In this article
We all talk about artificial intelligence. Meanwhile, the next big disruption is moving along without making much noise, inside a lab at 273 degrees below zero.
I’ve been explaining it wrong for years
When people ask me what the next big disruption for business is going to be, I almost always give the same answer: quantum computing. And for years, when they asked me to explain it, I trotted out the version we’ve all read a thousand times.
A classical bit is a zero or a one. A qubit can be both at once. And because it can be both at once, it tries every possible answer in parallel and lands on the right one while your laptop is still working through the third.
It’s a very simple explanation. You get it first time. It plays well in a meeting.
And it’s false.
The argument that takes it apart has been going round for almost twenty years, it’s signed by Scott Aaronson and it fits in two sentences: if a quantum computer really did try every answer at once, measuring it would hand you one at random. Which is exactly as useful as rolling a die. Superposition on its own gets you nothing.
The day you understand why that explanation fails, you can understand two more things. Why this technology has spent thirty years stuck at “ten years away”. And why what has happened in the last two years does matter, even if you never saw it on the evening news. I did catch it myself, tiptoeing past on some bulletin or other.
Five words you are going to see a lot
The first three explain the machine. The last two explain why it concerns you.
The machine
Physical qubit
The actual hardware. Fragile, noisy, breaks down on its own.
Logical qubit
A reliable one, built out of many physical qubits that correct each other. Today it takes dozens to hold up a single one.
Threshold
The point beyond which adding qubits lowers the error instead of raising it. It was crossed in 2024.
And why it concerns you
Q-day
The day a machine exists that can break today’s cryptography.
PQC
Post-quantum cryptography: the new algorithms, built to withstand that machine. It is already being deployed.
What has changed
For three decades, the objection to quantum computing was always the same one: noise.
A qubit is an absurdly delicate object. One vibration, one stray photon, one thousandth of a degree too many, and the information evaporates. So every time someone proposed adding more qubits to build a more powerful machine, the same problem turned up: the new qubits brought more error than they brought capacity. Adding muscle made the whole thing worse. And until that was solved, everything else was just talk. And once again, what fascinates me about the way this problem was attacked is how readily we reach for the everyday, and how far we simplify the pieces to get there.
The theoretical solution has been known since the nineties: use a lot of imperfect physical qubits to build, between them, one logical qubit that behaves. The same way a team of mediocre people can make better decisions than any of its members if it has a decent process for checking each other. The catch is that the theory only works below a certain error level, a point physicists call the threshold. Above it, the cure is worse than the disease.
In December 2024, Google published in Nature the first system that crossed that threshold. For the first time, adding more physical qubits reduced the error of the logical qubit instead of increasing it.
Think of it as a leaking bucket. For years, every extra bucket you put down leaked more than it caught. Now, for the first time, putting down more buckets makes the whole set lose less.
That’s the whole game. It isn’t a spectacular headline, it doesn’t fit in a tweet, and almost nobody outside the field understood it. But since then it’s been replicated by the University of Science and Technology of China, a Harvard team with QuEra and MIT, and Quantinuum. Four different platforms, three continents, same conclusion.
The thirty-year objection stopped being theoretical. And it happened quietly.
How it actually works
Fine. So how does a quantum computer win, if it isn’t by trying everything?
Start with the part that’s true. A bit is a coin that has already landed: heads or tails, whether you look at it or not. A qubit, while nobody is watching it, is more like a coin still spinning in the air. It isn’t that it’s heads and tails at the same time. It’s that the question “which of the two is it?” has no answer yet.
So far, the usual analogy. Now comes the part almost nobody tells you, and it’s the part that explains everything.
What makes quantum physics special isn’t uncertainty. There’s uncertainty in a spinning coin too, and that’s nothing more than ordinary statistical ignorance. What makes it special is that the state of a qubit is described with amplitudes, and amplitudes can be negative. Probabilities can’t. There is no minus thirty per cent chance of rain.
That sentence, on its own, is the difference between the two.
Because if something can be negative, it can cancel out. And there’s the whole trick: a quantum computer doesn’t win by trying every path, it wins by getting the paths that lead to the wrong answers to cancel each other out, and the ones that lead to the right answer to reinforce each other. Like noise-cancelling headphones, which never turn down the volume of the plane: they emit the opposite wave so the noise destroys itself and leaves us in peace.
Programming a quantum computer is coordinating those cancellations.
A bit, a qubit, and what actually does the trick
Bit
It lands on one side.
It is 0 or it is 1. It always was, whether you look or not.
Qubit
Spinning.
While nobody is looking, the question “which one is it?” has no answer yet.
And here is the trick: interference
positive amplitudenegative amplitude
cancels out
cancels out
cancels out
✓ the answer
The paths to wrong answers cancel each other. The right one reinforces.
