated by Kurpiers and colleagues in 2018, and universal gate sets for itinerant microwave photons followed in 2022. The Tsinghua result extends this trajectory from point-to-point operations to the synthesis of extended entangled photonic structures, a capability that connects naturally to proposals for microwave quantum networks and to hybrid architectures in which superconducting processors interface with flying qubits.

At the same time, the achievement sits within a broader international effort to make photonic graph states deterministic. Single quantum emitters, including quantum dots, trapped atoms in optical cavities, Rydberg superatoms, and solid-state emitters, have all been used to generate cluster states sequentially, with notable demonstrations by Cogan and colleagues, Thomas and colleagues, and Yang and colleagues between 2022 and 2025. Thomas, Ruscio, Morin, and Rempe showed in 2024 that deterministically generated photonic graph states from a single atom can themselves be fused, and temporal fusion of emitter-produced resource states was reported in 2025. The Tsinghua experiment adds a distinct architectural idea to this landscape: instead of growing one large state from a single emitter, it fuses independently generated small states through a deterministic Bell-state projection, combining the strengths of on-demand sources with the modularity of fusion-based quantum computation, a model formalized by Bartolucci and colleagues in 2023.

The implications reach toward several long-sought applications. Measurement-based quantum computing, proposed by Raussendorf and Briegel in 2001 as the one-way quantum computer, consumes a large entangled graph state as its computational substrate, and fusion-based schemes with high fault-tolerance thresholds have been proposed for one-dimensional cluster states. All-photonic quantum repeaters, introduced by Azuma, Tamaki, and Lo in 2013, replace matter-based quantum memories with graph states, removing the memory bottlenecks of conventional repeater chains. Graph states also serve as resources for quantum metrology, as Shettell and Markham showed in 2020. A deterministic, programmable fusion engine that can be reconfigured at will brings each of these visions closer to practical realization, particularly if the underlying photonic qubits can eventually be converted to the optical telecom band for long-distance distribution. The work was supported by China’s Quantum Science and Technology major project, the National Natural Science Foundation of China, Tsinghua University, and the Ministry of Education, and the underlying data have been made openly available, offering the community a concrete foundation on which to build the next generation of entangled photonic resources.

Subject of Research: Deterministic fusion of microwave photons on superconducting circuits for scalable photonic graph-state generation

Article Title: Deterministic and programmable fusion for the scalable generation of photonic graph states

Article References: Li, Y., Yang, Y., Bao, Z., Yang, J., Sun, L., Song, Y., Wu, Y., Zhang, H., & Duan, L. (2026). Deterministic and programmable fusion for the scalable generation of photonic graph states. Nature Physics. https://doi.org/10.1038/s41567-026-03471-5

Image Credits: AI Generated

DOI: 10.1038/s41567-026-03471-5

Keywords: photonic graph states, quantum entanglement, superconducting qubits, microwave photons, fusion operation, cluster states, measurement-based quantum computing, quantum repeaters, time-bin encoding, circuit quantum electrodynamics, stabilizer measurements, genuine multipartite entanglement

The experiment, led by Ziqian Li and Eesh Gupta, who contributed equally, together with senior author David I. Schuster, showcases a device in which seven distinct memory modes can be individually addressed through a single transmon qubit acting as a classical control intermediary. Crucially, the memory modes are not raw qubits exposed to the noise of the processing layer. Instead, they live inside a multimode storage cavity, and a carefully engineered buffer layer sits between the processor and the cavity, isolating the fragile stored quantum states from the nonlinear elements that make computation possible but also make errors likely.

The need for such a device is easy to appreciate when one considers how modern superconducting quantum processors actually operate. In today’s machines, every physical qubit is tied to its own dedicated control lines, readout resonators and filters. A logical qubit protected by a surface code can demand dozens of physical qubits, each with its own coaxial cable running down into the dilution refrigerator. The wiring burden grows so quickly that scaling to useful machines has become as much a problem of engineering and cryogenics as of quantum physics. A random access memory flips this picture: if quantum information can be parked in high-coherence storage and retrieved on demand through a small number of shared control channels, the number of lines required per logical qubit falls dramatically, and the logic and storage subsystems can each be optimised separately.

The conceptual foundations for quantum random access memory stretch back to 2008, when Giovannetti, Lloyd and Maccone formalised the idea of a structure that could be queried in superposition, an operation at the heart of several quantum algorithms with proven exponential advantages. Practical implementations, however, have lagged far behind their classical counterparts. Demonstration efforts in other platforms, including a 105-mode random access memory realised with atomic ensembles in 2019, showed that the principle could be realised, but superconducting processors, the leading candidate for fault-tolerant quantum computation, notably lacked an equivalent. Architectural proposals from computer scientists, including work by Baker, Schuster and Chong on memory-equipped quantum architectures, argued that the power of random access could reshape how error-corrected machines are organised, but the hardware to realise such a vision did not yet exist.

