IBM says three demonstrations on its Heron quantum processor prove quantum advantage has been reached, and one of them ran in seven hours on commercially available hardware where the Fugaku supercomputer burned 700,000 CPU hours without reproducing the result. The claim, detailed this week in a series of research papers with partners including Algorithmiq, Qedma and the University of Chicago, is the strongest assertion yet from Big Blue. It is also a claim the scientific community has challenged before, every time a vendor has made it.

Jay Gambetta, IBM Fellow and director of IBM Research, told journalists on a conference call: "These demonstrations prove that we can scale quantum computing forward with confidence." He said the work shows quantum computers solving problems beyond the reach of the largest classical systems, with results that can be rigorously validated. Every prior advantage claim has faced exactly that validation problem. Gambetta himself flagged verification as the key obstacle in random circuit sampling benchmarks.

What is actually being claimed?

Quantum advantage, as IBM defines it with French startup Pasqal, requires four things: the quantum system must work more accurately, more cheaply and more efficiently than a classical supercomputer, and outsiders must be able to validate the output. IBM runs a platform-agnostic Quantum Advantage Tracker to collect and test such claims. That framework is a year old. The three demonstrations are its first showcase.

The distinction matters. Google claimed quantum supremacy in 2019 with its 53-qubit Sycamore processor. D-Wave touted its Advantage 2 annealing system last year. Both claims were disputed, and supremacy only requires beating a classical machine at any task, useful or not. Advantage demands a real-world problem. Q-CTRL said in May it had "achieved evidence of practical quantum advantage" in material science, and even that hedged wording drew debate.

Eight months of challenge, no classical match

Algorithmiq, a startup building quantum software for life sciences and chemistry, submitted its demonstration to the tracker eight months ago. It simulated what co-founder and CTO Sabrina Maniscalco called "real-world disordered matter" — materials such as catalysts and battery electrolytes whose irregular structures resist idealized modeling.

"Most of the physics that we are taught in school describe really something that is rather idealized: idealized materials, perfect crystals, clean symmetric structures," Maniscalco said. "But the real world isn't like that."

In eight months, she said, no classical system has reliably reproduced the results at scale. IBM reran the experiment on quantum processors with different noise profiles and calibrations; Algorithmiq brought in classical simulation firms including the Flatiron Institute to attack its own findings. Two IBM chips produced consistent results within the margin of error every time. "This consistency is not coincidence, it's evidence," Maniscalco said. Roughly half of the classical simulation methods could not match the quantum results at all. "They didn't even agree with each other," she said.

Seven hours versus 700,000 CPU hours

Qedma Quantum Computing ran the most concrete comparison. Its demonstration studied materials driven by external laser fields, a problem relevant to optical electronics. The researchers first ran simulations on Fugaku, the supercomputer powered by 48-core AMD CPUs and Nvidia H100 GPUs, on materials of 28 and 35 electrons using finite-size scaling. Real materials contain billions of electrons.

The two classical simulation types, tracking time and magnetization, initially agreed, then diverged. "Actually, the H100 server runs out of memory and starts getting really bad results," said Netanel Linder, Qedma's co-founder and CTO. "You cannot see these oscillations."

Qedma then ran the simulations on a 51-qubit IBM system, with and without its Quantum Error Suppression and Error Mitigation (QESEM) software, and saw the oscillations. To validate, it reran the tests on Quantinuum's H2 and Helios trapped-ion systems — a different modality from IBM's superconducting hardware — and got matching results. It then scaled to two 74-qubit systems representing materials with billions of electrons.

The numbers are stark. Fugaku ran for 700,000 CPU hours and could not reproduce the oscillations. IBM's Heron r3 "Boston" systems ran for seven hours. Pushing Fugaku further, Linder said, would take 80,000 GPU servers running for half a year. "Clearly, this is not possible," he said, adding: "In the end, this is the first quantum advantage experiment that was done on commercially available hardware using commercially available software."

Where is the catch?

The University of Chicago work addresses the trust question directly. Researchers built a structured alternative to random circuit sampling using spacetime codes, running about 70 logical qubits with 21 insular qubits and up to 2,415 logical two-qubit gates, plus 468 logical T gates. Logical error rates came in ten times lower than physical error rates. The demonstration ran in about 16 minutes. Verification held as T gates were added up to 75, with no change in error-correction properties.

That is a verified run on 70 logical qubits, not a commercial deployment. The demonstrations cover material science and benchmark problems, not business workloads, and the eight-month challenge window is short by scientific standards. Still, if the claims survive the same scrutiny that broke Sycamore's supremacy claim and D-Wave's annealing assertions, IBM has moved the argument from whether quantum machines can win to whether anyone can catch them.