China Set a 420 km Quantum-Memory Entanglement Record. The Two Ends Were in One Lab.

August 11, 2026 – Physical Review Letters published a result from the University of Science and Technology of China (USTC) reporting entanglement between two atomic-ensemble quantum memories through 420 kilometers of optical fiber. The paper is by Xi-Yu Luo, Chao-Yang Wang, Ming-Yang Zheng and co-authors, with Xiao-Hui Bao and Jian-Wei Pan listed last, and appears as Phys. Rev. Lett. 137, 070801.

A preprint of the same work has been on arXiv since April 2025. Contributing institutions include the Jinan Institute of Quantum Technology, the Shanghai Institute of Microsystem and Information Technology, and Yangtze Optical Fibre and Cable.

The previous published record for matter-to-matter entanglement in fiber was 50 kilometers, set by the same group in 2020 and reported in Nature 578, 240. Two later results from the group bear on the record differently. In 2022 they entangled two atomic ensembles physically separated by 12.5 kilometers, with the entanglement postselected rather than heralded. In 2024 they held that same 12.5-kilometer separation in a three-node metropolitan network, with heralded pairwise entanglement and memory lifetime exceeding round-trip communication time.

Both memory nodes, named Alice and Bob in the paper, sit in a single laboratory at USTC. Two deployed field fibers, each about 10.1 kilometers, carry their photons to a middle node, Charlie, at Hefei Software Park. Ultra-low-loss fiber coiled on spools inside the laboratory makes up the rest of each arm.

The distances quoted throughout the paper are the two arms summed. At the 420-kilometer setting, each write-out photon travels about 210 kilometers to reach Charlie.

Each node holds a cloud of laser-cooled rubidium atoms in an optical cavity. A weak write pulse produces a scattered photon and a collective atomic excitation with a probability of about 6 percent, leaving the atoms and the photon in a joint state. Difference-frequency generation in a periodically poled lithium niobate waveguide, pumped at 1600 nanometers, then converts the photon from 780 nanometers to 1522 nanometers in the telecom S-band. At 780 nanometers the signal would meet about 3.5 decibels of attenuation per kilometer. At 1522 it meets roughly 0.17 on the coiled ultra-low-loss fiber and 0.31 on the deployed field sections.

The write-out modes from both nodes interfere at a beam splitter at Charlie. A click in either of two superconducting nanowire single-photon detectors heralds entanglement between the two atomic ensembles.

The protocol is the Duan-Lukin-Cirac-Zoller scheme, proposed in 2001. It succeeds more often than two-photon protocols because only one photon has to survive the journey.

Holding the relative optical phase steady across hundreds of kilometers of fiber required two stabilization loops. A continuous 1600-nanometer probe beam suppresses high-frequency noise. An intermittent dual-band beam, detuned symmetrically around the signal frequency by roughly 675 megahertz, compensates slow drift in the path-length difference. Residual phase error came to about 7 degrees.

At 420 kilometers, total fiber loss reached 78.7 decibels. The authors report a concurrence of 0.046 with an uncertainty of 0.022 on the channel heralded by one of the two detectors, alongside a signal-to-noise ratio of 3.5 and an entangling probability of 1.09 in a million per trial.

At 320 kilometers the same figures are 0.052 with an uncertainty of 0.007, a signal-to-noise ratio of 15, and an entangling probability of 6.86 in a million. Interference visibility on the retrieved photons falls from 0.72 at 320 kilometers to 0.64 at 420.

The paper attributes the visibility limit chiefly to higher-order excitations rather than to distance, calculating a ceiling of 0.769 at the 6 percent excitation probability used. Lowering that probability to 2 percent would raise visibility at 420 kilometers to roughly 0.72, at the cost of reducing both signal-to-noise and entangling probability to a third of their reported values.

Comparable and greater distances have been covered before by other means. Zhuang and colleagues distributed entangled photons over 404 kilometers of fiber, posted as a preprint in 2024 and published in Physical Review Letters in June 2025, and the Micius satellite delivered entangled pairs to ground stations 1,200 kilometers apart in 2017. The claim in this paper is specific to matter-to-matter entanglement, in which the quantum state is held in atoms rather than carried by photons in flight.

The authors also report that entangling probability scales with the square root of channel transmittance rather than linearly with it. That places the result above the Pirandola-Laurenza-Ottaviani-Banchi bound on direct entanglement distribution once transmittance falls below one part in 30,000, corresponding in this setup to distances beyond roughly 230 kilometers. The paper notes that twin-field quantum key distribution achieves comparable scaling.

In their outlook, the authors state that the experiment can be implemented between nodes physically separated by a similar distance once phase stabilization for separated nodes is integrated. They point to high-rate operation near 100 kilometers as the regime relevant to quantum repeaters, and identify high-order excitations as the dominant limit on visibility, with Rydberg blockade schemes as a route past it.

