IonQ Entangles a Trapped Ion With a Diamond Memory 1,000 Times a Second at 88% Fidelity. Its Own Paper Puts the Rate Needed Between Modules 10 to 100 Times Higher.

October 9, 2026 – IonQ said on Friday that its researchers had entangled a trapped-ion qubit with a solid-state quantum memory more than 1,000 times per second over a photonic link. The company announced the result in a press release. The preprint describing it, “Quantum Networking at the Speed of Quantum Computation: Kilohertz Entanglement in a Heterogeneous Quantum System,” was submitted to arXiv on 7 October by 39 authors from IonQ and Duke University.

The authors put the rate at about four times that of the fastest earlier photonic link between two trapped ions, the 250 pairs per second reported in 2024 by IonQ co-founder Christopher Monroe’s group at Duke. They also cited estimates that continuous logical operations between modules need 10,000 to 100,000 pairs per second, at an error rate about ten times that of local two-qubit gates. One pair per millisecond, they wrote, matches the cycle time of planned trapped-ion processors.

The link connects a single barium-138 ion to a silicon-vacancy center, an atomic-scale defect in diamond, coupled to a nanophotonic cavity in a dilution refrigerator at about 100 millikelvin. The ion emits a 493.5-nanometer photon whose polarization is entangled with the ion’s spin. Frequency conversion shifts the photon to 737.45 nanometers, and an interferometer re-encodes it into early and late time bins for the memory.

The cavity reflects the photon only when the memory’s electron spin is in one particular state. A single detected photon heralds the result, and each click marks one entangled pair between the ion and the memory. The authors reported an average of 1,032 pairs per second at a fidelity of 87.9% to the target Bell state, a maximally entangled two-qubit state.

The experiment ran 1.1 million attempts per second. A heralded pair took a mean of 731 attempts, or 969 microseconds once cooling and reset time are counted.

The authors left a 664-microsecond verification measurement of each pair out of the quoted rate. With that step included, they reported 612 measured pairs per second. They wrote that in a distributed quantum computer, gates would consume each pair in place of that measurement.

Carsten Robens, the corresponding author, and his colleagues attributed the lost fidelity to four subsystems. The ion accounted for 4.3 percentage points, the memory for 3.9, frequency conversion for 2.9 and undetected photons that had reached the memory for 1.9. Combining the four, they predicted a fidelity of 87.4%, in line with the measurement.

They wrote that none of the dominant errors is fundamental. They projected that better coupling of the ion’s light into fiber, a more efficient signal chain and a phase-based spin-photon gate would raise the rate above 5,000 pairs per second.

The authors also wrote that the memory’s silicon-29 nuclear spin can store entanglement during the generation of new links. They wrote that these capabilities can be combined to entangle ions in separate processors. In this experiment the nuclear spin was initialized but stored no quantum information, and the authors did not report ion-to-ion entanglement through the memory.

Strained silicon-vacancy centers have operated at up to 1.5 kelvin in earlier experiments, the authors noted, within reach of compact closed-cycle cryostats. They made the data and analysis code for every figure available on GitHub. They also disclosed IonQ’s patent applications on the work and their use of AI tools from Anthropic and OpenAI for readability and plotting code.

The authors reported a script error that affected the first attempt of each batch in several characterization datasets. The dataset behind the headline rate and fidelity was unaffected.

Niccolo de Masi, IonQ’s chairman and chief executive, described the result in the release as a breakthrough for interconnects. He likened IonQ’s plan to the growth of classical data centers through networked processors and memory. Mihir Bhaskar, IonQ’s senior vice president and general manager of quantum technologies at SkyWater, the company’s foundry subsidiary, said a photonic interconnect need not limit distributed quantum computing.

Monroe, IonQ’s chief scientist and a Duke professor, said any large-scale quantum computer will need to move qubits through photons. In the release, IonQ called the link the world’s first quantum memory-enhanced interconnect. IonQ also claimed the fastest interconnect rate between qubits of any platform. The company said the rate was “fast enough to support distributed quantum computing and other use cases.”

IonQ announced the completion of its acquisition of Lightsynq, a Boston quantum-memory startup co-founded by Bhaskar, in June 2025. The company said it uses the same approach in DARPA’s Heterogeneous Architectures for Quantum program. It announced sales of the memory platform to the University of Maryland in April and to SDT in South Korea in September.

My Analysis

In September I wrote that IonQ’s 20,000-ion estimate for breaking secp256k1 put the whole machine in one contiguous grid, with no boundary between modules. IonQ’s current roadmap lists a photonic interconnect from 2029. I asked IonQ how its modules would connect and at what cost. IonQ has now published the first measured numbers for one candidate link, from the quantum-memory team it acquired with Lightsynq.

