Table of Contents
August 19, 2026 – IBM has connected and cooled its first pair of modular cryogenic cells at its quantum facility in Poughkeepsie, New York. The test is an early step toward a refrigeration system that could house hundreds of interconnected quantum chips.
The joined cells reached 4 kelvin in under five days and fell below 15 millikelvin shortly afterward, according to IBM’s information sheet and announcement. Together, they stand more than eight feet tall and eight feet wide. No quantum processors were installed for this test; IBM plans to put one Nighthawk processor in each cell later in 2026.
The design uses box-shaped aluminum cells in place of the cylindrical cryostats familiar from today’s superconducting quantum computers. Each cell has its own vacuum chamber, cooling hardware and thermal shielding. Cables pass through openings between neighboring cells, with the shielding connected at each temperature stage to form a protected, cold tunnel.
The flat sides let the cells sit close together, shortening the cable runs between processors. IBM intends to connect the chips through its l-couplers: superconducting cables that carry quantum information over distances of about a meter inside the refrigerator. The company first demonstrated them in November 2024 with its two-chip Flamingo prototype.
Each module has a 2.75-cubic-meter vacuum chamber, including 0.25 cubic meters of payload space below 20 millikelvin, and 0.53 square meters available for flex wiring. IBM lists cooling capacity of 32 microwatts at 20 millikelvin and power input of 30 kilowatts. A module weighs 3,000 kilograms and can support 200 kilograms of hardware at cryogenic temperature. Compared with Quantum System Two, it provides four times the wiring area and 80 percent of the payload volume below 20 millikelvin. The announcement’s larger figure, up to twelve times the wiring space, uses System One as its reference.
Although the cells contained no processors, IBM tested them with an applied heat load to approximate operating conditions. Oliver Dial, IBM’s vice president of quantum systems, told diginomica that the team added about 30 microwatts of heat at the coldest stage to simulate the load from processors, wiring and electronics. The New Stack reported that the units operated at 23 millikelvin with that added load.
Separately, IBM reported a two-qubit operation across an l-coupler at 99.3 percent fidelity, with a target of 99.9 percent, according to EE Times, which attended the briefing. Dial said qubit coherence currently limits that performance. He also confirmed a practical limitation of the cryogenic design: replacing a cryogenic component still requires warming and cooling the connected cells as a whole. IBM is considering redundancy, including ways to leave failed components in place while the system continues operating.
IBM’s roadmap calls for at least 1,000 programmable qubits across linked processors by 2027. The longer-term goal is Starling, planned for 2029 with 200 logical qubits capable of running 100 million quantum operations. Dial said Starling would use about twelve cryogenic modules, with each cell intended to accommodate at least 2,000 physical qubits. IBM has also published a technical explanation of the architecture by Catherine Dundon, Matthew Hollister and Allie Lindler.
My Analysis
IBM published a press release and a specification sheet on August 19, and gave reporters further details at a briefing. The release emphasized up to twelve times the wiring space and a temperature below 15 millikelvin. The specification sheet put power input at 30 kilowatts per module. Coverage of the briefing reported 23 millikelvin with a simulated payload and 99.3 percent fidelity across an l-coupler. Those last three figures are absent from the press release, yet they tell me more about the system IBM is trying to build.
The maintenance disclosure is just as consequential: replacing cryogenic hardware still means warming the connected system. Having spent a fair amount of time writing about dilution refrigerators and helium-3 supply, I want to understand what these machines will take to operate. IBM has demonstrated that it can connect and cool the cells. The details outside the release show where the questions about running a complete machine begin.
Why IBM gives two wiring figures
The fourfold and twelvefold wiring figures use different baselines. Fourfold is the comparison with System Two, which uses the Bluefors KIDE platform. Twelvefold compares the new enclosure with System One, which IBM identified at the briefing as its most widely deployed design. Both comparisons are useful, provided the reference system is clear.
The specification sheet also shows that each new cell has 20 percent less payload volume below 20 millikelvin than System Two. That helps explain the design choice: IBM is gaining room for wiring while accepting a smaller volume at the lowest temperature. Given how difficult it is to bring more control and readout lines into superconducting machines, that seems a reasonable trade. The size of the outer enclosure alone tells us little about how much useful hardware it can support.
What the simulated load test showed
For a useful comparison, Bluefors specifies 30 microwatts at 20 millikelvin and 1,000 microwatts at 100 millikelvin for each KIDE cooling unit, equivalent to one XLD1000sl system. KIDE uses three such units. IBM’s 32 microwatts at 20 millikelvin is therefore close to the capacity of one commercial high-capacity dilution unit at that temperature.
That makes Dial’s account of the test particularly useful. He told 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.
