Table of Contents
August 19, 2026 – IBM said it has joined and cooled two cryogenic modules into a single operating environment for the first time, a step in the company’s plan to house hundreds of quantum chips in one connected refrigeration system.
The two coupled cells, assembled at IBM’s quantum facility in Poughkeepsie, New York, cooled to 4 kelvin in under five days and reached below 15 millikelvin shortly afterward, according to an information sheet IBM published on Zenodo alongside the announcement. The joined pair measures more than eight feet in both height and width. No quantum processors were installed for the test.
The architecture replaces the cylindrical cryostats that house today’s superconducting quantum computers with box-shaped aluminum cells built from solid panels and framing. Each cell has its own vacuum chamber, cooling hardware and thermal shielding. Technicians route quantum cables through an opening in the wall of one cell into the neighboring chamber, then connect the shield layers stage by stage. Those shields form a cold, electromagnetically sealed tunnel between the processors.
Flat faces on each cell allow two cells to touch along their full height, which shortens the cable run between chips in adjacent cells. IBM plans to make those connections with its l-couplers, meter-scale superconducting cables that operate inside the refrigerator and carry quantum information between separate chips. IBM first demonstrated l-couplers in November 2024 with a two-chip prototype called Flamingo.
IBM’s information sheet gives each module a vacuum chamber of 2.75 cubic meters, a payload space of 0.25 cubic meters below 20 millikelvin, and 0.53 square meters of area available for flex wiring. The sheet lists cooling capacity of 32 microwatts at 20 millikelvin, a power input of 30 kilowatts, a supported hardware mass of 200 kilograms at cryogenic temperature, and a total module mass of 3,000 kilograms. It puts the wiring area at four times that of IBM Quantum System Two per module, and the volume below 20 millikelvin at 0.8 times System Two per module. The press release describes the vacuum enclosure as offering up to 12 times more wiring space than IBM’s most widely used quantum systems, which the company identified as System One.
To approximate operating conditions, the team applied heat to the coldest stage during the test. “We put about 30 microwatts of heat onto that mixing chamber plate to simulate what it will be like when this refrigerator is operating with a set of installed quantum processors and wiring and electronics and all the heat that that generates,” Oliver Dial, IBM’s vice president of quantum systems and an IBM Fellow, told diginomica.
“Bringing fault-tolerant quantum computers to industries depends on several fundamental advances,” said Jay Gambetta, Director of IBM Research and IBM Fellow. Gambetta said the connection and operation of the modules would accelerate the company’s progress alongside work on hardware, software and algorithms.
IBM said it has demonstrated a two-qubit operation across an l-coupler at 99.3 percent fidelity and is targeting 99.9 percent, according to EE Times, which attended a pre-announcement briefing. Dial told the publication that qubit coherence is the main limit on current l-coupler performance, and that replacing any cryogenic component in the new system would still require warming and cooling the linked cells as a whole. IBM is studying operational redundancy, including a possible fail-in-place approach.
IBM said it will install one Nighthawk processor in each cryogenic module later this year to begin operational testing. Under its roadmap, the company plans to use l-couplers to link multiple processors into a machine with at least 1,000 programmable qubits by 2027, which IBM defines as qubits available directly for computation. Each cell will hold at least 2,000 physical qubits, and Starling will connect about twelve modules, Dial said.
The work supports IBM Quantum Starling, the fault-tolerant machine the company has committed to deliver in 2029 with 200 logical qubits running 100 million quantum operations. IBM published the accompanying technical explanation on its quantum blog, authored by Catherine Dundon, Matthew Hollister and Allie Lindler. IBM released the information sheet under a Creative Commons Attribution licence with a DOI.
My Analysis
IBM published two documents on 19 August, and briefed reporters on a third set of numbers that appeared in neither. The press release led with a wiring figure of up to 12 times. The information sheet on Zenodo gave four times against System Two and added the cold-stage specifications: 32 microwatts of cooling at 20 millikelvin, 0.25 cubic meters of payload volume below 20 millikelvin, and 30 kilowatts of power input, all per module. The briefing produced the rest, including the number that determines whether any of this becomes a computer.
That number is 99.3 percent.
I have spent a fair amount of time on dilution refrigerators and the helium-3 supply that runs them, so I read the spec sheet before the release. Reading both, and then the trade coverage, the picture is of a company that disclosed a great deal and distributed it so that no single source supports the argument. Three constraints show up across the three channels: the cooling margin, the service model, and the interconnect. None of them appears in the press release.
The two wiring numbers
Both figures are true and measure different things. Four times is against System Two, which runs on the Bluefors KIDE platform and is IBM’s newest deployment. Twelve times is against System One, IBM’s first commercial machine and still the most widely deployed design it sells, which the company confirmed at the briefing.
The same sheet reports that each module holds 0.8 times System Two’s volume below 20 millikelvin. Each cell is physically larger than a KIDE and colder inside a smaller fraction of itself. The design trades cold volume for wiring area, which is the correct trade if you accept that the wiring wall, not the chip, is what stops superconducting machines from growing.