IonQ’s link produces one entangled pair per millisecond between a trapped ion and a diamond memory, about four times the best photonic rate between two ions. In the fault-tolerant designs an MIT-led group surveyed in July, each module supplies tens to hundreds of pairs per error-correction cycle at about 99.9% fidelity, against the 87.9% IonQ measured.

Links between trapped ions have run at a few hundred pairs per second at best, limited by photon loss. IonQ uses the memory to cut the penalty for that loss. One link of this kind delivers a tenth to a hundredth of the rate IonQ’s own paper cites for continuous logical operations between modules, so matching that rate would take 10 to 100 such links in parallel, or faster ones. Its 12.1% error would have to fall about a hundredfold to reach the 99.9% those designs assume, or be distilled at a further cost in rate.

One Detected Photon Instead of Two

Most fast links between trapped ions use two-photon heralding. Each ion emits a photon and the two photons meet at a beamsplitter. The pair counts as entangled only when both photons are detected. The success probability therefore scales with the product of the two nodes’ collection efficiencies, which is the square of either one when they are equal. Monroe’s group reached 250 pairs per second in 2024 with a lens of numerical aperture 0.8 inside the vacuum chamber for each ion, at a million attempts per second, succeeding on about one attempt in 4,000.

In IonQ’s scheme only the ion emits. The memory is a silicon-vacancy center in a nanophotonic cavity, the kind of diamond defect whose coherence a Harvard group tripled using sound in a study I covered in August. The ion’s photon reflects off the cavity when the spin is in the right state. A single click at the detectors then heralds the pair, and the success probability scales with the collection efficiency itself, not its square.

On 3.3% of attempts, the barium ion’s photon reaches the fiber at the end of its collection optics. Frequency conversion, filtering, two passes through the time-bin interferometer, fiber losses and detection together pass 13.4% of those photons. The memory’s fiber coupling and reflection pass 35%. A photon already in the ion’s fiber therefore has less than a 5% chance of producing a herald.

IonQ’s link still heralded once per 731 attempts. The Duke link needed about 4,000 with no frequency conversion at all, because in a two-photon scheme the losses of both nodes multiply.

The ion side also needs no cavity and no optics inside the vacuum chamber. IonQ’s 0.75-numerical-aperture objective is outside the vacuum window. The authors note that dielectric surfaces near trapped ions degrade gate fidelity and coherence. For a module that must also compute, distance between the optics and the qubits is an advantage.

I think the memory-reflection scheme is the right direction for ion links. In July, Felix Knollmann, Isaac Chuang and colleagues at MIT, Oxford and MIT Lincoln Laboratory described an architecture with the same linear scaling, reached by another route. Their perspective proposes single-photon heralding directly between two ions, with one round of entanglement distillation and photonics integrated into the trap. They project 858 pairs per second per channel at 99.9% fidelity, with many channels per module and no separate solid-state memory. The two groups agree on the scaling and disagree on whether the diamond is necessary.

What IonQ Compared Its Link Against

The record IonQ compared against was set between two barium ions in the same trap, at a fidelity of at least 93.7% after correction for state-preparation and measurement errors. IonQ reported its 87.9% without that correction. The Duke link entangled two qubits a processor could compute with. In IonQ’s new link, one end is a diamond memory.

Linking two ion processors through the memory requires at least two ion-to-memory links and a joining operation. The authors wrote that the pieces exist and can be combined. We have no ion-to-ion number for the architecture yet. Each added step costs rate and fidelity, and under a simple depolarizing-noise model with a perfect swap, two links at 87.9% would join to about 78%.

In March, a team at the University of Science and Technology of China reported entanglement between a trapped ytterbium ion and a europium-doped crystal memory 75 meters away in a separate lab, at 89.2% fidelity and 0.2 pairs per second. The crystal stored the ion’s photon and released it for measurement, so the USTC team confirmed the entanglement only when the retrieved photon was detected. IonQ’s herald leaves an entangled pair in the ion and the memory spin, available for use.

The IonQ authors cite the USTC result and credit the kilohertz rate to their single-spin memory and to heralding on one detected photon. IonQ’s reported rate is about 5,000 times USTC’s, for a different kind of link. The IonQ authors do not say how far their ion trap was from their refrigerator.

IonQ left verification out of the headline rate. Each verification measurement took 664 microseconds. With that time counted, the authors reported 612 pairs per second. The Duke figure is likewise attempt rate times success probability, so the fourfold comparison is like for like. In a computer, the IonQ authors argue, gates would consume each pair in place of the measurement. For the ion-to-ion version, the number I’ll want is the sustained rate that reaches the consuming computation, with initialization, reset, storage, routing and joining included.

