IBM's 100-Qubit Chip Runs at 1.5 Kelvin—The Fridge Is the Catch

IBM shipped the chip. The fridge did not ship with it.

IBM’s Quantum team posted a hardware note this week for the Heron R3, a 100-qubit superconducting processor that the company says operates at a base temperature of 1.5 Kelvin. That number matters because the previous Heron generation ran in the 15–20 millikelvin range inside a dilution refrigerator, and 1.5 K is roughly 100 times warmer. If the spec holds, it changes which cooling plants can host the chip.

But IBM did not announce a new fridge. It announced a chip that tolerates a warmer one. That gap—between a device spec and a deployable system—is the whole story.

What 1.5 Kelvin actually buys you

Superconducting qubits are fragile. Thermal noise at higher temperatures kicks Cooper pairs apart and shortens coherence. The standard industry answer has been to go colder: dilution refrigerators from Bluefors and Oxford Instruments reach base temperatures around 10 mK and cost somewhere between $500,000 and $1.5 million depending on wiring and shielding.

A 1.5 K operating point is not cryogenic-free. It is still colder than liquid helium’s 4.2 K boiling point, and it still requires a cryostat. But it sits within reach of pumped helium-3 systems and some closed-cycle cryocoolers that are cheaper, smaller, and easier to service than a full dilution unit. IBM’s framing is that Heron R3 opens the door to “compact cryogenics”—the company has not published a bill of materials.

What IBM did publish: a 100-qubit count, a claimed two-qubit gate fidelity above 99.5%, and a coherence time the note describes as “materially extended” without giving a figure. That last omission is the one builders should watch. Fidelity and coherence are the numbers that decide whether a chip is useful, and “materially extended” is not a number.

Why a warmer chip is an engineering claim, not a marketing one

The mechanism behind the 1.5 K figure is worth spelling out, because it is easy to confuse with the marketing around it. A dilution refrigerator works by mixing helium-3 and helium-4 isotopes to pull heat out of the mixing chamber, and the whole apparatus—pumps, stills, heat exchangers, gas handling—is built around reaching and holding that millikelvin floor. A chip that tolerates 1.5 K does not eliminate any of that machinery by itself. What it does is relax the tolerance the machinery has to meet.

That relaxation compounds. Every stage of a dilution unit has to reject heat from the stage above it, and the coldest stage has the least cooling power to spare. Moving the operating point up by two orders of magnitude gives the system more headroom at every interface: the wiring, the attenuators, the filters, the connectors. A builder who has ever watched a fridge lose its base temperature after a rewiring job knows how little margin exists at 10 mK. At 1.5 K, that margin is larger, and the failure modes are less catastrophic.

The caveat is that none of this shows up in a qubit count. A chip that runs warmer but decoheres faster is not obviously better than a colder chip that holds its state. IBM’s 99.5% two-qubit gate fidelity claim is the number that matters most here, because gate fidelity is what determines whether error correction can work at all. But it is a claim, not a benchmark, and it is stated without the coherence figure that would let an outsider check the trade.

The wiring problem is older than the chip

Every superconducting qubit needs control lines. A 100-qubit processor needs hundreds of coaxial cables running from room temperature down to the mixing chamber, each one a thermal leak and each one a source of crosstalk. IBM’s own roadmaps have flagged this as the limiting factor for scaling past a few hundred qubits.

The warmer operating point helps here in a way that is easy to miss. Less aggressive cooling means less thermal margin consumed by each cable, so the same fridge can carry more lines before it runs out of cooling power. That is an engineering argument, not a benchmark, and IBM has not quantified it.

Competitors are attacking the same problem from the other direction. Intel has been pushing cryogenic CMOS control electronics that sit inside the fridge next to the qubits, cutting the number of wires that have to cross the temperature boundary. If that works, the advantage of a warmer chip shrinks. The two approaches are not mutually exclusive—a warm chip and in-fridge control could be combined—but each one, on its own, reduces the urgency of the other.

