IBM's chip-to-chip links reach 99% transfer fidelity, while error-correcting codes need more from them

IBM's new links pass a quantum state from one processor to another with about 1% error per transfer. A two-qubit gate across the same links, which is the operation an error-correcting code would actually run across a chip boundary, fails about 5% of the time. Those figures are roughly 10 and 50 times the 0.1% component error rate that IBM's own fault-tolerance blueprint assumed for links between chips. The public record does not yet say how much worse than on-chip operations a link can be before it becomes the main source of a logical qubit's errors.

The result comes from a preprint posted to arXiv on 21 October 2026, which has not yet been peer reviewed, and an IBM blog post dated 27 October. The team, which includes Neereja Sundaresan, connected two 120-physical-qubit IBM Nighthawk R1 processors with four links that IBM calls L-couplers. Each one is 80 cm of superconducting aluminum coaxial cable. At each end it attaches through a frequency-tunable L-qubit (a simplified qubit that sits at the cable end) and a tunable coupler. The cables mount perpendicular to the chip surface and are joined to it by an indium pressure bond made at room temperature, after both processors are already mounted. IBM says only a subset of each chip's qubits was wired up for the experiment.

What the 99% measures

State-transfer fidelity measures how faithfully a quantum state loaded on one chip arrives on the other. IBM's blog puts it simply: the state is correctly transferred about 99% of the time on average. The paper reports a mean of 99.00% across the four links, measured with a method it calls network benchmarking, and a best link of 99.29%, plus or minus 0.07%. The cable modes, near 5 GHz, had internal quality factors of 0.86 to 1.09 million. A quality factor measures how long a resonator holds energy before losing it.

The paper reports a second figure that most coverage left out. The team built distant CZ gates (a standard two-qubit entangling operation) between lattice qubits on separate chips by passing states across the cables. These gates reached a mean fidelity of 94.75%, with a best link of 96.42%, plus or minus 0.28%. The team used them to create entangled states shared between the two processors, including two pairs generated at the same time on adjacent L-couplers.

On a single chip, IBM reports much lower error rates. IBM says that on the third revision of its Heron processor, 57 of 176 two-qubit couplings make less than one error per 1,000 operations. IBM also says Nighthawk r2 keeps Heron-class gate fidelity. So the cross-chip gate's mean error of about 5.25% is more than 50 times the error of IBM's best on-chip couplings, and the 1% transfer error is about 10 times higher.

How good a link has to be

Error-correcting codes spread one logical qubit across many physical qubits and repeatedly check them for faults. Each code has a threshold, the physical error rate below which adding more qubits lowers the logical error rate instead of raising it. A threshold is a break-even point. Useful machines need to run well below it.

A 2023 arXiv paper modeled two surface-code patches (the most widely studied error-correcting code) joined along a seam of noisy links. It found the seam could reach threshold at about 10% error while the rest of the code sat near its usual 1% threshold. It also reported that the interface could tolerate 14 times more noise than the bulk with only a small effect on threshold and on behavior below it. By that measure, IBM's 5% cross-chip gate is under the 10% limit. Applied to on-chip error near 0.1% instead of 1%, though, the same 14-fold ratio would suggest a link budget near 1.4%. The 1% transfer figure falls within that range; the 5% gate does not. This is a rough extrapolation, and the model covers surface codes only.

IBM's plan uses a different code family. Its June 2025 architecture paper, posted to arXiv as a preprint, builds its design on bivariate bicycle codes, which IBM says can run about ten times larger logical circuits on a given number of physical qubits than surface-code designs. The paper assumes a link device it calls a Bell-coupler. This device would create a high-fidelity entangled pair between modules in less time than one round of error checks. The number of couplers needed between two modules grows with the code distance, which is 12 and 18 for the two codes studied. Its simulations gave every component, Bell-couplers included, the same error rate: 0.1% or 0.01%. The authors write that Bell-couplers may be as much as an order of magnitude worse than gates. They say relaxing that assumption would make the estimates worse by an unknown amount, and that operations between modules already look particularly limiting.

So the direct answer is that no published analysis gives a link-error threshold for IBM's own codes. The surface-code work suggests links can be about an order of magnitude noisier than on-chip operations. IBM's own resource estimates assumed they would be equally good. Two things would settle the question: a simulation of the bicycle architecture with a separate error rate for the links, or an experiment that runs a code across the L-couplers and measures logical error.

Other ways to join modules

Other hardware types face the same problem with different tools. Trapped-ion, superconducting and Rydberg-atom processors are each expected to hit size limits somewhere between 100 and 10,000 physical qubits per device. In February 2025, University of Oxford researchers reported in Nature on two trapped-ion modules about two meters apart, linked by photons. They used heralded entanglement (entanglement whose success is confirmed by detecting photons) to teleport a CZ gate between the modules at 86% fidelity, and ran Grover's search algorithm across both modules with a 71% success rate. For neutral atoms, Princeton researchers proposed, in a January 2024 preprint, a cavity-based link for ytterbium atoms. They estimate it could produce about 100,000 entangled pairs per second at fidelity near 99.9%; those numbers are projections, not measurements.

These results use different metrics and different distances, so they cannot be compared directly. Photons can connect many kinds of qubits and travel between separate apparatus, while IBM's microwave cables run between chips that share one cryogenic environment.

What IBM says comes next

IBM says it expects to connect up to nine Nighthawk modules with L-couplers by 2028, giving 1,080 connected physical qubits. It says a processor called Cockatoo will show entanglement between error-corrected modules in 2027. It also says Starling, a fault-tolerant machine designed to run 100 million gates on 200 logical qubits, will be built in Poughkeepsie, New York, by 2029. For the couplers themselves, IBM says it plans to improve yield, reproducibility and error rates, and to add more links that can be plugged and unplugged. It says it has room to improve every metric in the paper. The figures to watch are the cross-chip gate error compared with on-chip error, and the logical error rate of operations that span two chips.