The Bell-Pair Budget of an Error-Corrected Module

A modular fault-tolerant computer consumes Bell pairs, two-qubit entangled states shared across a boundary, to run error-corrected operations between modules. In their July perspective, Knollmann and colleagues worked through recent fault-tolerant designs, including the Pinnacle architecture and Cain and colleagues’ 10,000-atom Shor estimate. Each module, they found, must supply tens to hundreds of Bell pairs per error-correction cycle at 99.9% fidelity or better. At the one-millisecond cycle those designs assume, that comes to 10,000 to a million pairs per second per module.

The IonQ authors cite 10,000 to 100,000 pairs per second, at an error rate about ten times that of local two-qubit gates. Their headline claim, networking at the speed of computation, amounts to one pair per one-millisecond cycle of planned ion processors. A module needs tens to hundreds of pairs per cycle, one to two orders of magnitude more than one. Running tens to hundreds of links like this one in parallel for each module would supply the difference.

IonQ’s best two-qubit gate, from the Oxford Ionics team it acquired, has an error of $$8.4 \times 10^{-5}$$ on two ions. The ten-times relation the IonQ authors cite comes from Ramette and colleagues, who found that surface-code patches tolerate about 14 times more noise at the seam between them than in their bulk. For gates of IonQ’s quality, a link that barely affects the code would have an error near $$10^{-3}$$, or 99.9% fidelity, the level the MIT-led survey uses.

Ramette and colleagues also found that at a bulk error rate near 1%, the codes still reach threshold with seam errors near 10%. IonQ’s measured Bell-pair error is 12.1%, above that 10% and about 120 times the 0.1% level. Adding parallel channels raises the rate and leaves the error where it is.

IonQ’s own designs cycle more slowly than one millisecond. In the September secp256k1 design, each syndrome-extraction cycle takes 5.4 milliseconds. Holding the pairs needed per cycle fixed, which a different code and layout would change, that slower clock scales the 10,000 to 100,000 range down to roughly 2,000 to 20,000 pairs per second. The rescaling is arithmetic, and IonQ has not published the pair demand of its own module crossings.

Entanglement distillation turns several noisy pairs into fewer, better ones, and the cost depends on the kind of noise. I checked two simple cases with the standard two-to-one recurrence protocol. With depolarizing errors, one round lifts 87.9% to 90.9% and a second round to 93.4%. With pure phase errors, two rounds reach 99.96%, and after counting the rounds that fail, 1,032 raw pairs per second would yield about 200 distilled ones. Neither case is a forecast for IonQ’s link, and the preprint gives no distillation design to calculate from.

IonQ says in its release that the rate is already enough for distributed quantum computing. The authors call 1 kHz “a step toward the 10–100 kHz rates required for distributed quantum computation.” A demonstration across two modules can run on far fewer pairs per second. In a 2025 Nature paper, David Lucas’s group at Oxford ran Grover’s search algorithm across two trapped-ion modules on Bell pairs generated at 9.7 per second, at 96.9% fidelity. The release does not say which kind of distributed computing it means, and for error-corrected modules the authors’ own figure applies.

The Four Error Sources in IonQ’s Link

The ion costs 4.3 percentage points of fidelity, mostly through optical-pumping light that leaks through the modulators after a herald and through ion temperature and lens aberrations. The memory costs 3.9, about three of them from imperfect spin-dependent reflection and from nuclear-spin flips during a run. Frequency conversion costs 2.9 through noise photons that cause false heralds. Photons lost between the memory and the detectors cost 1.9. That term grows with each attempt before a reset.

The authors cap each batch at 15 attempts before resetting the memory. In their tests, more attempts per reset gave a higher rate and a lower fidelity. In the supplementary measurements, removing the pre-triggered pumping light raised the bound on the ion-photon fidelity from 94.4% to 97.0%, at roughly half the attempt rate.

Knollmann and colleagues named two weaknesses of silicon-vacancy links in July, lower-fidelity operations in solid-state qubits and coherent single-photon frequency conversion. In IonQ’s budget, the memory and the conversion account for about half of the four standalone subsystem estimates, 6.8 of 13 points.

The authors write that none of these errors is fundamental. They name a fix for each: narrower filtering of conversion noise, higher modulator extinction, better cavity contrast and more efficient return from the memory to the detectors. I agree none is fundamental.

If the four terms simply add, they would have to shrink about 13-fold together for a 99% link and about 130-fold for 99.9%, in whatever mix of fixes the authors can manage. The authors project a rate above 5,000 pairs per second and give no projected fidelity. For a link meant for error-corrected operations, I’d want the fidelity projection first.

