What is a qubit — Google needed 101 of them to make one reliable qubit, and why counts mislead

A qubit is the basic unit of information in a quantum computer. An ordinary bit is either 0 or 1. A qubit can hold a blend of both and can be entangled with other qubits. That is the textbook answer. The fact that matters more for understanding real machines comes next: current qubits make so many errors that it takes many of them to build one qubit you can trust.

On December 9, 2024, Google unveiled a 105-qubit chip called Willow and reported using 101 of those qubits to build a single error-corrected qubit. On November 11, 2026 (US Eastern time), IBM announced a next-generation chip design with 520 programmable qubits. Meanwhile, the estimated number of qubits needed to break RSA-2048 encryption fell from 20 million in 2019 to under 1 million in 2025. These three statements only make sense together once you know what each number is counting. This article sorts that out.

1. How a qubit differs from a bit

A bit is a switch: on (1) or off (0). A qubit is built from a very small physical system, such as an atom, an ion, a superconducting circuit or a particle of light, and differs in two ways.

Superposition means a qubit can sit in a blended state between 0 and 1. It carries values, called amplitudes, that set how likely a 0 or a 1 is. But when you measure a qubit you get only a 0 or a 1, and the blend is gone.

Entanglement means several qubits are tied into a single state that cannot be described one qubit at a time. Measure one of two entangled qubits and the result is correlated with the other. Describing n entangled qubits takes 2 to the power of n values. For 50 qubits that is about a quadrillion. This is why ordinary computers struggle to simulate quantum ones.

A common misconception: it computes every answer at once

Quantum computers are often described as trying all possible answers simultaneously. That is half right. A set of qubits can hold many cases in superposition, but the moment you measure, only one of them comes out, at random. Superpose and measure, and you have done no better than rolling dice.

What a quantum algorithm does lies in between. It arranges operations so that the amplitudes of wrong answers cancel and those of right answers reinforce, the way overlapping waves build up or flatten out. Measurement then returns the right answer with high probability. The catch is that only certain problems can be arranged this way.

A quantum computer is not a faster ordinary computer. It is a different kind of machine with an edge on specific problems.

2. Physical and logical qubits: why hundreds become one

Qubits are disturbed by heat, electromagnetic noise and imperfect control signals, and lose their state quickly. On good hardware a single operation fails about 0.1% of the time, or once in a thousand. Google puts the level needed for practical computation at about one in a trillion. Better devices alone are unlikely to close that gap.

The answer is quantum error correction. Information is spread across many physical qubits, the actual devices on the chip, while some of them continuously check whether an error has occurred so it can be fixed. The resulting unit is a logical qubit. A classical computer can copy a bit three times and take a majority vote. A qubit cannot be copied, and looking at it destroys its state. So the scheme never reads the information itself; it only measures whether neighboring qubits disagree. That detour is what consumes so many qubits.

Google's Willow result shows the structure in numbers.

The importance of the result is its direction, not the quality of that one logical qubit. If physical error rates sit above a certain threshold, bundling more qubits makes things worse, because errors appear faster than they can be fixed. Willow showed operation below the threshold, where a bigger grid means fewer errors.

The distance left is still long. An error of 0.143% per cycle is in the same range as a good physical operation, and Google wrote that at current physical error rates, reaching a logical error rate of one in a million takes more than 1,000 physical qubits per logical qubit. Useful computations need hundreds to thousands of such logical qubits. That multiplication is where figures in the millions come from.

There are efforts to shrink the ratio.

IBM's researchers wrote that building the same 12 logical qubits with the surface code would take nearly 3,000 physical qubits. In exchange, their code needs wiring between qubits that sit far apart on the chip. The low-ratio experiments in the table mostly used light error correction, meaning codes that catch only small errors. A large count of logical qubits does not mean each one can survive a long computation.

3. Five ways to build a qubit

A qubit is not one kind of object. Each physical approach has distinct strengths and weaknesses, and no winner has emerged.

