There is a ceiling problem in quantum hardware that doesn't get discussed as often as qubit counts do.
Every approach to building a quantum processor gets harder as the device grows. Superconducting chips need a control line for essentially every qubit, and those lines carry heat into a dilution refrigerator that has a finite cooling budget. Trapped-ion chains get slower and more fragile as you add ions to a single trap. Crosstalk between neighbouring qubits rises with density. Fabrication yield — the probability that every qubit on a chip is good — falls off a cliff as area increases. None of these are fundamental physics barriers. They are engineering walls, and engineering walls have a habit of arriving sooner than expected.
The alternative is the one classical computing took decades ago: stop building one bigger machine and start connecting smaller ones. That is what distributed quantum computing means, and it is why quantum networking has moved from a communications curiosity to a scaling strategy. When we read EO 14413 as a technical document, distributed quantum computing showed up as a named five-year planning target — which is a fairly direct admission that nobody is confident the single-big-chip route gets all the way there.
The catch is that connecting quantum computers is nothing like connecting classical ones.
Three different things people call "quantum networking"
Before anything else, a disambiguation, because these get conflated constantly and the confusion makes most coverage of the topic unreadable.
(a) Quantum networking for distributed computation. Linking two or more QPUs with a quantum channel so they behave as one larger computer. Quantum information genuinely moves between them. This is the subject of this post.
(b) Quantum key distribution and the "quantum internet." Using quantum states to establish a shared secret key whose security rests on physics rather than on computational hardness. This is a communications security technology, descending from Bennett and Brassard's BB84 protocol. It does not make any computer faster. It's also frequently misrepresented as the answer to the quantum threat against public-key cryptography — in practice that threat is being addressed by post-quantum cryptography, which is software you can deploy today, not hardware that needs new fibre.
(c) Classical networking of quantum computers. Running one circuit on a QPU in Maryland from a laptop in Berlin. This is what every cloud quantum service already does. It is ordinary TCP/IP carrying circuit descriptions and measurement results, and nothing quantum crosses the wire.
Press releases blur (a) and (c) especially often. If a headline says "networked quantum computers" and the mechanism is an API, it's (c).
Why you cannot just amplify a qubit
Classical networking works because signals can be copied. A repeater reads a degraded bit, decides whether it was a 0 or a 1, and transmits a fresh clean copy. Errors are stripped out at every hop, which is why a fibre link can span an ocean.
Quantum information forbids this. The no-cloning theorem says there is no operation that duplicates an arbitrary unknown quantum state. There is no such thing as reading a qubit to see what it is and sending a fresh one, because measurement collapses the state and destroys the very superposition you were trying to preserve. A quantum repeater that worked like a classical repeater would be a quantum repeater that deletes your data.
Meanwhile the physical channel is brutally lossy. Photons in optical fibre attenuate exponentially with distance, and unlike classical light you can't compensate by turning up the power — a single photon either arrives or it doesn't. Beyond a few hundred kilometres, the odds of a photon surviving the trip get small enough that you'd wait a very long time for one success.
Quantum repeaters solve this by a genuinely different mechanism. Rather than relaying the qubit, they relay entanglement. Split a long link into short segments. Establish an entangled pair across each segment independently — short hops, so the loss per hop is manageable, and success can be heralded, meaning you know when it worked and can simply retry when it didn't. Then perform entanglement swapping: at each intermediate node, jointly measure the two local halves of two adjacent pairs in the Bell basis. That measurement consumes both short pairs and leaves the two far-end qubits entangled with each other, despite never having interacted. Chain this along the route and you have entanglement spanning the whole distance.
The ingredients this demands are exactly the hard parts: heralded entanglement generation, quantum memories that can hold a half-pair coherently while the neighbouring segment keeps retrying, and some form of purification or error correction to keep fidelity from degrading across hops. Decoherence in the memory sets a hard clock on the whole procedure — if a stored qubit dies while its partner segment is still failing, the attempt is wasted.
Teleportation is the transport primitive
Once two nodes share an entangled pair, moving an actual qubit between them is a solved protocol: quantum teleportation, from Bennett and colleagues in 1993.
The sender performs a joint measurement on the qubit to be transmitted and their half of the entangled pair, obtaining two classical bits. Those bits are sent over an ordinary classical channel. The receiver applies one of four corrections determined by those bits, and their half of the pair becomes the original state. One entangled pair and two classical bits per qubit transported.
Two consequences are worth stating explicitly because both get mangled in popular coverage. First, the original state is destroyed by the sender's measurement — teleportation moves a qubit, it does not copy one, which is precisely how it stays consistent with no-cloning. Second, the classical channel is mandatory. Without those two bits the receiver's qubit is in a completely random state and useless. Entanglement alone signals nothing. Nothing travels faster than light, and no amount of entanglement between two labs lets them communicate without a conventional link.
That second point is the whole reason quantum networks are not a physics loophole. They are a way to move fragile information, not a way to move it instantly.
The interconnect is the bottleneck
Now the engineering reality of stitching QPUs together.