32 microwatts, and the 30 IBM put back
Bluefors publishes the number that makes the cooling comparable. Each KIDE cooling unit delivers 30 microwatts at 20 millikelvin and 1,000 microwatts at 100 millikelvin, which the company describes as the same cooling power as one XLD1000sl system. KIDE uses three of those units. IBM’s cell delivers 32 microwatts at 20 millikelvin, which matches one commercial high-capacity dilution insert at the coldest stage.
Then Oliver Dial, IBM’s VP of Quantum Systems, described the test itself to diginomica:
We put about 30 microwatts of heat onto that mixing chamber plate to simulate what it will be like when this refrigerator is operating with a set of installed quantum processors and wiring and electronics.
Thirty microwatts of applied load against 32 microwatts of published capacity.
IBM ran the demonstration at roughly 94 percent of the module’s cold-stage budget and called it true operating conditions, which is a more informative disclosure than the empty-box cooldown everyone reported.
It also runs against the rule I give integrators in my own procurement guidance: if the heat load at any stage exceeds 80 percent of available cooling power, there is no margin for a processor upgrade or added instrumentation. IBM chose the load figure to represent a fully populated cell, which makes it a design point and not a measured payload. Two microwatts of headroom at the coldest stage is the question I would put to IBM first, because continuous operation over years is where thermal margin disappears.
The four-kelvin cooldown deserves separate care. Under five days is a good figure and IBM measured it with no processors inside. Bluefors declines to publish a loaded cooldown specification at all, on the stated grounds that added mass varies too much between experiments. Whether a fully wired cell still reaches 4 K in five days determines whether commissioning a Starling-class machine takes a week or a season, and nobody outside Poughkeepsie can answer that yet.
Scaling by replication
Cooling capacity per module is fixed, so growth comes from adding modules. Dial told diginomica that Starling will connect about twelve of them, and IBM’s information sheet describes later systems reaching several tens of modules.
At the published 30 kilowatts per module, twelve modules draw 360 kilowatts for the cryogenic plant, before control electronics, before the classical decoding hardware that real-time error correction requires, before cooling the room. A later thirty-module machine implies 900 kilowatts on the same basis. Those are planning figures derived from IBM’s component specification, not a measured facility load, and the arithmetic is mine.
Helium-3 is the input I would ask about next. Each cell has its own dilution unit, so inventory grows with every module added. Bluefors publishes an optimal charge of 40 litres of helium-3 and 180 litres of helium-4 for an XLD1000sl. IBM has disclosed neither its charge nor its gas-handling topology, and comparable cooling power does not establish comparable isotope inventory, since circulation architecture and operating point both move the number. Until IBM publishes it, the helium-3 requirement of a twelve-cell machine cannot be calculated from public specifications, and I am not going to pretend otherwise. It remains a procurement and working-capital question for anyone assembling machines from many dilution units, and the supply chain concentration behind it is among the least-discussed risks in superconducting quantum computing.
Goldeneye explored the other end of the design space
In September 2022, IBM cooled Goldeneye, a 6.7-tonne experimental cryostat with 1.7 cubic meters of experimental volume, against 0.4–0.7 cubic meters for the fridges of the day. It held up to six dilution units and delivered about 10 milliwatts at 100 millikelvin. IBM put a qubit chip inside, measured 450-microsecond coherence times, and said the lessons would likely feed a future System Three.
The 2026 architecture points the other way.
One dilution unit per box, many boxes, each colder inside a smaller volume than the platform IBM ships today. IBM described Goldeneye as a proof of concept and said at the time that the shape of future systems remained open. The company explored one end of the design space and then built at the other.
One detail makes the arc concrete. Matthew Hollister, second author on IBM’s blog post, was the lead technical expert on Colossus at Fermilab and first author on its design paper. Colossus targets 300–500 microwatts at 20 millikelvin in a mixing-chamber volume two meters across, roughly nine to fifteen times IBM’s per-cell figure, delivered by multiple dilution circuits inside one continuous room-scale vessel. The engineer who specified that platform now works on the case for many small ones.
A caution on volumes. IBM’s 2.75 cubic meters is a vacuum chamber and Goldeneye’s 1.7 is an experimental volume. At least one analysis has set those two figures side by side to conclude that each new cell exceeds Goldeneye. They are different quantities, and the like-for-like number, payload below 20 millikelvin, is 0.25 cubic meters per cell.
IBM is also not the first to move a quantum signal between separate refrigerators. Andreas Wallraff’s group at ETH Zurich linked qubits in two spatially separated cryogenic systems in 2020, and Jerry Chow credits that work as inspiration. IBM claims scope: chip to chip instead of qubit to qubit, inside enclosures designed from the start to be manufactured, shipped and joined.
Bluefors announced a different modular platform in March
On 3 March 2026, Bluefors introduced its Modular Cryogenic Platform: self-supporting expandable vacuum chambers, modules that connect into a unified payload space, cooling inserts separated from wiring inserts so either can be upgraded alone, 800 kilograms of payload per module, up to 32 side-loading ports, and a low-height form factor built for data center rooms. First multi-module delivery is slated for late 2026. Bluefors says the platform will support hundreds of thousands of physical qubits, the same phrase IBM used five months later.
