If you work anywhere near AI strategy, someone has probably told you to get a quantum computing roadmap. Study it properly. Start now.
Good syllabi for that already exist. They're written by people who know the physics far better than I do. But a syllabus answers the second question. The first one is harder: how much of the urgency around this is real? That's what decides how deep you go, and in what order. And it's the one question the people selling quantum can't answer honestly, because the honest answer is that most of it doesn't touch your work yet.
So here's a method for that first question. It works on any new technology. Quantum is just the example, because it splits so cleanly.
One thing before we start. Nothing below is a prediction about quantum. If a result lands in 2028 that nobody saw coming, this should still read as correct. The whole argument is about how to decide what to study, and when.
The pitch you've probably heard
It goes like this. Waiting is the real risk. The organizations that start now will set the pace. Often it's attached to a real announcement, like a large company starting quantum training for its staff.
Some of that is true. The trouble is which part. Hold that thought. We'll come back to it once the method is on the table.
Question one: is a cost piling up while you wait?
Real urgency has a mechanism. Something specific accumulates while you do nothing, at a rate you could describe. "You'll fall behind" isn't a mechanism. It's a feeling about competitors.
If I do nothing for two years, does my exposure grow, or stay flat?
Flat means you can wait. Growing means a clock is running.
Is damage being done now and discovered later?
Some costs are taken today and only show up on the bill years from now.
How long does the fix take once I start?
If the fix takes years, the clock started years before the threat shows up. That's how something immature can still be urgent.
Is someone else creating my cost, whatever I do?
An attacker, a regulator, a customer contract. Their move can start your clock without you.
If you can't name what's piling up, the answer is "nothing yet."
Question two: will what you learn now survive the technology maturing?
Learning has a shelf life. Some of it keeps for decades. Some of it expires with the next hardware generation.
Am I learning the shape of the problem, or today's tools?
The shape transfers. The tools get replaced.
If the hardware got a hundred times better tomorrow, how much of this transfers?
Be honest about the answer. Often it's "not much."
Is it written down in a published standard, or in one vendor's product?
A standard outlives the company that sold you the training.
Would someone starting three years from now be behind you, or ahead of you because the tools got better?
That's the sharpest question in the set. I call it the latecomer test. If the latecomer would be ahead, studying now costs you more than it returns. Unless a cost is piling up while you wait.
Four answers
Put the two questions together and every piece of a technology lands in one of four boxes.
Act now, in depth
The clock is running and what you build keeps its value.
Act now, minimally
You have to move. Buy containment and skip mastery. Expect to redo it.
Study at leisure
Nothing's accruing and the knowledge keeps. Go deep, no hurry.
Skip, or sample cheaply
Nothing accrues and the knowledge dies. Most manufactured urgency lives here.
The step that makes this work happens first. Split the technology into the separate things you could actually spend time on. One verdict for a whole field is almost always wrong.
Quantum, split three ways
1. Moving off today's encryption: act now, in depth
This is the piece with a real clock. It's worth being exact about the mechanism.
Someone can record encrypted data today and keep it. Once a capable enough quantum computer exists, they can decrypt it. NIST, the US standards body that runs the replacement effort, calls this "harvest now, decrypt later." The IETF, which writes the internet's protocols, describes the same risk: encryption "broken retroactively" by anyone who recorded the traffic.
That's damage done now and discovered later. Then look at the fix. NIST's draft transition report says the move from a new standard to full deployment has historically taken 10 to 20 years.
Think about who holds data that has to stay secret that long. A bank, with encryption buried in payment systems, partner links, archives and vendor products it doesn't control. A drug company, with trial data and formulas that stay valuable for decades. A manufacturer, shipping equipment that sits in the field for twenty years with its keys built in. For each of them the clock started before the machine shows up.
And what you learn keeps. In August 2024 NIST finalized its first three post-quantum standards: FIPS 203 (ML-KEM), FIPS 204 (ML-DSA) and FIPS 205 (SLH-DSA). They're published standards. No single vendor owns them.
So a bank treating this as urgent is right. Look at what the urgent work actually is, though. An inventory of where your encryption lives. A migration plan. Questions for your vendors. None of it involves programming a quantum computer.
You don't need the math. You need to know what harvest now, decrypt later means for the data your organization keeps, and to be able to ask whether anyone has mapped where that data's encryption lives. That's a few hours of reading, and it keeps.