It does not try every answer at once: it makes the wrong ones cancel out.
Source: www.josejaviernicolas.com
And here comes the awkward consequence, the one that hardly ever shows up in the enthusiastic articles about this technology: to design that choreography of cancellations, you need to know the mathematical structure of the problem very well. Piles of data won’t do it, and neither will wanting to go faster. Which is why the list of things a quantum computer is good for is surprisingly short: factoring numbers, simulating quantum systems, some search problems and little else. And the search one, the famous Grover algorithm, only offers a quadratic improvement, not an exponential one: if something takes a million steps, it comes down to a thousand. Better, yes. Revolutionary, no.
I’ll say it early because it’s the most useful thing I can write today: most of a company’s problems are not going to get better with this. Not now and not when the technology matures. Anyone selling you the opposite either hasn’t understood it or is selling you snake oil.
AI doesn’t fix bad data. It amplifies it. Quantum works the same way: it doesn’t fix a badly framed problem, it leaves it exactly where it was.
Why “how many qubits” is the wrong thing to measure
Every few months a headline turns up in the press announcing a chip with more qubits than the last one. And the detail that turns the number into information is almost always missing: whether they’re physical or logical qubits.
Today you need a lot of physical ones to hold up a single reliable logical one. Google has got one logical qubit out of the 105 physical ones on its Willow chip. Atom Computing and Microsoft announced 24 logical from 1,180 physical. Quantinuum has presented the best published ratio, 48 logical on 98 physical, though it’s worth adding that this mixes corrected qubits with merely detected ones, which is not the same thing, and that one manufacturer’s “logical” qubits aren’t comparable with another’s.
And how many would you need for the thing that would really be frightening, breaking 2,048-bit RSA encryption? On the order of a million physical qubits.
Why “how many qubits” is the wrong thing to measure
Between what exists and what would be needed there are three or four orders of magnitude.
The bars use a logarithmic scale. One vendor’s “logical” qubits are not comparable to another’s.
Source: www.josejaviernicolas.com
Between what exists today and what’s needed there are three or four orders of magnitude. When you read the next qubit figure, ask yourself the question: physical or logical? That’s the difference between counting a company’s payroll and counting everyone who has ever walked through that company’s door.
The scoreboard, taking nobody’s word for it
Google is ahead on error correction and has just opened a second front with neutral atoms, which is an elegant way of admitting it isn’t sure its own technology is the one that wins. IBM has the most detailed roadmap in the field and promises a fault-tolerant machine, Starling, for 2029. Quantinuum shows the best fidelities and has gone public at a valuation that makes your head spin a little if you look at its revenue. IonQ buys competitors. The neutral atoms from QuEra and Harvard have achieved the longest demonstrations of fault-tolerant circuits. And China is running in parallel, on its own technical path and without telling us very much.
Microsoft deserves a paragraph of its own, because it’s the most interesting case and the one that warrants the most caution. Its bet, topological qubits, would be the most elegant of the lot: in theory it would cut the number of physical qubits needed drastically. The problem is that the very existence of the particles it rests on is still under discussion. Microsoft announced its Majorana 1 chip in 2025, and in June 2026 Nature published a peer-reviewed critique arguing that the signals observed can be explained by conventional physics. It isn’t the first time: an earlier paper by the company on these particles was retracted in 2021.
They may end up being right. But if they pull it off, it’ll be a verified headline, not a press release.
And now the caveat that applies to everything above: those dates are dates the manufacturer gives you. They’re set by whoever is selling the product and nobody audits them. They deserve exactly the same trust as any other commercial deadline you’ve watched slip in your career.
Three calendars that are not worth the same
The gap between the three lanes is the argument: only the last one has consequences if you miss it.
Source: www.josejaviernicolas.com
The awkward question: has anyone made money out of this?
No. Not yet.
There isn’t a single company in the world that has solved a real business problem with a quantum computer in a way it couldn’t have solved better, cheaper or faster with an ordinary computer. Not one.
What there is, is a pattern that repeats with almost comic punctuality. Somebody announces a quantum advantage. The community starts giving the problem a bit more love. And a few months later somebody turns up who solves it on a classical computer. In March 2025, D-Wave published a result in Science presented as impossible to match classically. Researchers at the Flatiron Institute reproduced it on a laptop. IBM’s own quantum advantage tracker records cases where a supposed advantage goes from hours to seconds in a matter of months. The phenomenon even has its own name in the literature, dequantization, and a measured average: the classical shortcut usually arrives in about eighteen months.
In July 2026 IBM announced verifiable quantum advantage. One of the field’s leading physicists replied publicly that they shouldn’t be making claims that strong. IBM’s share price barely moved, which is the market’s way of saying “we’ll see”.
