Google’s 2019 supremacy claim, seven years on: what remains?

Status as of September 24, 2026. In October 2019 Google announced that its 53-qubit Sycamore processor had done in 200 seconds what it estimated would take the Summit supercomputer 10,000 years (Nature 574, oct-2019). Seven years later, that sentence splits cleanly into a part that survived and a part that did not. This post walks the clock — date by date, number by number — and reads what is left in 2026.

What did Google measure in October 2019 — and what did it estimate?

Two different things, and the difference is the whole story. What was measured: Sycamore sampled the output of random quantum circuits on 53 qubits at 20 cycles, collecting one million samples in about 200 seconds (Arute et al., Nature 574, oct-2019). What was estimated: that reproducing those samples classically would take about 10,000 years on Oak Ridge’s Summit, then the fastest supercomputer (same paper). The measurement was a real machine doing a real (synthetic) task. The margin — the part that made «supremacy» a headline — was a claim about the best classical algorithm known to the claimant that year. Rosetta Quantum has a standing rule for exactly this split: advantage vs supremacy are different words for different evidence.

How did the classical side answer?

In escalating steps, each with a number attached. Days after the announcement, IBM argued the task would take about 2.5 days on Summit using secondary storage (Pednault et al., arXiv:1910.09534, oct-2019) — a counter-estimate on paper: the 2026 review of the episode notes it was never actually executed (LaRose, Quantum 10, 2198, sep-2026). In March 2021, Pan & Zhang produced samples that pass the XEB benchmark with tensor networks on GPUs (arXiv:2103.03074) — the method behind most reversals in this field (our tensor-networks explainer). In November 2021, a Sunway supercomputer team won the ACM Gordon Bell Prize for simulating the circuits in real time — sampling one bitstring from the 2019 experiment in 304 seconds on 41,932,800 cores (arXiv:2110.14502; prize nov-2021). In 2022, Pan, Chen & Zhang generated uncorrelated samples that pass XEB in roughly 15 hours on 512 GPUs (PRL 129, 090502). And in June 2024, Zhao et al. used 1,432 GPUs to produce 3 million uncorrelated samples in 86.4 seconds — 7× faster than the Sycamore run itself, at a higher linear cross-entropy score, and two orders of magnitude more energy-efficient than earlier simulations (arXiv:2406.18889; Natl. Sci. Rev. 12, nwae317, 2025). Their own abstract calls it «the first unambiguous experimental evidence to refute Sycamore’s claim of quantum advantage».

What survives of the 2019 claim in 2026?

The run survives; the margin does not. Nobody disputes that Sycamore sampled its circuits in 200 seconds in 2019 — that measurement stands. But the 10,000-year gap that carried the word «supremacy» is gone: a classical cluster now does the same task 7× faster, with higher fidelity, on measured hardware (arXiv:2406.18889, jun-2024), and the first peer-reviewed history of the field formally lists the 2019 experiment as refuted in its verdict table (Quantum 10, 2198, sep-2026). The claim did not die, though — it moved to bigger circuits. Google’s SYC-67 experiment (67 qubits, 32 cycles, XEB fidelity 1.5×10⁻³) estimates 12 years on Frontier under generous secondary-storage assumptions and about 6×10⁴ years from RAM alone (arXiv:2304.11119; Nature 634, oct-2024) — and the same 2026 review lists it as unrefuted to date. Willow (dec-2024) pushed to 105 qubits: an RCS benchmark in under five minutes against a 10²⁵-year Frontier estimate — with Google itself stating the estimate rests on «conservative assumptions», that «we expect classical computers to keep improving on this benchmark», and that RCS «has yet to demonstrate practical commercial applications» (Google blog, dec-2024). We found no public classical attack on the Willow estimate as of this writing — an absence in the record, not a proof.

What did seven years of this teach?

Four things, each of them checkable in the table above. First: a supremacy margin is a statement about the best classical algorithm known to the claimant on announcement day — the 2019 margin fell repeatedly without Sycamore getting one qubit worse. Second: estimates and measurements belong in different ledgers. Every number that killed the margin was measured on real hardware (304 s, ≈15 h, 86.4 s); both the 10,000 years and IBM’s 2.5 days were estimates, and one of them was never even run. Third: verification is the structural weak point. At these sizes the output distribution cannot be checked directly; XEB is a proxy, and producing samples that beat the proxy is exactly what the counterattacks did — the same pattern Rosetta documented in boson sampling and in how benchmarks get gamed. Fourth: the task is synthetic. Seven years on, no random-circuit-sampling experiment has produced a commercial application — Google says so itself (dec-2024). The 2019 episode is the founding case of the claim-vs-counterattack clock this ledger keeps.

What we don’t know

• Whether the SYC-67 and Willow estimates would survive a dedicated classical attack. The 2019 estimate looked solid for about a year and then fell to algorithms that did not exist when it was published; the 2026 review lists SYC-67 as unrefuted, which is a statement about work done, not about work possible.
• Where the true classical floor for the 2019 task is. 86.4 seconds is the best published measured time; nothing says it is final.
• The complete fidelity accounting of the 2024 classical samples beyond the published comparison («higher linear cross entropy score» than Sycamore’s, per the authors); independent replication of that run has not been published to our knowledge.
• Whether any group is currently attacking the Willow estimate in private. Absence from the public record is all we can report.
• Rosetta Quantum has not run any of these computations in house. This post is a dated reading of the public record, with every number sourced beside it.

Rosetta Quantum publishes verdicts with raw, reproducible data. This is educational content, not a product claim.