IBM’s blog includes a section headed “Why isn’t IBM using existing modular approaches for cryogenics?” The answer given is that IBM’s design is built specifically for IBM quantum computers, tailored to its readout wiring, processor designs and cryoelectronics. Read that as what it is, a make-or-buy decision, explained.
The two architectures make different trades. Bluefors joins modules into one continuous cold volume. IBM keeps separate cold volumes and connects them through shielded tunnels engineered to minimize thermal interaction, which is why IBM can claim that cooldown time and stability stay flat as cells are added.
That is thermal decoupling during operation, and it is not maintenance isolation. I assumed at first that separate cold volumes would let one cell warm without its neighbors, which is the argument the geometry invites. Dial told EE Times the opposite: replacing anything cryogenic still requires warming and cooling the linked system as a whole, and IBM is studying redundancy and a possible fail-in-place strategy. The boxes look like rack units and do not yet behave like them. For a machine expected to run for years, that service model decides operating cost more directly than the vacuum geometry does.
Modular without being open
I have argued that the most valuable change available to the field would be a Quantum Open Architecture, where qubit modules, control stacks and cryogenic platforms come from different vendors and interoperate. IBM has built the most modular cryostat in the industry and made it the least interchangeable. Cells connect to IBM cells, house IBM processors, and route IBM l-couplers. Modularity here is a manufacturing property, not an interface property.
The in-house part needs qualifying. IBM’s blog says the cells use established dilution refrigerator technology, and reporting from the briefing describes IBM leveraging Bluefors cooling technology so it can concentrate on the system level. What IBM designed is the enclosure, the shielding, the wiring harness and the integration architecture around a cooling technology it still sources. That is a narrower claim than displacing its supplier, and it supports the engineering scale and manufacturability argument just as well.
IBM made three moves this year and the cryostat is the second. On 21 May, IBM and the Department of Commerce announced a letter of intent for Anderon, a planned 300-millimeter pure-play quantum wafer foundry in Albany, supported by a proposed billion-dollar CHIPS award and a matching billion in IBM cash. On 19 August, the cryostat. On 26 August, IBM completed its acquisition of HRL Laboratories from Boeing and General Motors, buying silicon-spin qubits along with cryogenics, control electronics, packaging and interconnect capability. I wrote about HRL’s cryo-CMOS silicon processor three weeks before that deal closed.
Wafers, packaging, control electronics, cryogenic systems engineering. In three months IBM moved to control or internalize most of the stack below the qubit, which is the mainframe playbook applied to quantum by the company that wrote it.
What has to happen next
The August milestone validated vacuum, thermal, shielding and mechanical engineering. The computational question turns on the l-coupler, and here IBM has moved further than the last published figure suggested.
When IBM demonstrated l-couplers with Flamingo in November 2024, it reported its best benchmarked cross-chip CNOT gates at 3.5 percent error for a 235-nanosecond operation, on test devices. Against Heron’s best on-chip two-qubit error of 8 × 10⁻⁴, that ran roughly forty times worse, a comparison Du and colleagues make from the same two IBM figures. At the August briefing IBM said its best two-qubit operation across an l-coupler now reaches 99.3 percent fidelity, against a target of 99.9 percent. Error has fallen roughly fivefold in under two years.
What IBM has disclosed is a best result. What error correction consumes is a distribution, across many couplers, over long runs, in the presence of whatever the link does to the processors on either end. Dial said the current limit is the coherence of the qubits attached to the coupler, which points the next improvement back at the processor and not at the cable. He also said the planned Nighthawk demonstration will test state transfer and whether activating the link degrades either chip. Those are the right tests, and none of the results exist yet.
Three checkpoints follow, all inside eighteen months. IBM installs one Nighthawk processor in each joined cell before the end of 2026, which will be the first operation with live qubits. A reproducible inter-module benchmark rather than a best-case number, which is the disclosure I want and the one most likely to stay internal. And Cockatoo in 2027, tasked with entangling two encoded modules, alongside the separate target of at least 1,000 programmable qubits.
Counting those qubits requires care, and IBM’s newest processor shows why. Nighthawk r2, released on 31 August, carries 120 programmable qubits supported by 218 couplers and 120 reset elements, for 458 physical quantum elements in total. Programmable qubits, physical elements and the 200 logical qubits promised for Starling are three different counts, and press coverage collapses them constantly.
IBM has been hitting its published dates, and the cadence continued through the summer. Nothing in the cryogenic announcement contradicts that record. What it does is move the interesting uncertainty from mechanical engineering, which IBM has now shown it can do, to three constraints the company disclosed in three different places: two microwatts of thermal headroom, a service model that still warms the whole machine, and an interconnect running 99.3 percent against a 99.9 percent requirement.
The engineering sheet told me more than the press release. The briefing told me more than either. The next document that matters is a benchmark table.