But my favourite figure, the one I use when somebody gets too excited in a meeting, is this: Goldman Sachs worked out how many qubits it would need for one of its real portfolio problems. The answer was eight million logical qubits and a runtime of millions of years. They disbanded the team.
Eight million. The best machines in the world are somewhere between one and forty-eight today.
And there’s one last honesty test I’m very fond of. McKinsey puts the market at between 60 and 100 billion dollars by 2035. BCG talks about between 2.5 and 5 billion by 2030. When two top-tier consultancies are a factor of twenty apart, what they’re selling you isn’t a forecast. It’s a range of faith.
The one thing that already has a date in the calendar
And now the part you really should take to your management committee.
First, precisely, because almost every article gets this wrong and generates the useless kind of panic. A big enough quantum computer breaks public-key cryptography: RSA, elliptic curve, Diffie-Hellman. That is, the cryptography protecting secure connections, digital signatures and certificates. Symmetric cryptography it only weakens: AES-256 holds up fine. “We have to migrate” is a different sentence from “the world is ending”.
What breaks and what holds
IT BREAKS
Shor’s algorithm · public-key cryptography
- RSA
- Elliptic curve (ECC)
- Diffie-Hellman
It protects HTTPS connections, digital signatures, certificates, digital identity and blockchain.
It has to be migrated. And it already has an official deadline.
IT HOLDS
Grover’s algorithm · symmetric cryptography and hashes
- AES-256
- SHA-384 / SHA-512
It protects data at rest, backups, passwords and file integrity.
Grover only weakens it: longer keys are enough.
“We have to migrate” is not the same as “the world is ending”. Confusing the two produces the useless kind of panic.
Source: www.josejaviernicolas.com
Second, the most useful tool in this whole article, and you can use it in a meeting without knowing any physics. It’s called Mosca’s inequality and it’s a sum of three numbers. Take three numbers: X, the years a piece of your data has to stay secret. Y, the years it would take you to migrate your cryptography. And Z, the years left until a machine capable of breaking it exists. If X plus Y is greater than Z, you’re already late.
Do the sum with a medical record. It has to stay confidential for twenty-five years, and that’s a conservative figure. A large organisation takes between eight and fifteen years to migrate completely, and that isn’t me talking, it’s the studies and the experience of previous cryptographic transitions: moving from SHA-1 to SHA-2 took seven years, and it only got finished because browsers started rejecting certificates outright. Total: thirty-five years. Now put next to it the experts’ median estimate for that machine to exist, which is around a decade or a little more.
It doesn’t add up. And it doesn’t add up by a lot.
Mosca’s inequality: do the sum before you relax
X · years the data must stay secretY · years it takes you to migrate
X + Y = 35 years, and Q-day arrives around year 13. That is 22 years of exposure you cannot get back.
If X + Y > Z, you are already late. The machine does not need to exist: it is enough that someone stores today what they will be able to decrypt tomorrow.
Source: www.josejaviernicolas.com
That’s where the concept you need to understand comes from: harvest now, decrypt later. The machine doesn’t have to exist for you to have a problem. All it takes is for someone to capture your encrypted traffic today and keep it until they can open it fifteen years from now. I’ll be honest here, because I think it matters: there isn’t one public, attributed incident of this. The urgency doesn’t rest on a demonstrated breach, it rests on the logic of that sum and on how the regulators are behaving. Which is not nothing.
Because the regulators have already moved, and these dates are mandatory. NIST published the new standards in August 2024. In June 2026 the United States signed an executive order requiring its agencies to have encryption migrated by 2030 and authentication by 2031, and it drags their contractors along with it. The European roadmap asks member states for cryptographic inventories and national plans by the end of 2026, and high-risk systems migrated by 2030. The UK sets 2028, 2031 and 2035. And the NIST timetable leaves RSA and elliptic curve deprecated in 2030 and disallowed in 2035.
Notice the detail: those deadlines don’t depend on the technology arriving on time. They apply either way.
And this is already happening, again with no noise. More than two thirds of the human web traffic crossing Cloudflare now travels protected by post-quantum cryptography. Your browser is probably using it already. Nobody told you.
One last note, and it’s the one that unsettles me most: the deadline is shortening through software, not hardware. In 2019 it was estimated that breaking RSA-2048 would require twenty million qubits. A 2025 paper brought the figure down to under a million without assuming any improvement in the machines, purely with better algorithms. A hardware advance can be delayed. A mathematical advance, once published, can’t be un-published.
In pharma: the most credible promise and the most overblown at the same time
If quantum computing is genuinely good for anything commercial, chemistry is the best candidate. And not because it’s fashionable: simulating how molecules behave is a quantum problem by nature, which is why classical computers scale so appallingly badly at it. It’s the one place where the advantage is a mathematical consequence rather than a hope.