There is a second-order effect too. Cables are not just thermal loads; they are antennas. Every line that crosses the temperature boundary is a path for noise, and the filtering that suppresses that noise is itself a heat source. A design that needs fewer or shorter runs is a design with fewer places for a stray photon to land. IBM has not said whether Heron R3 changes the wiring count, only that it tolerates a warmer environment.

Where Heron R3 sits against the field

IBM’s own published roadmap puts Heron in the “utility-scale” tier, with the company targeting 200 qubits per chip and 2,000 qubits across multiple chips by the end of its stated window. Heron R3 at 100 qubits is a step along that path, not a leap off it.

Google’s Willow chip, announced in late 2024, demonstrated below-threshold quantum error correction on 105 qubits—a different claim, measured differently, on a different architecture. Comparing raw qubit counts across vendors is close to meaningless. The useful comparison is error rate per gate at a given operating temperature, and neither IBM nor Google publishes that in a directly comparable form.

IonQ and Quantinuum use trapped ions, which operate at room temperature in some configurations but trade that for much slower gate times. The superconducting camp—IBM, Google, Rigetti—is betting that speed and fabrication maturity outweigh the cryogenics tax. Heron R3 is a bet that the tax can be reduced.

What makes the comparison hard is that the vendors are optimizing different parts of the same stack. Google’s headline is an error-correction result, which is a statement about what a logical qubit can do. IBM’s headline here is an operating condition, which is a statement about what a physical qubit costs to run. Both are real, and neither answers the other’s question. A reader trying to decide which platform to build on will not find the answer in either press release.

The price nobody has quoted

IBM sells quantum access through its cloud service, not as boxes. The company’s pay-as-you-go plans have historically been priced per second of circuit execution, with a free tier capped at a small number of minutes per month. For Heron R3, IBM has not published a new rate card.

That silence is telling. If the warmer chip really does reduce the cost of the surrounding plant, IBM has an incentive to say so loudly. It hasn’t. The likely reason is that the cost of a dilution refrigerator is amortized across many users on a cloud service, so a cheaper fridge barely moves the per-second price a developer sees.

For a builder, the practical question is not 1.5 K versus 10 mK. It is whether Heron R3 shows up in the cloud at a price that makes running a 100-qubit circuit cheaper than it was last quarter. IBM has not answered that yet.

What would make this a real shift

Three things would turn Heron R3 from a spec sheet into a platform change. First, a published coherence number in microseconds, not an adjective. Second, a named cryostat partner and a stated base-temperature tolerance, so buyers know what they are installing. Third, a cloud price for Heron R3 executions that undercuts the previous generation.

IBM has delivered one of the three. The company has a history of shipping hardware notes ahead of the systems that make them useful—its 2023 Condor chip hit 1,121 qubits while the surrounding infrastructure lagged. Heron R3 looks like the same pattern at a smaller, more honest scale.

The chip is real. The fridge is the question.

Frequently asked questions

Is 1.5 Kelvin warm enough to skip a dilution refrigerator? No. It is still below liquid helium’s 4.2 K boiling point, so a cryostat is required. It is warm enough to consider cheaper pumped helium-3 or closed-cycle systems instead of a full dilution unit.

How does Heron R3 compare to Google’s Willow? They measure different things. Willow’s headline result was below-threshold error correction on 105 qubits. Heron R3’s headline is an operating temperature and a qubit count. Raw qubit counts are not comparable across architectures.

Can I use Heron R3 today? Only through IBM Quantum’s cloud service, and IBM has not published a new rate card for it. Check the IBM Quantum platform for current access tiers before planning a workload.

Does a warmer chip mean cheaper quantum computing for me? Not automatically. The savings land in the capital cost of the refrigerator, which a cloud provider absorbs and amortizes across users. Whether that reaches your per-second execution price depends on a rate card IBM has not published.

What should I watch next? A coherence figure in microseconds, a named cryostat partner with a stated base-temperature tolerance, and a Heron R3 rate card. Any one of those would move this from a hardware note to a deployable system.