IBM Says Quantum Advantage Has Arrived. Its Three Papers Don’t Carry Equal Weight.

Editor’s note: This replaces our 10 September article, which said IBM’s quantum-advantage claim was still to come. IBM made the claim on 30 July. We missed it, and this version corrects that.

The Deadline Came Early

IBM’s roadmap promised quantum advantage by the end of 2026 and a fault-tolerant machine, Starling, with 200 logical qubits running more than 100 million operations per job, in 2029. It declared the first milestone met five months early.

On 27 and 28 July, IBM and its partners posted three preprints to arXiv, and on 30 July the company announced them together. It says quantum computers “can outperform leading classical methods” and presents the three results as the start of a “quantum advantage era.”

The three papers are not equally strong. Reading them separately is the whole job.

Three Claims, Three Strengths

IBM and the University of Chicago: the strongest. The team ran a 70-qubit circuit with 468 non-Clifford gates on IBM’s Heron processor. They wrapped it in a “spacetime code” that detects errors and discards bad runs. From roughly 3.5 million shots they kept 2,051 samples and certified a state fidelity of at least 0.284 at 95% confidence. This is the only one of the three that makes an explicit advantage claim backed by a complexity-theory argument that the problem is hard classically.

Some coverage called this a 70-logical-qubit result. IEEE Spectrum’s reading is that these are physical qubits with error detection layered on, not logical qubits encoded for error correction. That distinction is exactly the one that separates today’s machines from Starling.

Qedma: empirical. Working with RIKEN and BlueQubit, Qedma simulated a magnetic material under periodic pulses on up to 74 qubits of a Heron R3. It cross-checked the results on Quantinuum’s trapped-ion H2 and Helios machines. The classical methods it tested stopped agreeing with the quantum results at about 51 qubits. That is “beyond the classical methods we tried,” which is a weaker statement than “beyond classical computing.”

Algorithmiq: the weakest. A 56-qubit circuit on Heron, built to be hard to simulate, with a framework for establishing trust in results no classical computer can check. The analyst Marin Ivezic notes that its hardest-regime estimate still lacks a quantitative accuracy bound.

The sceptics’ reading is measured rather than dismissive. ETH Zurich’s Dominik Hangleiter told IEEE Spectrum the papers “seem to say, ‘this seems to be hard to simulate,’ which I think is good. They shouldn’t make much stronger claims than that.” Ivezic’s summary is that IBM packaged three different levels of evidence as proof that an era has begun. His narrower conclusion: quantum computing now has better tools for building and testing credible advantage candidates.

Where the Classical Side Stands

Advantage claims usually die the same way. A classical team writes a better algorithm and closes the gap. The interval between claim and counter-result is the best test of whether a claim was real.

That contest has started. A July preprint by Oh proposes a “frozen-tree” classical algorithm that reproduces the statistics sample-based checks rely on. It does not directly attack the University of Chicago doped-Clifford result, but it targets the verification style the field leans on. Expect more of these; that is how the field works, not a sign of bad faith.

What It Does Not Change

If you manage security rather than research, the answer is short. None of these experiments is a reason to revise estimates of when a quantum computer can break today’s encryption, or to change a post-quantum cryptography migration plan. Ivezic makes the point directly: the University of Chicago work uses virtual T gates, not the fault-tolerant logical operations cryptanalysis requires. Starling, and machines like it, are the relevant threat, and IBM dates Starling to 2029.

What to Watch

• Classical rebuttals to the University of Chicago result specifically. It is the claim with a hardness argument. If it survives six months of serious classical attack, the “era” framing earns some credibility.
• Logical qubits, stated as logical. The next roadmap update should report error-corrected logical qubits and their error rates, not physical counts with detection layered on.
• Whether “advantage” reaches a useful problem. All three demonstrations were chosen for hardness, not usefulness. The first advantage result on a problem someone would pay to solve is the milestone that matters commercially.

The practical reading: IBM hit its own deadline, with one strong result and two suggestive ones. Treat July as the start of a verification contest rather than the end of one. Judge each follow-up claim by three numbers: logical qubits, sustained operations, and how hard the classical side was allowed to try.