2. Knowing which problems are quantum-shaped: study at leisure
This is the reasoning layer. Which kinds of problems actually get faster on a quantum machine, and what has to be true for that to happen.
In 2015 Scott Aaronson took the best-known quantum algorithm for solving systems of linear equations and listed its fine print. Four caveats. The data has to be loaded into the machine fast. The matrix needs a special structure, usually sparse. It has to be well-conditioned. And the output is a quantum state, so reading out the actual answer can cost you the speedup. Any one of them can erase the advantage in practice.
In 2018 Ewin Tang looked at a quantum recommendation algorithm that had been seen as one of the strongest candidates for an exponential speedup in machine learning. She found a classical algorithm that matched its scaling, once the classical side was given comparable sampling access to the data. Much slower in practice, with large polynomial overheads, but the exponential gap was gone. Follow-up work did the same for many low-rank machine-learning algorithms.
It does not show that quantum machine learning has no advantage. Its own authors say it leaves sparse, high-rank problems alone and does not rule out large polynomial speedups. What it shows is narrower: specific claimed speedups, under specific assumptions about data access, were matched classically once the classical algorithm got comparable access.
That's the useful habit. When someone claims a speedup, find the assumption about the data. That habit survives every hardware generation, and nothing piles up if you learn it next year. Study it at leisure. Go as deep as you like.
3. Programming today's machines for deep learning: wait, or sample cheaply
This is usually the piece being sold. Run the two questions on it.
Is a cost piling up while you wait? For most people doing deep learning, nothing specific. Nobody is taking anything from you while you hold off. As of October 2026 I couldn't find an accepted demonstration of quantum advantage on a commercially useful machine-learning task. There are credible advantage claims on specialised scientific and benchmark tasks, and the field moves fast. That's why this verdict comes with a trigger instead of a forecast.
Will what you learn survive? Today's largest machines hold hundreds to a few thousand physical qubits. With today's main error-correction approach, one reliable logical qubit takes on the order of a thousand physical ones. Published estimates for a machine that could break RSA-2048 run from roughly 100,000 to about a million physical qubits, and they keep falling. Whatever you'd program for real work sits on the far side of that gap. How much of today's craft carries across it isn't known.
Now the latecomer test. Someone who starts in three years gets better tools and a clearer picture of what works. They're probably ahead of you on this piece. So it lands in the bottom right: wait, or sample cheaply. A weekend with a simulator, to know what the thing feels like, is a fine sample.
Now back to the pitch
Look at what "waiting is the real risk" does. The waiting risk is real for the first piece. Encrypted data can be taken today and read later, and the fix takes years. What's on offer, though, is usually the third piece: training on today's machines.
Both halves are true on their own. The pitch never says one causes the other. It just puts them next to each other and lets you connect them. That's why it survives a fact check.
One piece of a technology has a real clock. That clock gets lent to a different piece, the one being sold. Nothing false is said. The link is implied.
Five tells of manufactured urgency
Hold these up against any claim in front of you.
The urgent thing and the sold thing share a name
"Quantum" covers both an encryption migration and a training course. Only one of them has a clock.
The cost of waiting is stated as a race
"You'll be behind." Compare that with "this specific thing builds up at this rate." Only the second one is a mechanism.
Nobody names what accrues
Ask. If there's no answer, there's no clock.
The remedy is training or tooling, never containment
A real accruing cost usually has a cheap containment step. An inventory. A plan. A contract clause. If that's never offered, notice it.
Whoever says it's urgent sells the fix
That doesn't make them wrong. It means you check the mechanism with someone who doesn't sell anything.
One tell is a reason to look harder. Several together usually mean the urgency is real somewhere else, for a different piece than the one being sold.
So, in what order
If you've committed to learning quantum, good. Here's the order the method gives.
The encryption piece, now
Enough to know what harvest now, decrypt later means for your organization's data and what a migration involves. A few hours of reading. Skip the math.
The concepts, at your own pace
Which problems are quantum-shaped. Aaronson's fine print. Tang's result and its limits. Take your time. It keeps.
Today's hardware for deep learning, when a trigger fires
Sample it if you're curious. Otherwise wait for an accepted advantage on a task like yours, or for the tools to stop changing every year.
That order is the deliverable. It doesn't need quantum to fail or succeed. It moves when the facts move.
Who this is wrong for
Every method has people it fails. This one has at least three.
Anyone whose curiosity is the point. The method ranks study time by return. If you're learning quantum because you find it beautiful, you're not making the mistake it describes. Ignore me.
