← Back to blog

Below Threshold: What Google's Willow Actually Proved, and the Septillion-Year Number It Didn't

Willow ran a benchmark in under five minutes that a supercomputer would need ten septillion years to match — a task its own makers call useless. Buried under that number was a genuinely historic result: quantum error correction finally crossed a threshold it had chased for thirty years. Here is how to tell the milestone from the marketing.

Published July 2026 · 8 min read

In December 2024, Google put a number into the world that was impossible to ignore. Its new quantum chip, Willow, had run a benchmark in under five minutes that would take the fastest supercomputer on Earth about ten septillion years. That is a 10 followed by 25 zeros, a span so far beyond the age of the universe that the comparison stops meaning anything. The number went everywhere, usually attached to the phrase "quantum is here."

Here is the detail that almost none of the coverage carried. The benchmark that produced the septillion-year figure has, in Google's own words, "no known real-world applications." Google's own executives described it as "not useful for an application" and "just the entry point." The headline number measures a task chosen precisely because it is hard for classical computers and useless for everyone else.

That would make this a simple story about hype, except for one inconvenient fact: buried under the septillion-year headline was a genuinely historic result, and it was a completely different experiment. Willow cleared a bar that quantum computing had been chasing for thirty years. The trouble is that the real milestone and the marketing number got fused into one, and the fusion is where the physics quietly left the room.

This piece separates the two. It is worth doing carefully, because the real achievement deserves the credit, and the inflated one deserves the scrutiny, and telling them apart is the whole skill of reading a quantum headline.

The real result: the curve finally bent the right way

Start with what Willow actually proved, because it is the part worth being excited about, and undervaluing it would be its own kind of dishonesty.

Quantum bits are fragile. A qubit holds its state for a flicker before noise corrupts it, and no amount of careful engineering makes a physical qubit reliable enough to run a long computation. The field's answer, worked out in theory decades ago, is error correction: spread the information for one good "logical" qubit across many noisy physical qubits, and use the redundancy to detect and fix errors faster than they accumulate.

There has always been a catch, and it is brutal. Error correction only helps if your physical qubits are already good enough. Below a certain error rate, adding more physical qubits to a code makes the logical qubit better. Above that rate, adding more qubits makes it worse, because each new part brings more failure than the redundancy can clean up. That dividing line is called the threshold, and for thirty years real hardware sat on the wrong side of it. You could build bigger codes, and they would rot faster. The theory said scaling was possible. The machines said not yet.

Willow crossed the line. Google built a series of surface-code memories at increasing sizes, code distance 3, then 5, then 7, and measured what happened to the error rate as the code grew. The result is a single number that carries the entire achievement: Λ = 2.14. Each time they stepped the code distance up by two, the logical error rate fell by a factor of about 2.14. The curve bent the right way. On a real device at this scale, a bigger quantum memory was now a more reliable one, the below-threshold crossing that thirty years of hardware had never reached.

The best of those memories, the distance-7 code, reached a logical error rate of 0.143% per cycle of error correction. It also passed a second landmark that matters more than it sounds: it lived longer than its own best physical qubit, by a factor of 2.4. Before this, the abstraction never paid for itself. You assembled 101 fragile qubits into one logical qubit and got something flimsier than the single best qubit you started with. Willow's logical qubit outlasted its parts. The overhead finally bought something. Google's team also showed the error correction running in real time, with a decoder fast enough to keep up with the hardware as it ran, which is a separate practical hurdle that had to be cleared and was.

That is a real milestone, and a hard-won one. The honest version of the Willow story leads with it. Scalable fault tolerance requires being below threshold, and until December 2024 nobody had demonstrated it on a real machine at this size. Google did.

The other number is a different experiment

Now the septillion years.

The ten-septillion-year figure did not come from the error-correction work. It came from a separate experiment called Random Circuit Sampling, run on the same chip but measuring something entirely unrelated. In Random Circuit Sampling you apply a long sequence of random operations to the qubits and then sample the output distribution. The task is engineered to be maximally awkward for a classical computer to simulate, and it succeeds at that. Willow did it in minutes. Estimating how long the Frontier supercomputer would need to reproduce the same sampling gives the headline number.

