Quantum computing companies keep asking investors to accept one large assumption: that recent progress in hardware puts the industry on a reasonably clear path to commercially useful computing. On that account, the science is largely settled and the remaining work is engineering, manufacturing, capital and time.
The evidence does not yet justify that confidence.
The physics works. Hardware is improving quickly. Quantum processors can now perform some calculations that are extremely difficult to reproduce on a classical computer. None of this establishes that they will solve an important commercial problem faster or more cheaply than the best classical alternative.
Those are three different tests:
- Can the company build a larger, more reliable quantum processor?
- Can that processor perform a calculation beyond the practical reach of a classical computer?
- Can it produce an answer of commercial value, at an acceptable cost, that a customer could not obtain more easily another way?
Pitch decks often treat these as stages on the same roadmap. They are separate technical and economic problems. Progress on the first does not guarantee the second, and neither guarantees the third.
The hardware progress is real
Google's Willow processor was a major advance. In December 2024, Google reported below-threshold surface-code memories on a 105-qubit superconducting processor. As the error-correcting code became larger, the logical error rate fell rather than rose. The reported suppression factor was about 2.14 for each increase of two in code distance (Nature). That is an important result because useful quantum computers will need error correction that improves as more physical qubits are added.
Other approaches are advancing too. Quantinuum's 98-qubit Helios trapped-ion processor reported average two-qubit gate fidelity of about 99.92 per cent (Nature). A Harvard, MIT and QuEra team operated a neutral-atom processor using as many as 448 atoms and demonstrated key parts of a fault-tolerant architecture with as many as 96 logical qubits (Nature). These results matter. They also measure different things, so the raw qubit numbers cannot sensibly be ranked as though they were processor clock speeds.
Google's Quantum Echoes result in October 2025 was another serious milestone. Google reported that an out-of-time-order correlator ran about 13,000 times faster on Willow than its estimate for the best known classical method on a leading supercomputer (Google). Unlike a random sampling benchmark, the experiment was designed so that another quantum computer of similar quality could repeat it.
The commercial interpretation needs more care. Google's announcement also described a molecular-structure experiment based on nuclear magnetic resonance. It was a separate proof of concept, used only nine to fifteen qubits and remained classically simulable. Google's team said at the time that the molecular result was not beyond classical computing (Science News). The beyond-classical experiment and the proposed application were connected by a research programme, not by a demonstrated commercial advantage.
That distinction is easily lost. A demanding physics experiment can show that a processor is doing something genuinely difficult without showing that the result is useful to a customer.
Classical computing keeps changing the comparison
Quantum advantage is measured against the best available classical method. That benchmark moves whenever researchers find a better algorithm, approximation or way to exploit the structure of the problem.
Google's 2019 Sycamore experiment is the best-known example. Google estimated that its random-circuit sampling task would take a leading classical supercomputer around 10,000 years. Improved tensor-network methods cut the classical cost dramatically. A later implementation using 1,432 GPUs reported sampling at a specified fidelity seven times faster than the quantum processor (Physical Review Research). That did not make the original experiment scientifically worthless. It showed that the size of the claimed advantage depended on what classical researchers knew how to do at the time.
IBM's 2023 Eagle experiment was described more cautiously as evidence of quantum "utility" before full fault tolerance. IBM did not claim that it had beaten every approximate classical method. Soon afterwards, researchers at the Flatiron Institute used a tensor-network method on a MacBook Pro to produce results that were more accurate than the quantum processor. Their method worked because the correlations created on IBM's heavy-hexagonal lattice were unusually close to a tree structure (PRX Quantum). The geometry of the experiment gave the classical algorithm an opening that earlier comparisons had missed.
Several prominent advantage claims have therefore been narrowed or beaten after publication. It would go too far to say that every previous claim has fallen. The more useful conclusion is that advantage is provisional. It belongs to a particular machine, problem, accuracy level and classical comparison at a particular time.
IBM now reflects this in its Quantum Advantage Tracker, which classifies candidate results and records when later classical work supersedes them. This is a better model than treating advantage as a line crossed once and forever.
The language has not kept up with that caution. Quantum supremacy, utility, computational advantage and commercial advantage are still used as though they mean roughly the same thing. They do not. Nature noted in late 2025 that systems remained difficult to compare because research groups used different hardware, algorithms and performance measures (Nature). Ambiguous terminology is useful in marketing because a result proved under one definition can be heard as a claim under a much stronger one.