The new device closes that gap with an elegant layered design. At its heart is a multimode microwave storage cavity, a single physical object that supports many electromagnetic modes at distinct frequencies, each capable of holding a quantum state for a long time. Seamless cavities of this kind have previously achieved photon lifetimes measured in tens or even hundreds of milliseconds, far exceeding the coherence of planar transmon qubits. A single transmon, the workhorse qubit of the superconducting world, provides the address mechanism: by tuning the transmon into resonance with a selected cavity mode, the experimenters can swap a quantum state into or out of that mode while leaving the others untouched.

The crucial innovation is the cascaded architecture with its buffer layer. Directly coupling a transmon to a multimode cavity creates a problem: the strong nonlinearity of the qubit, which enables control, also pushes unwanted photons and residual excitations back into the memory, degrading the stored states through induced dephasing and other many-body effects. The team solved this by inserting an intermediate buffer mode between the processor and the storage cavity, cascading the interactions so that the storage modes are only weakly, virtually engaged during operation. The single transmon then classically addresses the seven memory modes through this buffer, addressing them one at a time while the full stack remains protected from processor nonlinearities.

The performance figures reported are striking for a first demonstration. Arbitrary random access across the eight-bit device, meaning the ability to write to or read from any chosen memory mode on demand, was achieved with an average infidelity of less than 1.5 percent per mode. The researchers did not stop at headline numbers: they characterised the dominant error processes in detail, showing that residual many-body interactions within the multimode cavity set the present error budget. This kind of honest error accounting matters enormously for the path forward, because it tells engineers exactly which physical mechanisms must be suppressed as the architecture is scaled to more modes and longer storage times.

Beyond mere storage, the architecture supports operations performed transversally within the memory module, meaning operations that act across the encoded information without spreading errors from one component to its neighbours, a key requirement for fault tolerance. Combined with the drastic reduction in control lines per logical qubit, this positions the cascaded random access memory as a candidate unit cell for fault-tolerant quantum architectures. In such a scheme, the memory module would hold many logical qubits in long-lived cavity modes, a modest number of transmons would provide address and control, and dedicated processors would operate on retrieved states, each subsystem optimised on its own terms. The authors describe the result as enabling resource-efficient control of logical qubits, separating the problems of high-speed logic and long-duration storage that currently fight for the same physical qubits.

The work also resonates with a broader trend in the field towards bosonic and multimode approaches to quantum information. Error-corrected logical qubits encoded in oscillator states, dual-rail cavity qubits with erasure detection, and high-quality-factor niobium and coaxial cavities have all advanced rapidly in recent years, with several experiments now reaching or exceeding the break-even point for quantum error correction. What has been missing is a way to organise many such high-performance memories into an addressable, switchable whole. The cascaded random access memory provides exactly that missing layer of organisation, and it does so using components, transmons, cavities and parametric couplers, that are already compatible with existing superconducting processor fabrication.

There remain, of course, substantial challenges between this eight-bit demonstration and the megabyte-scale ambitions of quantum computing. Error rates must fall further, storage times must lengthen relative to operation times, and the many-body interaction effects that dominate the current error budget must be engineered away or incorporated into error models. The researchers have made their data and simulation codes openly available through figshare, an invitation for the community to probe, reproduce and extend the results. Yet the conceptual milestone is unambiguous. Classical computing took its decisive step towards scalability when memory was separated from logic and made randomly accessible. Quantum computing has now taken an analogous step, demonstrating that a single control qubit can reach into a protected reservoir of quantum states and retrieve any one of them on demand. If the approach scales as its designers hope, the random access quantum memory may come to be seen as the moment quantum hardware began to acquire the architectural maturity that classical machines have enjoyed for decades.

Subject of Research: Demonstration of an eight-bit cascaded random access quantum memory using superconducting circuits, transmons and multimode storage cavities for scalable fault-tolerant quantum computing.

Article Title: A cascaded random access quantum memory

Article References: Li, Z., Gupta, E., Zhao, F., Banerjee, R., Lu, Y., Roy, T., Oriani, A., Vrajitoarea, A., Chakram, S., & Schuster, D. I. (2026). A cascaded random access quantum memory. Nature Physics. https://doi.org/10.1038/s41567-026-03418-w

Image Credits: AI Generated

DOI: 10.1038/s41567-026-03418-w

Keywords: quantum memory, random access memory, superconducting circuits, transmon qubit, multimode cavity, quantum computing, fault tolerance, quantum error correction, circuit quantum electrodynamics, logical qubits, cascaded architecture, quantum architecture