Inside a single processor, two-qubit gates take microseconds or less and land fidelities that, on good hardware, exceed 99.9%. Between two processors, you must first generate a shared entangled pair over a photonic link — a probabilistic, lossy process — and only then teleport a gate or a qubit across it. Inter-node entanglement generation is currently orders of magnitude slower than local gates, and arrives at meaningfully lower fidelity.
This changes how you have to think about programming such a machine. A distributed QPU is not a flat pool of qubits. It is a strongly non-uniform architecture where some pairs of qubits are cheap to entangle and others are extremely expensive, and where the expensive operations also happen to be the noisy ones. Circuit compilers have to partition algorithms to minimise cross-node operations, in much the same spirit as minimising communication in classical HPC — except the penalty for getting it wrong is not just latency but error accumulation. And because entanglement must be produced faster than the memories holding it decohere, a slow interconnect doesn't merely reduce throughput; past a certain point it stops working at all.
There is a genuinely appealing upside, though, and it's the reason serious people pursue this: error correction needs physical qubit counts in the millions for useful algorithms, and no one has a credible plan to put a million high-quality qubits in one enclosure. Modularity is one of the few routes to that number that doesn't require a single manufacturing miracle.
Modality matters here more than usual
The photonic interface is where hardware platforms diverge sharply.
Trapped ions have a natural optical interface. An ion can be excited with a laser and made to emit a single photon whose polarisation or frequency is entangled with the ion's internal state — the qubit is already coupled to light at optical wavelengths that fibre transmits well. IonQ argues that this photonic compatibility was a central reason it chose trapped ions in the first place, and that the platform is particularly suited to multi-QPU networking. That is an architectural argument from a vendor whose commercial strategy depends on modular scaling, and should be read as advocacy rather than a neutral survey — but the underlying physics of atom-photon coupling is standard, well-studied quantum optics and is not in dispute.
Superconducting qubits face a harder problem. They operate at microwave frequencies inside a millikelvin refrigerator, and microwave photons cannot travel any meaningful distance at room temperature — thermal noise swamps them. Networking them optically requires microwave-to-optical transduction: coherently converting a gigahertz photon into a telecom-wavelength one without destroying the quantum state. This is an active and legitimate research field, with electro-optic, optomechanical, magneto-optic and atomic-ensemble approaches all under investigation, but efficiency and added noise remain the limiting factors. IonQ makes this contrast a talking point, characterising the conversion process as slow and not yet well matched to quantum computers; the general difficulty is widely acknowledged in the literature, though the framing is naturally sharper coming from a competitor. NIST, among others, is building optical channels for remote microwave entanglement precisely because the problem is considered tractable rather than hopeless.
Neutral atoms sit closer to the ion story optically; photonic qubits are their own case, since the information is already in flight. Our modality comparison covers the broader trade-offs.
Where this actually stands
Plainly: early.
The building blocks are real. Ion-photon entanglement — the first milestone, where a photon leaves a trapped ion carrying entanglement with it — has been demonstrated, and IonQ reports having done so outside an academic setting. Remote ion-ion entanglement has followed, with IonQ describing a setup that collects photons from two separate trap wells, interferes them at a shared detection hub, and leaves the two distant ions entangled. IonQ frames these as the first and second of four milestones on its own roadmap to photonic interconnects, with the remaining two being the transfer of that remote entanglement onto computation qubits via swap gates, and programmable multi-QPU entanglement using single-photon switching. In 2026 the company announced it had photonically interconnected two independent trapped-ion systems.
Read those carefully. They are milestone announcements on a vendor's internal roadmap, and the public write-ups notably do not disclose entanglement rates, fidelities, or success probabilities — the numbers that would let anyone judge whether the link is fast and clean enough to compute across. IonQ itself acknowledges low entanglement rates as a current limitation. Its broader case — that networking is what carries quantum computing from research demos to commercial scale, and that modular systems are faster to build and easier to service than monolithic ones — is a coherent argument, but it is a company describing why its own architecture wins.
What is not in doubt is that the demonstrated capability sits far below what distributed computing requires. Getting two qubits in different traps entangled occasionally is a real physics achievement. Sustaining entanglement generation at rates and fidelities that let two processors execute a single fault-tolerant algorithm together is a different order of problem, and nobody has done it.
The takeaway
Quantum networking is not a faster internet and not a security product — though it shares physics with one. It is a proposed answer to a hardware scaling wall, built on entanglement distribution, entanglement swapping, and teleportation, all constrained by the fact that quantum information cannot be copied and therefore cannot be amplified.
If you're evaluating claims in this space, three questions do most of the work. Is the link actually quantum, or is it an API? What is the entanglement generation rate and fidelity between nodes — not whether entanglement was achieved once? And is the demonstration between two nodes in adjacent racks, or over a distance where photon loss genuinely bites?
Answers to those tend to be absent from the announcements, which is itself informative. The glossary covers the vocabulary if you want to read the primary sources yourself.