My Analysis

Where the 420 Kilometers Are

The 420 kilometers is the sum of the two optical arms running from Alice and Bob to Charlie, not the distance between the two memories. Alice and Bob are in the same laboratory at USTC. At the longest setting, each arm runs about 10.1 kilometers of deployed field fiber and roughly 200 kilometers of coiled ultra-low-loss fiber. A write-out photon therefore covers about 210 kilometers to reach Charlie, and the two arms together reproduce a 420-kilometer channel at 78.7 decibels of total loss.

That is more than a paper calculation, and calling it a mere loss budget would undersell it. The attenuation was real, the Raman noise was real, the propagation delay was real, and holding the relative phase across that much glass is the hard part of the experiment. What the apparatus did not do is operate two cold-atom laboratories hundreds of kilometers apart under independent control. The result is a channel-length benchmark rather than a node-separation benchmark.

Building the first version on spools is also the rational order of work. It tests attenuation, propagation delay, Raman noise and long-path phase stabilization before anyone takes on the thermal, mechanical and operational instability of a deployed route between two cold-atom laboratories.

The difficulty appears in the coverage rather than in the paper. Certified fuel-economy figures come off a chassis dynamometer, where the engine is real, the load is real, and the measurement is honest to three significant figures. Nobody drives on a dynamometer. The remedy has never been to distrust the dynamometer, only to say which of the two numbers is being quoted.

Luo and his co-authors are explicit about which one they are quoting. Figure 1 puts both memories in one laboratory, and the abstract describes a testbed for studying network applications from metropolitan to intercity scale. By mid-August the South China Morning Post had it as groundwork for intercity quantum networks, and ScienceAlert had already run it as a shattered record.

One number puts the gap in proportion. The largest geographic separation between memory nodes this group has ever published is 12.5 kilometers, first reached in a postselected two-node experiment in 2022 and held again in the three-node metropolitan network of 2024. The new paper doesn’t beat it, and doesn’t claim to.

The other qualifier that keeps falling out of the coverage is the word matter. Entangled photons crossed 404 kilometers of fiber in 2024, and Micius put entangled pairs on the ground 1,200 kilometers apart in 2017. Photon distance stopped being the frontier years ago.

What is new here is that the entanglement ends up stored in atoms at both ends instead of expiring in a detector, and storage is what would make this architecture a repeater segment rather than a postselected transmission experiment. Reported as “China pushes entanglement to 420 kilometers,” the achievement reads as a distance record it isn’t, while the part that genuinely advanced goes unmentioned.

The Weakest Row in the Table

Table I lists six distances, and the 420-kilometer column is the least supported entry in it.

Concurrence at 420 kilometers is 0.046 against a quoted uncertainty of 0.022, which puts it about 2.1 times that uncertainty above zero. At 320 kilometers, 0.052 against 0.007 gives a ratio of 7.4. Concurrence is bounded below at zero and the paper reports no null-hypothesis test, so those are ratios rather than sigmas.

The second herald channel at 420 kilometers reports 0.016 against the same 0.022 and cannot be distinguished from no entanglement at all, which is why Luo and colleagues claim verification on one detector rather than on both. Signal-to-noise over that final hundred kilometers falls from 15 to 3.5, as dark counts in the superconducting detectors overtake what remains of the signal.

My reading of that table is that the 320-kilometer row is the result and the 420-kilometer row is the headline. Both are in the paper, and both were put there by the authors, which is the behavior anyone should want from a group setting a record. It is not what a reader takes away from a story built entirely on the larger number.

A smaller error is still young enough to correct. Interesting Engineering reported in mid-August that 420 kilometers exceeds a roughly 320-kilometer threshold for direct transmission, and ScienceAlert put the crossover at the same figure. The paper puts the crossover at transmittance below one part in 30,000, which it states corresponds to distances over 230 kilometers in this setup. The 320-kilometer figure is a row label from Table I that appears to have been attached to the wrong claim.

Storage Time Against Transit Time

The constraint I haven’t seen raised anywhere is the one that decides whether this is a network component or a physics measurement.

A quantum memory earns its place in a network by holding a state until the herald arrives and reports what happened. Charlie sits about 210 kilometers of fiber from each node, which the write-out photon covers in roughly 1.03 milliseconds. Luo and colleagues retrieve the atomic excitations 750 nanoseconds after creating them, and measure the resulting photons in a delayed-choice arrangement.

At the moment the memory is emptied, the write-out photon has covered about 150 meters of its 210-kilometer path. Retrieval happens before the heralding click at Charlie has even occurred. In a genuinely separated topology at this distance, a memory would have to survive the outbound millisecond, a return millisecond, and detector and control latency on top, holding fidelity well enough to be useful afterwards. Nothing in this apparatus attempts that.

The entanglement is therefore verified after the fact rather than held and used. The same group’s 2024 metropolitan paper stated explicitly that memory lifetime exceeded round-trip communication time across its 12.5 kilometers. The 2026 paper makes no equivalent claim, and the experimental sequence shows why none follows from this result.

Bao and Pan have published sub-second storage in optical lattices, and their outlook section points at that combination for entanglement swapping between segments, so the capability exists inside the group. What has not happened is putting long storage and long distance into one apparatus at one time. Until that happens, calling this a repeater segment overstates it by the one property a memory-based repeater needs.

