Hadamard Get the app

IBM’s L-Coupler Explained: Linking Quantum Chips on the Road to 2029

IBM's L-Coupler Explained: Linking Quantum Chips on the Road to 2029 — Hadamard

An IBM l-coupler is a long-range interconnect that links separate quantum processors. IBM uses the name to set it apart from c-couplers, which connect qubits on a single chip. IBM described the approach in a 27 October 2026 blog post on connecting quantum processors with l-couplers. From 2027 onward, its roadmap depends on these links to join processor modules into larger machines.

That makes the IBM l-coupler one of the most important parts of IBM’s plan for fault-tolerant quantum computing. This explainer covers what an l-coupler is, what IBM has already shown with meter-scale links, how to read the fidelity numbers that come with announcements like this, and where the part fits on the road to IBM’s Starling system in 2029.

The October result in brief: IBM linked two 120-qubit Nighthawk processors with four l-couplers and moved quantum states between them at an average fidelity of 99.00%. Gates across the link averaged 94.75%. Both figures come from IBM’s preprint, and the gap between them is the most important thing to understand about the announcement.

What is an IBM l-coupler?

According to IBM’s l-coupler blog post, by Ryan Mandelbaum and Neereja Sundaresan, “long-range interconnects” between processors are called l-couplers. On-chip couplers are called c-couplers. IBM says the l-couplers are connected after the processors are mounted in the dilution refrigerator.

That order matters. The link is made while the system is assembled, not while a single chip is fabricated. Each processor stays its own piece of hardware, and the l-coupler joins processors inside the same cold environment.

“Long-range” alone does not mean l-coupler. IBM’s experimental Loon chip, announced in November 2025, uses long-range c-couplers and qubit reset to support qLDPC error-correcting codes. Those long-range connections stay on one chip.

c-coupler l-coupler
What it connects Qubits within one processor (including long-range on-chip links, as on Loon) Qubits on separate processors
When it is made As part of the chip After the processors are mounted in the dilution refrigerator
Roadmap role Loon’s error-correction experiments Cockatoo (2027) and IBM’s 1,000-qubit target for 2027

Why link quantum processors with cables instead of building one bigger chip?

A single superconducting chip cannot grow forever. Fabrication equipment caps the physical size. Yield gets worse as chips get larger, because one defect can ruin a bigger area. Each qubit also needs control wiring that has to fit through the cryostat.

IBM’s current processors show how much one chip already carries. Nighthawk, announced on 12 November 2025, has 120 qubits and 218 tunable couplers arranged in a square lattice, so each qubit connects to four neighbors. According to The Next Platform’s coverage, that is more than 20% more couplers than IBM’s Heron. IBM claims about 30% more circuit complexity.

Modularity avoids the single-chip limit. You build processors of a size you can manufacture reliably, then connect them with meter-scale links inside shared cryogenic hardware. The trade-off is that the link becomes a new source of error, and it has to be good enough to be worth using.

What IBM has shown so far: meter-scale gates and a shared cryostat

A remote CNOT over a meter-scale link (2025)

In an IBM preprint by Heya et al. (arXiv 2502.15034, first posted 20 February 2025 and revised 23 October 2025), researchers reported a remote CNOT gate with 93.3% fidelity, measured by randomized benchmarking over a meter-scale microwave link. The gate was not a single physical interaction. It was built from local CZ gates plus state transfers across the link.

That structure explains why chip-to-chip gates are harder than they sound, which the next section covers.

Two cryostats, one cold environment (August 2026)

On 19 August 2026, IBM said it had joined two modular cryogenic systems into one environment and cooled them. They reached 4 K in under five days and later went below 15 mK. Each module’s enclosure has up to 12 times the wiring space of IBM’s most widely used systems.

The same report gives IBM’s target: at least 1,000 programmable qubits by 2027, using l-couplers to connect processors. The shared cryostat provides the cold space, and the l-coupler provides the quantum link between chips inside it.

Four l-couplers between two Nighthawk chips (October 2026)

The l-coupler result itself is described in the preprint “Meter-scale quantum interconnects between large quantum processors” (arXiv 2610.29986), submitted on 21 October 2026 by Neereja Sundaresan and 41 co-authors. It has not been peer-reviewed yet. The team connected two 120-qubit IBM Nighthawk R1 processors with four l-couplers. Each link joins two Nighthawk lattice qubits through two frequency-tunable “L-qubits”, two RF-SQUID-based couplers and 80 cm of superconducting aluminum coax.

Two engineering details matter for scaling. Each cable is mounted perpendicular to the chip surface, which leaves room for denser connections. It is joined to the chip with an indium-to-indium pressure bond at room temperature, made after both modules were already mounted. According to the IBM blog post, the experiment wired out only a subset of the qubits for development purposes, although the chips themselves are full Nighthawk processors.

The headline numbers from the preprint:

  • State transfer: mean fidelity of 99.00% across the four links (network benchmarking), best link 99.29 ± 0.07%.
  • Distant CZ gates between qubits on different chips: mean fidelity 94.75%, best link 96.42 ± 0.28%.
  • Entanglement: the distant CZ gates were used to create entangled states across both processors, including two Bell pairs generated simultaneously on two adjacent l-couplers.

IBM says the next steps are better yield and reproducibility, lower error rates, and more l-couplers that can be plugged and unplugged between larger processors.

State transfer fidelity vs gate fidelity: how to read the numbers

Announcements about connecting quantum chips usually lead with one number. Which number it is matters a lot.