The two results share a chip and nothing else. One is a memory that stores a qubit well. The other is a sampling stunt with no output anyone can use. Random Circuit Sampling does not run an algorithm, does not solve a problem, and produces no answer that matters outside the benchmark itself. That is not a criticism smuggled in from a skeptic. It is Google's own description: no known real-world applications.

The sleight of hand, and it was mostly a sleight of the coverage rather than an outright claim, was to let the septillion-year number stand in for the achievement. A reader comes away believing that Willow performed some astronomically valuable computation ten septillion years faster than a supercomputer. It did not. It performed a deliberately useless one quickly, and separately, it stored a single qubit reliably. Fuse the two and you get "quantum is here." Keep them apart and you get the truth, which is more interesting: quantum error correction crossed a threshold it had never crossed, and the flashy speed number is a well-understood benchmark that even its makers call an entry point.

There is a further wrinkle worth knowing. The classical side of that comparison is not fixed. The "how long would a supercomputer take" estimate depends on the best known classical algorithm, and those algorithms keep improving. Every previous quantum-supremacy headline has had its classical estimate revised downward afterward, sometimes drastically, as researchers found smarter ways to simulate the task. Treating a single supremacy number as a fixed fact about the universe misreads what it is. It is a snapshot of a moving contest.

The overhead is the real story

Return to the memory, because the size of what Willow built is where the honest sense of scale lives.

Willow's chip has 105 physical qubits. The distance-7 code consumed 101 of them, split into 49 data qubits, 48 measurement qubits, and 4 for removing a specific kind of error, and out of all that machinery came exactly one logical qubit. One. And that logical qubit is a memory. No logical gates were performed on it. No algorithm ran. It held a quantum state, checked it, and corrected it. Holding a qubit still is the necessary first step, and it is genuinely hard, but it is not computing with the qubit.

Now put that against what a useful quantum computer needs. Algorithms that would actually matter, factoring large numbers, simulating complex molecules for drug discovery or materials, want logical error rates somewhere around one in a million to one in a trillion. Willow's memory sits at roughly one in seven hundred per cycle. Closing that gap means much larger codes, which means more physical qubits per logical qubit, on the order of a thousand or more each. And a real computation needs not one logical qubit but thousands to millions of them, all held below threshold at once, with gates running between them.

Stack those factors and the shape of the road becomes clear. Willow is one rung on a very tall ladder. It is the rung that proves the ladder can be climbed at all, which is not nothing, and it is nowhere near the top. The distance from "one logical qubit that remembers" to "a million logical qubits that compute" is measured in years of engineering and in fabrication costs that remain largely unpriced.

What actually changed, and how to read the next headline

It would be easy to end on the deflation, but that would repeat the original error in the opposite direction. Something did change, and it is worth naming precisely.

Before Willow, the open question was whether quantum error correction could scale at all, whether the threshold was reachable on real hardware or whether some unforeseen physics would keep machines stuck on the wrong side of the line forever. That question is now answered. The answer is yes, demonstrated, on a chip you can point to. The question that replaces it is an engineering and economics question: how many chips, how many years, and at what overhead until a logical qubit runs a useful algorithm. That is a better question to be asking. It is also not the same as "quantum is here."

For anyone who has to make decisions downstream of quantum headlines, and that increasingly includes people budgeting for post-quantum security or fielding vendor pitches, the Willow episode offers a clean, reusable test. When a quantum result arrives wrapped in a giant number, ask two things. First, what exactly does the big number measure, and does the thing it measures do anything useful? For the septillion years, the answer was no, by the maker's own admission. Second, is the impressive result a computation, or a component? Willow's real achievement was a component, a reliable memory, which is essential and which is not the same as a machine that computes.

Those two questions would have let any reader take the correct lesson from December 2024 without the whiplash. A genuine, thirty-years-coming milestone in error correction, reported alongside a benchmark its own creators call useless, stapled together by a number too large to feel. The science was real. The framing oversold it. And the ability to hold both of those thoughts at once is exactly what the coverage asked its readers to give up.

← Back to blog