Meter-scale L-couplers advance IBM’s modular path to fault-tolerant quantum computing

IBM researchers have published results showing four superconducting interconnects, called L-couplers, linking two 120-qubit IBM Quantum Nighthawk R1 processors over 80 cm of aluminum coaxial cable. The connection achieved an average relay-based state-transfer fidelity of 99.00% across the four links, with the best link reaching 99.29%.

The result answers whether separate processor modules can exchange quantum information with sufficient quality to function as a single machine. IBM also reports distant CZ gates between qubits on separate chips with a mean fidelity of 94.75%, below the roughly 99.5% typical of on-chip transmon gates. 

The 99% figure measures state transfer, while the gate figure better indicates how much engineering remains.

Technical Details

Each L-coupler links two Nighthawk lattice qubits through a chain of dedicated components that sit outside the Nighthawk square lattice and preserves compatibility with Nighthawk processors already available on IBM’s cloud.

The key details include:

• Link topology: Each L-coupler connects two lattice qubits via two frequency-tunable L-qubits, two RF-SQUID-based i2m couplers, and 80 cm of superconducting aluminum coax. The links operate near 5 GHz.
• Cable quality: The standing-mode internal quality factor averages 0.94 million, and one test cable reached 2.0 million. IBM identified trapped magnetic flux as the aluminum passes through its superconducting transition as a loss mechanism, with the quality factor falling from 1.5 million at zero field to 250,000 at 14 Gauss. Multi-layer high-permeability magnetic shielding braids restore quality factors near one million.
• Assembly: Cables mount perpendicular to the chip and join via indium-to-indium pressure bonds, held by spring-loaded fixtures. IBM installs them at room temperature after both modules are in the cryostat, and the links survive three thermal cycles.
• State transfer: Relay-based transfer, measured via network benchmarking, averages 99.00% across the four links. The best link reaches 99.29 ± 0.07%.
• Distant gates: A distant CZ gate combines two state transfers, three local CZ gates, and four single-qubit gates. It achieves an average fidelity of 94.75%, with a best link of 96.42 ± 0.28%. IBM also generated Bell pairs simultaneously across two adjacent L-couplers.
• Multipartite entanglement: A four-qubit GHZ state across the two chips measured 77.2% raw and 91.7% with error mitigation. A six-qubit GHZ ring state measured 67.7% raw and 90.9% with mitigation.

IBM attributes 53% of the distant CZ error to the two state transfers and 40% to the three local CZ gates. It states that dephasing in the L-qubits, which IBM built without its standard coherence-optimization steps to speed development, drives most of the transfer error.

IBM projects that integrating the coupler junctions directly into lattice qubits and extending L-qubit lifetimes to 100 microseconds would reduce transfer error below 0.1% and increase distant CZ fidelity above 99%.

The demonstration also includes a development shortcut. IBM wired and measured only a subset of qubits around the L-couplers, so the full 120-qubit chips did not operate as complete systems during the experiment.

Analysis

IBM has shifted its quantum messaging from qubit counts to connectivity, circuit depth, and error correction, and this result extends that shift. Linking chips of manageable size lets IBM avoid building impractically large single processors and gives the company a physical foundation for its Cockatoo milestone, which entangles two Kookaburra modules in 2027, and for Starling in 2029.

The experiment uses production-class 120-qubit Nighthawk R1 processors, distinguishing it from earlier interconnect work that used fewer qubits or employed architectures incompatible with large processors.

Competitive Impact

IBM competes on modular superconducting scaling with chiplet builders, academic groups, and vendors that favor single-chip scaling. Photonic links form a longer-term alternative for very large systems.

The table below summarizes the alternatives.

IBM’s differentiation is strongest in scale and packaging. Few interconnects join processors of this size, and post-mount indium bonding addresses a manufacturing problem that bench-top demonstrations avoid.

Differentiation is weakest in gate quality. SUSTech reported a transfer fidelity equal to IBM’s headline, and prior cable results were near 80% to 90% before work like this raised the bar.

Rigetti’s chiplet position and the long-run case for optical links give competitors credible alternatives if IBM’s coherence improvements arrive late.

Final Thoughts

The announcement shows IBM delivering a credible hardware result. It connected two production-class 120-qubit processors with four meter-scale links, assembled them after cryostat mounting, and achieved 99% average state transfer during thermal cycling.

The error analysis is candid about where fidelity is lost, and the researchers identify yield and reproducibility as open problems.

Questions remain on two fronts:

• Distant-gate fidelity of 94.75% leaves a gap to the 99% level that fault-tolerant codes favor, and IBM’s path to closing it depends on L-qubit coherence improvements it has projected but not yet demonstrated.
• The experiment wires a subset of qubits, and four links are a small number compared with the connectivity a 1,000-qubit system needs.

For customers and investors tracking IBM’s quantum program, the practical test is delivery against the roadmap over the next two years, starting with Kookaburra and Cockatoo. IBM has shown that the interconnect works at the scale of its production processors, and 2027 will show whether it works at the scale of a fault-tolerant machine.