Better hardware may still not produce a useful application
Large, fault-tolerant machines would remove one major constraint. They would not supply the missing algorithms or make every theoretical speed-up worthwhile.
Many proposed applications offer a quadratic speed-up rather than an exponential one. Quadratic gains can disappear once the full cost of error correction, data preparation and repeated measurement is included. A 2021 study by Google researchers modelled this under a specific set of hardware assumptions. It estimated an error-corrected Toffoli gate at about 170 microseconds and found that some proposed quadratic speed-ups would take years or centuries to reach break-even against conventional computing. One model of quantum-accelerated simulated annealing on a 512-variable problem produced a break-even time of roughly 880 years (PRX Quantum). These numbers are not physical constants; better hardware can change them. Their importance lies in showing that an asymptotic speed-up on paper does not guarantee a practical one.
Torsten Hoefler, Thomas Häner and Matthias Troyer made the same point more broadly. Quantum computers have limited input and output bandwidth. They are most plausible for problems that require immense computation over relatively small amounts of data and deliver more than a quadratic speed-up. Tasks that require loading very large conventional datasets may lose much of their theoretical advantage before the quantum calculation begins (Communications of the ACM).
That qualification matters for machine learning. Some early quantum machine-learning proposals offered exponential improvements only if data could be prepared efficiently in a quantum state, the problem had the right mathematical structure, the answer required limited precision and the output could be extracted cheaply. Scott Aaronson set out these restrictions in 2015 (Nature Physics). In 2018, Ewin Tang then found a classical algorithm matching the speed-up of a prominent quantum recommendation method under comparable data-access assumptions (arXiv). This did not disprove quantum machine learning. It removed one of its most persuasive examples.
Optimisation has a similar problem. A 2024 review involving 55 researchers from academia and industry found no demonstrated quantum advantage for practical optimisation and set out several barriers, including input-output costs, limited circuit depth and the strength of modern classical solvers (Nature Reviews Physics). There may eventually be valuable quantum optimisation methods. The evidence does not support treating them as a natural consequence of adding more qubits.
Chemistry and materials science remain stronger candidates because quantum systems may be well suited to modelling other quantum systems. Even here, the case is narrower than the standard pitch. A 2023 study involving Ryan Babbush, John Preskill, Garnet Chan and others found that convincing evidence for a generic exponential advantage in ground-state quantum chemistry had not yet been established (Nature Communications). Preparing a useful starting state can itself be extremely difficult. Resource estimates for important molecules remain large and depend heavily on assumptions about error rates, code cycles, state preparation and the required accuracy.
Cryptography provides the clearest important algorithm. Shor's algorithm would break widely used public-key encryption on a sufficiently capable fault-tolerant machine. Craig Gidney's 2025 estimate reduced the projected resources for factoring RSA-2048 to fewer than one million noisy physical qubits and under a week of runtime under its assumed architecture (arXiv). That is a major improvement. No universal quantum computer operating today is close to that scale.
Verification also depends on the application. Some useful answers, including the factors of a large integer, are easy to check classically. Certain sampling experiments are deliberately hard to verify by direct classical calculation; repeating them on another quantum processor is not the same as independently proving every output. Verification is a serious issue for some applications, not a single ceiling over the whole field.
The unresolved questions are now clear. Can fault-tolerant hardware of the required size and depth be built economically? Are there algorithms that produce a large enough gain on problems customers value? Will that gain survive the best classical comparison available when the machine arrives? Can the answer be checked cheaply enough for the customer to rely on it? Qubit count answers none of them.
Why the roadmaps sound more certain than the science
Researchers inside the field speak openly about these gaps. John Preskill and Jens Eisert described a "fraught road to quantum advantage" in 2025 and separated progress in experiments from the harder task of producing broad practical value (arXiv). DARPA has built an entire programme around independent tests of whether proposed machines can create more value than they cost.
Public company presentations leave less room for doubt because much of the valuation rests on future milestones.
IonQ's March 2021 investor presentation forecast US$237 million of revenue and US$61 million of positive EBITDA in 2025. It placed "early quantum advantage" in machine learning and financial services in the same year (IonQ investor presentation). IonQ reported US$130 million of 2025 revenue and an adjusted EBITDA loss of US$186.8 million (IonQ). The company also broadened through acquisitions into networking, sensing, security and semiconductor manufacturing. IonQ says its 2025 organic revenue grew by nearly 80 per cent, so it would be misleading to describe the full increase as acquisition-driven. The original commercial and advantage forecasts were nevertheless missed by a wide margin.