Which is why Boehringer, Roche, Moderna, AstraZeneca and Novo Nordisk are in, the last of them having bought themselves an entire machine. And why their press releases are worth reading with a magnifying glass.
A real example: in May 2026 the simulation of a 12,635-atom protein with quantum computing was announced. The work exists and it’s serious, but the bulk of the calculation was done by classical supercomputers, and the quantum part was a fragment. Another: a 234-fold speed-up was published for simulating an enzyme. Read the detail and it turns out it wasn’t run on any real machine: they’re estimates for computers that don’t exist yet.
The line that best sums up the state of the sector I heard from a Merck executive: “it’s always three years away”. And another, from Novo Holdings, plainer still: you can’t go to a client and tell them you use quantum molecular docking, because the computer doesn’t exist right now.
It’s a credible promise. But it simply hasn’t arrived.
What to do on Monday morning
Ordered by real return, not by how good it sounds.
Start with the cryptographic inventory. Which algorithms you use, in which systems, and at which suppliers. It’s properly boring and it’s the only thing in this entire story you can put an official date on. If you don’t know what you encrypt and where, you can’t migrate anything.
Then classify your data by shelf life, not by sensitivity. They’re different things. Whatever has to stay secret for more than ten years is what you should consider migrating first: medical records, genomic data, trials, long contracts, industrial property.
Demand cryptographic agility in new contracts. If you’re buying industrial equipment today with a fifteen-year service life, ask how its cryptography, or any algorithms it uses, get updated. If the supplier can’t answer, you already have your answer.
Run a pilot in the cloud without buying anything, if you ever have a clear use case. But only if it’s clear: the hourly price for some of these emulators can reach 3,000 euros per QPU hour. IBM Quantum, Amazon Braket and Azure Quantum give you access to this kind of machine. The goal isn’t to extract value, because you’re not going to. Just to experiment and reach your own conclusions about this technology in practice.
And now the least glamorous advice, and probably the most profitable: look at quantum-inspired algorithms, which run on classical hardware. The tensor networks knocking down supremacy announcements belong to that family. They work today, on your current infrastructure, and they need neither a cryostat nor a PhD in physics.
Finish with the only question that really matters: what would have to be true in my sector for this to change the business? If you can’t answer it, any investment you make will be an act of faith. And faith is extremely expensive when it’s publicly traded.
The essentials
The news of the last two years isn’t more qubits: it’s that error correction stopped being theory. The popular explanation, the one about trying everything at once, is false; it works by interference, cancelling out the wrong answers, which is why it’s good for so few problems. Nobody has made money with a quantum computer yet. And what already has a date in the calendar isn’t the machine. It’s your cryptography.
Meanwhile, the revolution carries on in rooms at absurd temperatures and in papers nobody tweets. There probably won’t be a day when someone announces that quantum computing has arrived, in the same way nobody told you your browser already encrypts with it in mind.
And I still think it’s the most interesting thing out there right now. Precisely because it makes no noise.
In case you fancy digging deeper
- Google Quantum AI and Collaborators, Quantum error correction below the surface code threshold. Nature 638, 920–926 (2025); published online in December 2024. Open access — the milestone of the decade.
- Scott Aaronson, What Makes Quantum Computing So Hard To Explain? Quanta Magazine, June 2021 — the right analogy, from the man who has spent twenty years correcting us.
- Craig Gidney (Google Quantum AI), How to factor 2048 bit RSA integers with less than a million noisy qubits. arXiv:2505.15917 (2025), preprint without peer review — why the deadline shortens through software.
- D. Bluvstein et al. (Harvard, MIT and QuEra), A fault-tolerant neutral-atom architecture for universal quantum computation. Nature 649, 39–46 (2026); online since November 2025. Open access.
- NIST, the FIPS 203, 204 and 205 standards (August 2024) and Transition to Post-Quantum Cryptography Standards (NIST IR 8547), which has been sitting in initial draft since November 2024.
- Michele Mosca and Marco Piani, Quantum Threat Timeline Report 2025. Global Risk Institute and evolutionQ, March 2026 — the annual expert survey on “Q-day”.
- Bas Westerbaan, State of the post-quantum Internet in 2025. Cloudflare blog, October 2025 — how much of the web already travels encrypted with this in mind.
- McKinsey, Quantum Technology Monitor 2026: A commercial tipping point (April 2026) and BCG, Quantum Is Getting Real. CEOs Need to Shape Where It Creates Value (June 2026) — read them together, the contrast is the lesson.
All eight sources are freely available: both Nature papers are open access with copies on PMC and arXiv, and the reports download without registration.