Anyone in a field where the clock can't be read yet. The method needs you to know whether a cost is accruing. For some technologies nobody knows. Forcing an answer gives you false confidence, and that's worse than the uncertainty it replaced. Mark it "unreadable," sample it cheaply, and decide what news would make you look again.
Anyone whose value comes from being early. Being the person who can talk to a board about quantum is worth something, whether or not quantum pays off. This method scores that at zero. It isn't zero. If that's your job, the conceptual piece is your accruing cost, and it keeps.
Run it on your own situation
Nothing to buy. Pick a technology you've been told is urgent and give it twenty minutes.
Split it
Write down the separate things you could spend time on. A migration, a concept, a tool, a skill.
Ask both questions of each piece
What piles up while I wait, and who causes it? Would a latecomer be behind me or ahead of me?
Put each piece in a box
Act now in depth, act now minimally, study at leisure, or skip.
Write what would move it
A standard gets published. A regulator sets a date. The tools settle down. When that happens, run it again.
You end with a written order. What to learn first, what can wait, and what would change your mind.
One way this goes wrong
Using it to win an argument. "Skip" in the bottom box is a ranking of your time. It says nothing about whether the field matters. The day someone uses this to tell a colleague quantum is hype, it's stopped being the method.
Sources
Every figure here was pulled from the source below on October 6, 2026. Hardware counts and resource estimates change fastest, so they're given as ranges with that date.
NIST, "NIST Releases First 3 Finalized Post-Quantum Encryption Standards," August 13, 2024. www.nist.gov
NIST IR 8547 (Initial Public Draft), "Transition to Post-Quantum Cryptography Standards," November 2024. Source of the harvest-now, decrypt-later description and the 10 to 20 year deployment history. Still a draft as of this writing. csrc.nist.gov
IETF RFC 9954, "Hybrid Key Exchange in TLS 1.3," July 2026, Section 1.2. www.rfc-editor.org
Scott Aaronson, "Read the fine print," Nature Physics 11, 291 to 293, 2015. Author's copy. www.scottaaronson.com
Ewin Tang, "A quantum-inspired classical algorithm for recommendation systems," STOC 2019. arxiv.org
Chia, Gilyen, Li, Lin, Tang and Wang, "Sampling-based sublinear low-rank matrix arithmetic framework for dequantizing quantum machine learning," STOC 2020. Source of the stated limits on dequantization. arxiv.org
Fowler, Mariantoni, Martinis and Cleland, "Surface codes: Towards practical large-scale quantum computation," Physical Review A, 2012. arxiv.org
Gidney, "How to factor 2048 bit RSA integers with less than a million noisy qubits," 2025 preprint, and Webster et al., "The Pinnacle Architecture," 2026 preprint from a quantum startup, not yet peer reviewed as far as I could find. The two ends of the estimate range. arxiv.org
Manetsch et al., "A tweezer array with 6100 highly coherent atomic qubits," Nature, 2025. Upper end of today's physical qubit counts. arxiv.org
Questions people ask
Should I learn quantum computing now?
Split it first. Learn what harvest now, decrypt later means for your organization's data now, because encrypted data can be recorded today and decrypted later and migrations take years. Learn which problems are quantum-shaped at your own pace, because that knowledge keeps. Programming today's machines for deep learning can wait, or be sampled cheaply, unless your job depends on it.
Is the urgency around quantum computing real?
For one piece, yes. Encrypted data can be harvested now and decrypted once capable machines exist, and NIST notes that moving from a new standard to full deployment has historically taken 10 to 20 years. That urgency belongs to the cryptographic migration. It does not transfer automatically to training on today's quantum hardware.
What is the latecomer test?
Ask whether someone starting three years from now would be behind you, or ahead of you because the tools improved. If they would be ahead, studying now returns less than it costs, unless a cost is piling up while you wait.
Did dequantization prove quantum machine learning doesn't work?
No. Ewin Tang and later collaborators showed that specific claimed exponential speedups, mostly for low-rank problems, could be matched classically once the classical algorithm got comparable access to the data. The same authors say their results leave sparse, high-rank problems alone and do not rule out large polynomial speedups.
How do I tell if a technology's urgency is manufactured?
Check five things. Is the urgent thing the same thing being sold? Is the cost of waiting described as a race or as a mechanism? Does anyone name what accrues? Is the remedy training or tooling with no containment step? Does the party claiming urgency sell the fix? Several together usually mean the urgency is real somewhere else.