The scale of what has been shown differs just as much. Researchers at Caltech reported in March 2024 that they had trapped more than 6,100 neutral atoms and measured a coherence time of 12.6 seconds. That experiment held the atoms; it did not entangle them or run gates. Xanadu's photonic machine Aurora, announced on January 22, 2025, implemented 12 qubits across 35 photonic chips and 13 kilometers of fiber, and the company named optical loss as its next hurdle. Diraq and the Belgian research center imec said in September 2025 that two-qubit silicon devices made on a 300mm production line exceeded 99% operation fidelity. That 6,100, 12 and 2 can all be headline numbers in roughly the same period shows why counts alone do not compare.

4. Why the qubit count alone is not enough

On December 4, 2023, IBM unveiled Condor, a chip with 1,121 qubits. On the same day it presented the 133-qubit Heron with a lower error rate, and it shifted away from doubling qubit counts each year toward machines that resist errors rather than larger ones. Nighthawk, which IBM introduced on November 12, 2025, has 120 qubits. Going from 1,121 to 120 is not a retreat. What is being counted changed.

Simple arithmetic shows why. If one operation, or gate, succeeds 99.9% of the time, the chance of getting through 1,000 gates without a single error is 0.999 multiplied by itself 1,000 times.

This assumes errors are independent, and real systems use error-mitigation techniques to extract better results than the table suggests. The point holds anyway. A thousand qubits are of no use if the result drowns in noise after a few hundred operations. So the following metrics are read alongside the count.

Quantum volume was devised by IBM. A machine runs random circuits on n qubits with n layers; the largest n for which the results pass a statistical test (a heavy-output probability above two thirds) is reported as 2 to the power of n. A value of 2 to the 25th means a 25-qubit, 25-layer circuit ran reliably. A machine with hundreds of qubits and high error rates scores low. Companies also favor different metrics: IBM cites gate counts, Quantinuum cites quantum volume, and IonQ uses its own measure called algorithmic qubits (#AQ). That is one reason cross-company comparison is hard.

5. What the main companies have published

These are figures each company announced. They were not measured by an independent body under the same conditions, and the announcement date is given for each.

IBM Nighthawk r3 is the most recent case. On November 11, 2026 (US Eastern time), IBM announced r3 as designed to run 10,000 quantum operations on 520 programmable qubits, and said it plans to link two r3 chips into a 1,040-qubit system in 2027. The announcement contains the same distinctions drawn above. The 520 is the number of qubits a user computes with; the chip carries far more devices, 2,034 in all. The IBM press release calls these total physical qubits, while SiliconANGLE describes them as physical quantum components including couplers and reset elements. Either way, the 520 are not error-corrected logical qubits. IBM gave no ship date or error rate for r3. Our same-day article on IBM Nighthawk r3 covers the details.

IBM's target for a machine with full error correction is a separate one. In a roadmap released on June 10, 2025, the company said it would build Starling in Poughkeepsie, New York, by 2029 to run 100 million operations on 200 logical qubits. That 520 and 200 come from the same company, and that the smaller number is by far the harder goal, is the reason to keep physical and logical apart.

6. How many qubits would break RSA-2048?

The main reason qubits reach general news is cryptography. A quantum computer large enough to run Shor's algorithm would break the RSA and elliptic-curve cryptography that the internet relies on. Estimates of how many qubits that machine needs have kept falling.

Three points are worth noting.

First, the twentyfold drop between 2019 and 2025 came from algorithms, not hardware. Gidney's 2025 paper assumes exactly the same hardware as the 2019 one. It shrinks the working space the calculation needs and stores idle logical qubits more densely. The runtime rose in exchange, from 8 hours to under a week. Qubits were traded for time.

Second, the under-100,000 figure from February 2026 rests on different assumptions. It replaces the surface code with qLDPC codes, which need direct links between distant qubits, and such links have not been built at scale. The Quantum Insider noted that the architecture has been validated only in numerical simulation, and reported that Gidney has said he does not see another tenfold reduction without changing assumptions.