The New Stack supplies the temperature for that test: the units operated at 23 millikelvin with the added 30-microwatt load. That is the figure I would use when discussing the simulated operating conditions. The below-15-millikelvin cooldown result in IBM’s announcement does not describe the result reported for this added-load test. No processors were installed in either case.
The 32-microwatt capacity specification is quoted at 20 millikelvin. Subtracting the applied load from it does not establish a remaining cooling margin: the temperatures differ, and the reports do not provide enough detail to reconcile the specification with the test conditions. They also do not establish that these are two comparable measurements on a single cooling curve.
For continuous operation, I would want to see the temperatures and stability achieved with processors, wiring and electronics installed, together with the acceptable operating range for that hardware. The heater test is useful evidence of performance under a simulated payload. It cannot yet tell us how much room a populated cell will have for upgrades or additional instrumentation.
The cooldown result also needs to be read in context. Reaching 4 kelvin in under five days is encouraging, but there were no processors installed. Bluefors does not publish a general loaded cooldown specification because the added mass varies between experiments. IBM’s result with fully installed processors, wiring and electronics will be more useful for estimating commissioning time and the downtime after a repair.
What adding modules means for the facility
IBM’s approach adds cooling capacity by adding cells. Dial described a Starling system of about twelve modules, while the information sheet envisages later systems with several tens of modules. That gives facility planners something concrete to work with, even at this early stage.
Using IBM’s figure of 30 kilowatts per module, twelve modules would require 360 kilowatts for the cryogenic equipment. Thirty would bring that to 900 kilowatts. The control electronics, classical hardware for real-time error correction, and room cooling would add to the total. These are estimates from the module specification; they are not measured loads for a complete facility.
I would also ask about helium-3. Each cell has its own dilution unit, so adding cells increases the required inventory. For context, Bluefors lists an optimal charge of 40 liters of helium-3 and 180 liters of helium-4 for an XLD1000sl. We cannot apply that figure to IBM’s design simply because the cooling capacities are similar. The circulation system and operating point affect the charge, and IBM has not disclosed its inventory or gas-handling arrangement.
That leaves the helium-3 requirement of a twelve-cell system unresolved. For a buyer, it is a procurement and capital-planning question, particularly given the concentration of the supply chain. A design that can be replicated still needs a dependable supply of everything that goes into each module.
What changed since Goldeneye
IBM has already tried building a much larger refrigerator. In September 2022, it cooled Goldeneye, a 6.7-tonne experimental cryostat with 1.7 cubic meters of experimental volume. Typical systems at the time offered 0.4–0.7 cubic meters. Goldeneye could accommodate up to six dilution units and delivered about 10 milliwatts at 100 millikelvin. IBM tested a qubit chip inside it, measured coherence times of 450 microseconds, and said the work would likely inform a future System Three.
The new architecture takes a different approach, distributing the hardware among smaller cold spaces, each with its own dilution unit. I would be cautious about reading this as a reversal. IBM described Goldeneye as a proof of concept and left the design of future systems open. Testing a very large shared cryostat was a way to learn what that approach could offer; it did not commit IBM to putting every future machine in one.
There is an interesting connection here through Matthew Hollister, a co-author of IBM’s new technical blog. He was the lead technical expert on Fermilab’s Colossus and first author of its design paper. Colossus targets 300–500 microwatts at 20 millikelvin in a mixing-chamber space two meters across, using multiple dilution circuits in one large vessel. That is roughly nine to fifteen times the cooling capacity of one IBM cell at the same temperature. Hollister’s work spans both the large shared-vessel approach and the smaller connected cells IBM is now developing.
The volume figures need care, too. IBM’s 2.75 cubic meters describes the new cell’s vacuum chamber, while Goldeneye’s 1.7 cubic meters describes its experimental volume. Comparing those numbers directly does not establish that the new cell offers more usable cold space. For equipment that needs to operate below 20 millikelvin, the figure IBM gives is 0.25 cubic meters per cell.
Connecting quantum hardware across refrigerators also has a history. Andreas Wallraff’s group at ETH Zurich linked qubits in two separate cryogenic systems in 2020, work that IBM’s Jerry Chow has credited as an inspiration. IBM is trying to extend that idea to connections between processors in enclosures designed to be manufactured, shipped and assembled into larger systems. The engineering required to make that repeatable is a substantial part of the task.
How it compares with Bluefors’ modular platform
Bluefors introduced its own Modular Cryogenic Platform on March 3, 2026. Its self-supporting vacuum chambers connect into a shared payload space, with separate cooling and wiring inserts so either can be upgraded independently. Bluefors specifies 800 kilograms of payload per module, up to 32 side-loading ports, and a low-height design intended for data center rooms. The first multi-module delivery is planned for late 2026. Like IBM, Bluefors describes an eventual system supporting hundreds of thousands of physical qubits.