Rigetti's October 2021 presentation forecast US$288 million of revenue, US$142 million of positive EBITDA and more than 1,000 qubits by 2025. Actual 2025 revenue was US$7.1 million, with a US$50.5 million non-GAAP net loss. In March 2026 the company said deployment of its 108-qubit system was expected in the second half of that year (Rigetti).
Quantinuum shows that strong technology does not resolve the valuation question. Its June 2026 IPO priced at US$60 a share and implied a valuation of about US$15.8 billion, before the shares rose on their first day (Quantinuum). The company had reported US$30.9 million of revenue for 2025. In August it guided to US$28 million to US$32 million for 2026, while reporting strong second-quarter growth and continued technical progress (Quantinuum). The IPO was a substantial bet on its technical lead and future market, not a valuation supported by current revenue.
These misses do not prove that the companies will fail. Frontier hardware businesses often spend for years before revenue catches up. They do show that quantum roadmaps have been poor forecasts of commercial timing. Investors should treat them as plans with scientific dependencies, not conventional product schedules.
IBM offers a useful contrast. It delivered processors with 127, 433 and 1,121 physical qubits across 2021, 2022 and 2023, then changed its roadmap as the field shifted from raw qubit count towards modular systems and error correction. The revision was sensible. It also showed why a roadmap should be allowed to change when the evidence changes. A company valued mainly on meeting that roadmap has a harder time making the same admission.
Capital has continued to arrive. McKinsey estimates that investment in quantum technology startups reached US$12.6 billion in 2025, 6.3 times the 2024 level, with about 60 per cent concentrated in the ten largest deals. It estimates worldwide quantum-computing revenue at a little over US$1 billion for the year (McKinsey). The investment figure includes large late-stage and capital-markets transactions, so it should not be read as venture funding alone. The gap still shows how far present valuations reach beyond present sales.
The Australian bet
Australia has made one of the largest public bets in the sector. The Commonwealth and Queensland governments committed about A$940 million to PsiQuantum in 2024 through equity, loans and grants. The support is conditional on technical, project, financing and employment milestones, with independent technical review before payments (Queensland Audit Office). That structure matters: the full amount was not an unconditional grant paid upfront.
PsiQuantum originally planned a Brisbane Airport site and said it was working towards operation by the end of 2027. It later moved the project to Moreton Bay and began construction there in June 2026 (PsiQuantum). Jeremy O'Brien moved from chief executive to executive chair and Victor Peng became chief executive. The 2027 target remains exceptionally ambitious and has not yet been demonstrated.
Australia also has two companies in Stage B of DARPA's Quantum Benchmarking Initiative: Diraq and Silicon Quantum Computing. They are among eleven companies whose plans DARPA is examining in detail. Stage B lasts a year and focuses on the R&D plan, risks and prototypes required to reduce those risks. Selection is evidence that the approaches deserve serious examination, not that DARPA has validated a working machine (DARPA).
DARPA's wording is unusually useful. Its programme manager said in March 2026 that a utility-scale machine by 2033 now appeared likely, while the team or technical approach most likely to deliver it remained unclear (DARPA). The agency is optimistic enough to invest and uncertain enough to test the claims independently. Investors should take the same position.
What an investor should ask
The useful question is not how many qubits a company expects to have. It is what exact customer problem the machine will solve better than the best classical alternative, measured across the whole job.
That comparison should include the physical and logical qubits required, error-correction overhead, circuit depth, runtime, energy, data preparation, repeated measurements, verification and total cost. The classical baseline should be the best method available, tuned by someone with an incentive to make it win.
Each technical milestone should then be assigned to one of the three tests: better hardware, beyond-classical computation or customer value. A sentence that begins with an error-correction result and ends with drug discovery has probably moved between them without supplying the missing evidence.
Revenue needs the same treatment. Hardware sold to a government laboratory, a research contract, cybersecurity software, consulting work and payment for a computational result unavailable classically are all legitimate revenue. They do not prove the same thing. Public filings rarely separate them cleanly enough to show how much revenue comes from demonstrated computational advantage.
A pitch that depends on future hardware, a future algorithm and a classical comparison that has not been run is a research hypothesis, not a proven product roadmap. It may still be worth backing. The technical progress is real, the strategic value could be enormous and one or more of today's approaches may work. The price should reflect the unresolved science, the shifting classical benchmark and the possibility that a useful machine arrives later, serves a narrower market or costs more than the roadmap assumes.
Quantum computing is an option on a major scientific and industrial breakthrough. It should be valued as one.