State transfer fidelity measures how faithfully a qubit’s state arrives after being sent from one processor to another. It tests the channel: does the information survive the trip?

Gate fidelity measures how closely an entangling operation, such as a CZ or CNOT between a qubit on chip A and a qubit on chip B, matches the ideal operation. Error-corrected algorithms need gates, not just transport.

A cross-chip gate is usually a composite of several steps. In the Heya et al. work, the remote CNOT combined local CZ gates with state transfers. Errors from each step add up. As a rough illustration, not a measured result: three steps at 98% each multiply to about 94%. So a link can move states very well and still produce noticeably weaker gates.

When you read a chip-to-chip fidelity claim, these questions help:

  1. Which operation is measured? State transfer, a two-qubit gate across the link, or an entangled multi-qubit state?
  2. Average or best case? A headline may quote the best link. The average across all links tells you more about a system.
  3. Raw or error-mitigated? Error mitigation is classical post-processing. It improves the reported number but does not make the hardware better.
  4. How does it compare with on-chip gates? A link is useful for computation only if cross-chip operations come close to the quality of local ones.
  5. How was it measured, and where was it published? Randomized benchmarking, tomography and other methods are not directly comparable. A preprint has not yet been peer-reviewed.
  6. Does it survive real operation? Repeatability across devices and through warm-up and cool-down cycles matters for a production machine.

The same care applies to other headline metrics. Our piece on what Quantinuum’s 99.9995% logical fidelity record measures walks through a similar gap between a striking number and the question it answers.

Where l-couplers sit on IBM’s roadmap

IBM set out its path to fault tolerance in June 2025, in a post on large-scale fault-tolerant quantum computing and at the launch of a new quantum data center (The Quantum Insider, 10 June 2025). L-couplers enter that path explicitly in 2027.

System Timing Role Status (early November 2026)
Nighthawk Announced Nov 2025 120-qubit processor with 218 tunable couplers Announced
Loon Announced Nov 2025 Experimental fault-tolerance chip: long-range c-couplers, qubit reset for qLDPC codes Announced as experimental
Kookaburra 2026 First modular processor to store and process encoded information Scheduled for 2026; no confirmed operational announcement found
Cockatoo 2027 Entangles two Kookaburra modules through “universal adapters”, parts of which use l-couplers Planned
Starling 2029 100 million operations on 200 logical qubits Planned
Blue Jay After Starling 1 billion operations on 2,000 logical qubits Planned

Cockatoo is the key dependency. Its stated job is to entangle two Kookaburra modules through what IBM calls universal adapters, which can partly be built from inter-module l-couplers. That means linking modules that hold encoded information, not just individual physical qubits. That is a harder goal than moving a single qubit’s state between two chips.

Kookaburra itself is scheduled for 2026, but as of early November 2026 we found no confirmed IBM announcement that it is running. IBM has made major hardware announcements in November before, including Nighthawk and Loon at its Quantum Developer Conference in 2025, so the next update may come soon.

Error correction also needs more than good links. Our explainer on magic state cultivation and Google’s Willow result covers another part of the fault-tolerance stack that has to work at scale.

What l-couplers do not settle yet

A working link between processors is necessary for IBM’s modular plan, but it is not enough on its own. Several questions stay open:

  • Gate quality across the link. The October 2026 distant CZ gates averaged 94.75%, after IBM’s 2025 remote CNOT reached 93.3%. On-chip gates are considerably better, so cross-chip gates need to get much closer to on-chip quality before a modular machine behaves like one large processor.
  • Scale. A 1,000-qubit system and later Starling need many links that work reliably, not a handful of good ones.
  • Encoded operations. Cockatoo requires entangling error-corrected modules. That has not been shown, and the module it depends on, Kookaburra, has not been confirmed as operational.
  • Independent scrutiny. New results usually appear first as preprints and company posts. Peer review and independent replication come later.

For cryptography, the point is timing. A fault-tolerant machine depends on many engineering steps like this one, and each can slip or arrive on schedule. That uncertainty is why standards bodies and blockchains are planning migrations now. Our post-quantum migration tracker follows where those efforts stand.

At Hadamard we build BB84, an Ethereum Layer 2 that uses hash-based one-time addresses instead of ECDSA keys. The design goal is that account security does not depend on how fast milestones like Cockatoo or Starling arrive.

FAQ

Does the l-coupler announcement mean IBM can run programs across two chips as one machine?

Not yet. IBM moved quantum states between the two chips at about 99% fidelity, but cross-chip CZ gates averaged 94.75%, and the paper reports no general computation running across both processors. A working link is a prerequisite. Running general computations across modules also requires cross-chip gates close to on-chip quality. IBM’s roadmap places entangling two encoded modules in 2027, with Cockatoo.

Why connect quantum chips with cables instead of building one bigger chip?

Fabrication equipment, chip yield and wiring all limit how large a single chip can get. Joining modules with meter-scale links inside shared cryostats avoids those limits, at the cost of adding a new source of error at the link.

How is an l-coupler different from a quantum internet?

L-couplers are superconducting microwave links that stay inside the same cold environment. They do not convert signals into photons and send them over optical fiber between separate sites, which is what quantum networking usually means.

When will IBM have a fault-tolerant quantum computer?

IBM’s target is Starling in 2029, running 100 million operations on 200 logical qubits. Cockatoo, planned for 2027, is the l-coupler-dependent step before it. Blue Jay is planned to follow with 2,000 logical qubits.

Sources