What Beating the PLOB Bound Actually Buys

The distance record is not the only claim in the paper, and the second one will outlast it. The Pirandola bound sets the ultimate capacity of a repeaterless lossy channel per use, and at high loss that capacity falls roughly linearly with transmittance. Luo and colleagues measure their per-trial heralding probability against that benchmark. A single-photon heralded scheme with memories at both ends scales with the square root of transmittance instead, and past roughly 230 kilometers in this setup the square-root scaling gives the higher entangling probability. That scaling shows up across the distance settings, which is why it will outlive the endpoint everyone quoted.

Two qualifications follow, and the authors supply one of them. Twin-field QKD achieves the same scaling advantage with no quantum memory anywhere in the system, and twin-field links have run past 1,000 kilometers of spooled fiber already. An entangling probability of 1.09 in a million per trial, at hertz-level click rates, also describes a laboratory measurement rather than a service.

The engineering will still matter in five years when the record has moved. Converting at 1522 nanometers rather than deeper into the C-band, separated from the 1600-nanometer pump by 78 nanometers, keeps anti-Stokes Raman noise low enough for the filter stack to remove the rest. Other groups will copy that choice, and they will copy the dual-band locking scheme too. It solves a drift the supplement puts at about 1.16 radians per kelvin on a 100-kilometer write-out arm. A fifth of a kelvin between two arms therefore moves the phase by roughly 0.23 radians, or 13 degrees, against an overall stability budget the paper gives as about 7 degrees. Luo and colleagues note the scheme transfers to other platforms. I expect it does, and I expect that to be the citation this paper collects in 2031.

Nothing Here Moves a Migration Deadline

I write about Q-FUD often enough that my first instinct is to look for the vendor. Nobody is selling anything here. This is a clean academic result, with an industrial fiber manufacturer among the co-authors and no product attached, from a group that has been extending this particular number since 2020, and the paper is more careful about its own limits than any of the coverage of it has been.

The misreading to correct is a category error. Quantum networking is communication, and a cryptographically relevant quantum computer is computation. Entangling two clouds of rubidium through a spool of fiber factors nothing, changes no resource estimate for RSA-2048, and moves Q-Day by not one day in either direction. The distributed-computing line in the abstract is aspirational at a success probability of one in a million per trial, and nobody runs a distributed algorithm across a link that heralds a few times a second.

For a CISO the operational content of this paper is zero, and the strategic content is one line. Migration deadlines were never set by Chinese laboratory records. They were set by regulators, insurers, and clients, and those deadlines are already fixed on calendars anyone can read today. Data sent this year under a ten-year confidentiality requirement, protected only by quantum-vulnerable public-key cryptography, is exposed to harvest now, decrypt later regardless of what happens in Hefei.

There is also a reason Western agencies rank QKD where they do. The NSA has declined to recommend it for national security systems, citing authentication, cost, and dedicated-fiber requirements, and BSI, ANSSI and their Dutch and Swedish counterparts reached a similar conclusion in their joint position paper, finding QKD practical today only in niche cases. Memory-based entanglement is a different technology from the trusted-relay QKD those documents assess, and mature repeaters could remove the dependence on trusted relay nodes. They would not by themselves touch authentication, hardware cost, implementation assurance, or denial of service.

Where the result does belong is in the record I keep of China’s quantum networking program, as the next entry after the 2024 metropolitan network and the 2026 repeater building block. Four steps in six years, and each one moved a different variable: 50 kilometers of coiled fiber between co-located memories in 2020, 12.5 kilometers of genuine separation but postselected in 2022, a three-node metropolitan network with storage outlasting the round trip in 2024, and now a 420-kilometer channel with the memories back in one laboratory.

That is a pattern rather than a leap. It is a group that moves one limiting variable at a time, publishes the supporting table honestly, and lets everyone else supply the adjectives. We should be reading the table and ignoring the adjectives.

The sovereignty reading is narrower than the alarm and more durable. The author list puts USTC alongside the Jinan Institute of Quantum Technology, the Shanghai Institute of Microsystem and Information Technology, and Yangtze Optical Fibre and Cable. Frequency conversion, superconducting detection and specialty fiber are the three subsystems this experiment could not run without, and Chinese institutions working in each are named on the paper rather than thanked in a footnote.

That shows domestic capability across the subsystems. What it does not show is component-level sourcing: no author-contribution statement assigns each institution its part, and nobody has published a bill of materials for this apparatus. An author list says nothing about where the lasers, the cryogenics, the electro-optics, or the fabrication tools came from, and I have seen nobody produce component-level sourcing for this apparatus. The claim I will make is the modest one: the institutional capability is domestic and broad, and it will still be there when the 420-kilometer number has been beaten twice over.

What I’m waiting for is the paper where Alice and Bob are genuinely 200 kilometers apart, storage runs past the round trip, and the concurrence at the far end comes with an uncertainty bar too narrow to park a null result inside. On six years of evidence, that paper is coming. This isn’t it.