Third, every qubit in this table is a physical qubit that operates steadily at a 0.1% error rate. Google wrote in May 2025 that quantum computers with relevant error rates currently have on the order of 100 to 1,000 qubits. The requirement is 100,000 to 1 million; the supply is in the hundreds. The gap is somewhere between 100 and 10,000 times. It is shrinking, and it is still wide.

No one can put a date on when it closes. That uncertainty is why the US National Institute of Standards and Technology decided to deprecate RSA and elliptic-curve cryptography after 2030 and disallow them after 2035. Our explainer on post-quantum cryptography covers that timeline and the replacement algorithms, and our explainer on elliptic curve cryptography gives the background on why a 256-bit curve falls to fewer qubits than RSA.

7. Where South Korea stands

South Korea is a useful case of a country building its own hardware while buying foreign machines. Only confirmed items are listed.

The flagship project began research in October 2025 and is budgeted at about 360.4 billion won through 2029. A coherence time of 100 microseconds is in the same range as the 68 microseconds Google reported for Willow. Coherence is one metric among several, though, and as the table in section 4 shows, an entangling accuracy of about 99% is one or two nines short of what long computations require. The largest domestically built machine demonstrated so far has 20 qubits.

KISTI's Tempo was presented in a March 18, 2026 agreement with NVIDIA and IonQ as due to be linked to the country's sixth national supercomputer. No report confirming that installation is complete was found at the time of writing.

8. Frequently asked questions

Does more qubits mean a faster quantum computer? No. The count sets the size of problem that fits. Whether the computation works depends on error rates and on how many operations can be chained. IBM built a 1,121-qubit chip and then changed course to a 120-qubit one.

Does a quantum computer calculate every possibility at once? It can hold many possibilities in superposition, but measurement returns one. The algorithm has to arrange interference so the right answer is likely, and that is known to work for a limited set of problems such as factoring, search and simulating quantum systems.

How many physical qubits make one logical qubit? There is no fixed number. It depends on the physical error rate, the code and the target error rate. Google's 2024 experiment used 101 for one, and Google says a one-in-a-million target takes more than 1,000. IBM's qLDPC design gets 12 from 288.

Why are qubits kept so cold? Only superconducting ones are. Heat disturbs the qubit state, so they are cooled to around 0.015 kelvin. Trapped-ion and neutral-atom machines hold atoms with vacuum chambers and lasers, and photonic machines run at room temperature.

Are my passwords or online banking at risk now? No existing quantum computer can break RSA-2048. Estimates call for 100,000 to 1 million qubits, announced chips have hundreds, and those are qubits before error correction. Encrypted traffic captured now could be stored and decrypted later, however, so long-lived secrets are being moved to new algorithms first.

Is quantum advantage the same as fault tolerance? No. Quantum advantage means beating classical computers on some particular problem. Fault tolerance means running long computations reliably through error correction. Breaking cryptography needs the second.

9. What remains open or unconfirmed

• Cross-company comparison: fidelities and qubit counts here are self-reported. No independent like-for-like comparison was found. IonQ's 99.99% is a prototype result.
• The 2,034 figure for IBM r3: the IBM release says physical qubits, SiliconANGLE says physical quantum components including couplers. r3 is at the design stage.
• RSA-2048 estimates: all are theoretical. The under-100,000 figure from February 2026 is validated only by simulation and needs connectivity not yet built at scale.
• Original papers: the Nature versions of the Willow and Helios papers could not be opened; arXiv abstracts and company statements were used.
• South Korea: whether Tempo has been installed at KISTI and whether the KRISS 50-qubit system has been demonstrated could not be confirmed. The 127-qubit figure for Yonsei rests on one report.
• When it becomes useful: fault-tolerant machines have company targets (IBM says 2029) but no verified schedule.

For a different route to weakening existing cryptography, see our report on DeepMind AI agents finding two RSA-1024 weaknesses. The migration timeline is in our explainer on post-quantum cryptography, the principle behind current public-key systems is in our explainer on elliptic curve cryptography, and the latest IBM announcement is covered in the same-day article on IBM Nighthawk r3.