IBM addresses the obvious question in its blog: why develop its own architecture when modular cryogenic platforms already exist? Its answer is that the cells are tailored to IBM’s processors, readout wiring and cryoelectronics. That is a reasonable basis for deciding to build a system internally, although the advantages will have to show up in operation.
One difference is how the cold spaces connect. Bluefors creates a continuous payload volume. IBM connects separate cold volumes through shielded tunnels designed to limit thermal interaction. IBM says this will allow cooldown times and temperature stability to remain consistent as more cells are added.
I initially assumed that this separation would also let IBM warm one cell for maintenance while its neighbors stayed cold. Dial’s comments to EE Times make clear that it does not: replacing a cryogenic component still requires warming and cooling the connected system. IBM is studying redundancy and the possibility of leaving failed components in place.
For an operator, that is a significant limitation. Separate boxes suggest a degree of serviceability that the current design does not yet provide. How often components fail, which failures the system can tolerate, and how long repairs take will have a direct bearing on availability and operating cost. Those details deserve as much attention as the ability to add another cell.
A modular system built around IBM
I have argued for a Quantum Open Architecture in which customers could combine qubit modules, control systems and cryogenic platforms from different vendors. IBM’s design is built around its own hardware: IBM cells connecting IBM processors through IBM l-couplers. It should make IBM’s systems easier to manufacture and expand, but customers should not infer that the components will be interchangeable with another vendor’s.
That does not mean IBM is making every component itself. Its blog describes established dilution refrigerator technology, and coverage of the briefing says IBM is using Bluefors cooling technology while concentrating on the surrounding system. IBM’s contribution is the enclosure, shielding, wiring harness and integration architecture. There is considerable manufacturing and systems-engineering work in that, even with the underlying cooling technology supplied externally.
The cryostat also fits a broader pattern in IBM’s recent investments. On May 21, IBM and the Department of Commerce announced a letter of intent for Anderon, a planned 300-millimeter pure-play quantum wafer foundry in Albany. The proposal included a billion-dollar CHIPS award and a matching billion dollars from IBM. The cryogenic announcement followed on August 19.
On August 26, IBM completed its acquisition of HRL Laboratories from Boeing and General Motors, adding silicon-spin qubit technology as well as expertise in cryogenics, control electronics, packaging and interconnects. I had written about HRL’s cryo-CMOS silicon processor three weeks before the deal closed.
Taken together, these moves suggest IBM wants greater control over the infrastructure needed to turn its quantum processors into complete machines. That is familiar territory for a company with its mainframe history. It also explains why IBM might prefer to develop a cryogenic architecture around its own requirements, even as specialist suppliers offer more general-purpose platforms.
What I want to see next
The August test gives IBM an operating pair of cryogenic cells in which to begin testing processors. Whether those processors can work together reliably depends heavily on the l-coupler. The figures reported at the briefing suggest progress, although they leave plenty to establish.
In November 2024, IBM reported its best benchmarked cross-chip CNOT gates on test devices at 3.5 percent error for a 235-nanosecond operation. That was roughly forty times Heron’s best on-chip two-qubit error of 8 × 10⁻⁴, a comparison Du and colleagues discuss. The 99.3 percent fidelity reported in August corresponds to 0.7 percent error. If the measurements are comparable, that is about a fivefold reduction in error in under two years. IBM’s stated target is 99.9 percent fidelity, or 0.1 percent error.
A best result tells us what a link can achieve under particular conditions. To assess an error-corrected machine, I would want to see how performance varies across many links and over long runs, including whether using a link disturbs the processors at either end. Dial said the attached qubits’ coherence is the main limitation today, so further improvement will depend on the processors as well as the interconnect. The planned Nighthawk tests will examine state transfer and whether activating the connection degrades either chip’s performance.
The next scheduled step is to install a Nighthawk processor in each of the two cells before the end of 2026. For 2027, IBM plans Cockatoo, intended to entangle two encoded modules, alongside its separate goal of at least 1,000 programmable qubits. As those tests proceed, I would like IBM to publish reproducible inter-module benchmarks with enough detail to judge typical performance. That would make it much easier to assess the remaining work than another best-case fidelity figure.
The qubit counts also need to stay clearly labeled. Nighthawk r2, released on August 31, has 120 programmable qubits, 218 couplers and 120 reset elements, making 458 physical quantum elements in total. Those are different measures from the 200 logical qubits planned for Starling. A refrigerator’s capacity to hold quantum hardware does not, by itself, tell us how many qubits will be available for computation or how well they will perform.
IBM deserves credit for building and operating the connected cells, and for publishing enough detail to examine the design. The specification sheet and briefing coverage make clear how much remains to be tested in a complete machine. The next Nighthawk installation should show how the cells behave with processors installed and how reliably the links work between them. For anyone planning a facility around this architecture, those results and a workable maintenance plan will be more useful than enclosure size or the lowest cooldown temperature alone.
