Somewhere in the next five years, the most expensive hour in computing will stop being the one you buy from a cloud GPU vendor and start being the one you buy from a dilution refrigerator.
That sounds like science fiction until you look at the only number that actually matters in quantum computing: how many physical qubits you must pay for to get one that behaves. In 2021 that number was around a thousand. Today, on the best published results, it is under a hundred. Everything downstream — drug discovery timelines, battery chemistry, the cost of a cryptography migration, the shape of the cryogenics supply chain — is a function of that single ratio.
This article is a developer's map of the stack: what actually changed, what the numbers imply, and where the market arithmetic lands if the curve holds.
The number that governs everything
A physical qubit is a terrible computer. It decoheres, it drifts, it picks up noise from a nearby pump. A logical qubit is a fiction you maintain by measuring many physical qubits and correcting their errors faster than they accumulate. The ratio between the two is the overhead, and the overhead is the price of the whole industry.
Two things are happening in that chart at once, and they are not independent. The bars fall because the underlying gate fidelities improved enough to run below threshold — the regime where adding more qubits genuinely reduces the logical error rate instead of adding more noise. The line rises because, once you are below threshold, you can concatenate: each new generation of hardware converts a fixed number of physical qubits into a larger number of logical ones.
The practical consequence is a compounding effect that hardware people like to call "the other Moore's law" and that finance people should call what it is: a cost curve.
What a logical-qubit-hour actually costs
Vendors do not publish this number, so you have to build it. Take the overhead ratio, multiply by the cost of a physical qubit-year (device, wiring, cryogenic plant amortisation, duty cycle, and the fact that a dilution fridge runs whether or not you are computing), then divide by the fraction of wall-clock time the machine is actually coherent.
The interesting part is not the absolute level — it is the slope. On my model, a logical qubit-hour costs roughly four thousand dollars in 2024 and lands in the low tens of dollars by the early 2030s. That is the same order of magnitude as a well-provisioned classical core-hour, which is the point at which "should we run this on quantum hardware?" stops being a research question and becomes a procurement question.
Two caveats worth stating loudly, because the marketing around this field does not:
- The comparison is only fair for a narrow class of problems. A logical qubit-hour buys you simulation of a 100-spin system, not a database query. The crossover in the chart is a statement about price per unit of a specific kind of work, not a claim that quantum computers will run your web backend.
- The model is sensitive to duty cycle. If your machine is idle 90% of the time because calibration eats the schedule, the effective cost per useful hour is ten times higher. This is why the vendors pushing "quantum as a service" hardest are also the ones investing in automated calibration — it is a margin lever, not a convenience feature.
Where the money actually is
If you read only the hardware press, you would think this market is a race to build the biggest fridge. The revenue mix says otherwise.
Three observations from that split:
- Hardware revenue is lumpy and slow. A single dilution refrigerator is a multi-million-dollar capital item with an 18-month lead time and a customer list you can count on two hands. It is a great business for four companies and a terrible market to enter.
- Cloud access scales with developer adoption, not capex. The moment a chemist can rent 20 minutes of a 24-logical-qubit machine for the price of a nice dinner, the buyer set expands from national labs to every materials startup on earth. This is where the growth rate lives.
- Services and software is the quiet winner. Compilers, error-mitigation libraries, resource estimators, calibration tooling — the layer that turns "we have qubits" into "we have an answer" — carries software margins on hardware constraints. It is also the only layer a small team can enter without a cleanroom.
The arithmetic that follows: if the total market reaches roughly forty billion dollars by the early 2030s and services plus cloud take two thirds of it, then the addressable software market is larger than the entire hardware market by a factor of two. Most of the "quantum startup" pitch decks I see are still aimed at the smaller half.
What this does to an ordinary week
Here is the part that never makes it into the funding announcements, and it is the part that will touch you personally.
Drug discovery timelines compress unevenly. Quantum chemistry is the closest thing to a killer app: the threshold for beating the best classical method on a catalyst or a binding-affinity problem is a couple of hundred logical qubits, not a hundred thousand. When that lands, the early phase of drug discovery — the part that currently takes three to five years of wet-lab iteration — becomes a simulation problem with a wet-lab verification step. The bottleneck moves downstream to clinical trials, which are governed by biology and regulators, not by flops. Expect the industry's shape to shift accordingly: fewer chemists doing synthesis triage, more people doing protocol design.
Cryptography migration gets a deadline it can plan against. Factoring a 2048-bit key needs something like four hundred thousand logical qubits by most published estimates — an order of magnitude past anything on the roadmap. But "an order of magnitude past the roadmap" is exactly the kind of statement that turns into a compliance requirement five years early. If you maintain anything that signs with RSA or a non-post-quantum elliptic curve, the migration clock is not set by the day the machine exists; it is set by the day your auditor reads the same chart you just did.
The cryogenics supply chain becomes a strategic asset. Helium-3, dilution units, ultra-low-noise amplifiers, and the handful of fabs that make superconducting films are all single points of failure. When a national lab orders twenty machines, the constraint is not the qubit design — it is the number of people on the planet who can braze a mixing chamber.
Three numbers to watch
If you want to track this field without reading every preprint, track these:
- Physical-to-logical ratio on the best published result. When it drops below 50, error correction stops being the story and applications start being the story.
- Wall-clock duty cycle of publicly bookable machines. Vendors rarely publish it; you can infer it from queue times and calibration windows. Every ten points of duty cycle is a ten percent cut in effective cost.
- Number of logical qubits available in a single job, not in a single device. The distinction matters: twenty-four logical qubits split across four isolated registers is a very different machine from twenty-four that can be entangled.
The uncomfortable conclusion
The quantum industry has spent a decade selling a threshold event — "quantum advantage" — as if it were a light switch. It is not. It is a cost curve, and cost curves are boring, predictable, and far more consequential than press releases. The companies that win will not be the ones that announce the biggest number; they will be the ones that quietly make a logical qubit-hour cheap enough that a materials scientist can spend one without filing a grant application.
That is a market you can actually model. And on the current slope, it arrives before the end of the decade.
One more thing
If this analysis saved you a week of paper-reading — or gave you a number you can put in a budget slide — a tip keeps this work going. Pulling these figures apart and drawing the charts takes real effort.
SOL: HCqPmPJM3NVtS6YrmVaTZYnuP8XZ2PFgQZKT94VK1LkS
Creating this takes real effort — thank you for the support.
Top comments (0)