The Press Release and the Paper

IonQ states three things in the release more strongly than its preprint supports. The sufficiency claim conflicts with the authors’ own 10–100 kHz figure for error-corrected modules. The headline calls the link the world’s first quantum memory-enhanced interconnect, a category the release never defines. IonQ’s preprint notes that the USTC team had already entangled a trapped ion with a solid-state memory, and the release does not mention that result. The first IonQ can claim is heralded entanglement between a trapped ion and a solid-state memory, and the release could have said exactly that. The third claim is a speed record across all qubit platforms.

In 2018, Philipp Kurpiers and colleagues in Andreas Wallraff’s group at ETH Zurich entangled two superconducting qubits on separate chips, joined by about a meter of cable, at 50,000 pairs per second and 78.9% fidelity, deterministically and with microwave photons. In 2017, Mete Atatüre’s group at Cambridge heralded entanglement between quantum-dot spins about two meters apart, in separate cryostats, at 7,300 pairs per second and 61.6%.

IonQ’s link has higher fidelity than either and a lower rate. A speed record needs a defined category and the same accounting on both sides. The record IonQ’s numbers support is heralded entanglement involving a trapped ion, where its link is the fastest I know of, at four times Duke’s ion-to-ion rate.

De Masi called the result a breakthrough. The authors never use the word in the preprint.

Trade outlets repeated the release. Quantum Zeitgeist told readers the rate supports distributed quantum computing. Quantum Computing Report credited the link with the fastest heralded entanglement rate yet between separate nodes. Its write-up also gave the memory a 20-millikelvin refrigerator and a cooperativity of 92, where the paper reports about 100 millikelvin and 25.8.

The authors make their work easy to check. They give the rate with and without readout. The measured rate came out 14% below the 1.21 kHz they predicted from their efficiency budget, within the uncertainty. The error budget appears term by term, and the data and analysis code are public. They also report the script error in several characterization datasets. I’d rather read that admission in a preprint than discover the bug in a replication.

The Module Boundary on IonQ’s Roadmap

In its June 2025 roadmap announcement, IonQ described its 2028 system as two connected chips, a plan I analyzed at the time. The current roadmap lists a photonic interconnect from 2029. In April, IonQ linked two of its commercial systems by photon and published no rate, fidelity or distance. In the September secp256k1 estimate, IonQ’s authors placed all 19,397 physical qubits in one grid and priced no crossing between modules.

With the October preprint, IonQ has published numbers for half of such a crossing, an ion to a memory, at about one pair per millisecond and 87.9%.

Knollmann and colleagues note that published designs for factoring RSA-2048 need 10,000 to 10 million physical qubits. By their count, the most capable surface-electrode ion processors control a few hundred ions individually. Designers who bridge that difference with linked modules assume links much better than this one. Tian Xue and Jacob Covey assume 100,000 Bell pairs per second between modules, with one-millisecond measurements, in their modelled design for factoring RSA-2048 on a modular atomic processor. At that rate, they estimate, the factoring takes only 16% longer than on a single module.

IonQ’s new result does not move Q-Day. IonQ has now published measured numbers for a component its roadmap schedules for 2029. Designs for a cryptographically relevant quantum computer built from linked modules depend on the same component, and at 87.9% it is not ready for that role. My CRQC Quantum Capability Framework tracks modular links under qubit connectivity and routing, and this measurement is my baseline for IonQ’s next networking result.

What I Will Watch For

I’ll look for six results in IonQ’s next networking papers.

1. Ion-to-ion entanglement between two separate traps through silicon-vacancy memories, with rate and fidelity reported together, and the sustained rate that reaches a consuming computation once initialization, reset, storage, routing and joining are counted.
2. An ion-to-memory fidelity above 99% at kilohertz rates, or a distillation scheme with its rate cost stated.
3. Several channels running in parallel for one module. The authors propose addressing several pre-initialized memories in turn so the ion never pauses for a reset. That raises one channel’s rate and does not add channels.
4. A memory running outside a dilution refrigerator, at the 1.5 kelvin the authors cite for strained silicon-vacancy centers.
5. An IonQ resource estimate that prices a photonic module boundary, so the secp256k1 numbers can be rerun with the link included.
6. Remote entanglement generated while the computational qubits in each module keep their state and local processing continues.

In September I asked how IonQ’s modules would connect. The first measured answer is a diamond memory entangled with a single ion a thousand times a second, at about a hundred times the error that fault-tolerant designs assume. When the ion-to-ion paper arrives, I’ll read its fidelity before its rate.

From the author